refactor(shutdown): move shutdown logic in one place

- Monkey-patch transformers.AlbertModel in plugins/kokoro/__init__.py before kokoro imports it
- Works with transformers 4.x (no-op) and 5.x (adds moved symbol)
- No pinning, forks, or extra files needed

.gitattributes

@@ -0,0 +1,15 @@
+*.py text eol=lf
+*.md text eol=lf
+*.yml text eol=lf
+*.yaml text eol=lf
+*.toml text eol=lf
+*.json text eol=lf
+*.txt text eol=lf
+*.html text eol=lf
+*.css text eol=lf
+*.js text eol=lf
+*.sh text eol=lf
+*.cfg text eol=lf
+*.ini text eol=lf
+*.svg text eol=lf
+*.j2 text eol=lf

.github/FUNDING.yml

@@ -0,0 +1,15 @@
+# These are supported funding model platforms
+
+github: [jborza, jeremiahsb, mohangk]
+patreon: # Replace with a single Patreon username
+open_collective: # Replace with a single Open Collective username
+ko_fi: # Replace with a single Ko-fi username
+tidelift: # Replace with a single Tidelift platform-name/package-name e.g., npm/babel
+community_bridge: # Replace with a single Community Bridge project-name e.g., cloud-foundry
+liberapay: # Replace with a single Liberapay username
+issuehunt: # Replace with a single IssueHunt username
+lfx_crowdfunding: # Replace with a single LFX Crowdfunding project-name e.g., cloud-foundry
+polar: # Replace with a single Polar username
+buy_me_a_coffee: # Replace with a single Buy Me a Coffee username
+thanks_dev: # Replace with a single thanks.dev username
+custom: # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2']

.github/workflows/test_pip.yml

@@ -1,7 +1,9 @@
-name: pip install
-run-name: pip install
+name: CI
+run-name: CI
+
 on:
   push:
+    branches: [main]
     paths:
       - '**.py'
       - 'pyproject.toml'
@@ -11,23 +13,41 @@ on:
       - 'pyproject.toml'
       - '.github/workflows/**'
   workflow_dispatch:
+
 jobs:
-  install-and-run:
+  test:
     strategy:
       matrix:
-        os: [ubuntu-latest, macos-latest, windows-latest]
+        os: [ubuntu-latest, macos-14, windows-latest]
         python-version: ['3.12']
       fail-fast: false
-    continue-on-error: true
     runs-on: ${{ matrix.os }}
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v4
+        uses: actions/checkout@v7
+
       - name: Set up Python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: ${{ matrix.python-version }}
-      - name: Install from repository
-        run: python -m pip install .
-      #- name: Run abogen
-      #  run: abogen
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v8.3.1
+        with:
+          enable-cache: true
+          prune-cache: false
+          cache-dependency-glob: pyproject.toml
+
+      - name: Install system dependencies (Ubuntu)
+        if: runner.os == 'Linux'
+        run: sudo apt-get update && sudo apt-get install -y libegl1
+
+      - name: Install dependencies
+        run: uv pip install --system .[dev]
+        env:
+          UV_LINK_MODE: copy
+
+      - name: Run tests
+        env:
+          QT_QPA_PLATFORM: offscreen
+        run: pytest tests/ -v --tb=short

.github/workflows/test_publish.yml

@@ -18,7 +18,7 @@ jobs:
   build:
     runs-on: ubuntu-latest
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v7
 
     - name: Login to Github Container Registry
       # Only if we need to push an image

.gitignore

@@ -38,3 +38,4 @@ dist/
 .old/
 test_assets/
 dev_notes/
+.claude/

abogen/VERSION

@@ -1 +1 @@
-1.3.0
+1.3.1

abogen/book_parser.py

@@ -915,7 +915,11 @@ class EpubParser(BaseBookParser):
 
             if slice_html.strip():
                 slice_soup = BeautifulSoup(slice_html, "html.parser")
-                for tag in slice_soup.find_all(["p", "div"]):
+
+                # Add line breaks after block-level elements to ensure pauses in speech
+                for tag in slice_soup.find_all(
+                    ["p", "div", "h1", "h2", "h3", "h4", "h5", "h6", "li", "blockquote"]
+                ):
                     tag.append("\n\n")
 
                 for ol in slice_soup.find_all("ol"):

abogen/constants.py

@@ -63,64 +63,6 @@ SUPPORTED_INPUT_FORMATS = [
 # 384 if self.lang_code in 'ab':
 SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION = list(LANGUAGE_DESCRIPTIONS.keys())
 
-# Voice and sample text constants
-VOICES_INTERNAL = [
-    "af_alloy",
-    "af_aoede",
-    "af_bella",
-    "af_heart",
-    "af_jessica",
-    "af_kore",
-    "af_nicole",
-    "af_nova",
-    "af_river",
-    "af_sarah",
-    "af_sky",
-    "am_adam",
-    "am_echo",
-    "am_eric",
-    "am_fenrir",
-    "am_liam",
-    "am_michael",
-    "am_onyx",
-    "am_puck",
-    "am_santa",
-    "bf_alice",
-    "bf_emma",
-    "bf_isabella",
-    "bf_lily",
-    "bm_daniel",
-    "bm_fable",
-    "bm_george",
-    "bm_lewis",
-    "ef_dora",
-    "em_alex",
-    "em_santa",
-    "ff_siwis",
-    "hf_alpha",
-    "hf_beta",
-    "hm_omega",
-    "hm_psi",
-    "if_sara",
-    "im_nicola",
-    "jf_alpha",
-    "jf_gongitsune",
-    "jf_nezumi",
-    "jf_tebukuro",
-    "jm_kumo",
-    "pf_dora",
-    "pm_alex",
-    "pm_santa",
-    "zf_xiaobei",
-    "zf_xiaoni",
-    "zf_xiaoxiao",
-    "zf_xiaoyi",
-    "zm_yunjian",
-    "zm_yunxi",
-    "zm_yunxia",
-    "zm_yunyang",
-]
-
 # Voice and sample text mapping
 SAMPLE_VOICE_TEXTS = {
     "a": "This is a sample of the selected voice.",

abogen/conversion.py

@@ -1,16 +0,0 @@
-"""Backwards-compatible re-export of conversion module.
-
-The PyQt-based implementation lives in abogen.pyqt.conversion.
-The web-based implementation is in abogen.webui.conversion_runner.
-"""
-
-from __future__ import annotations
-
-# Re-export PyQt conversion classes for backwards compatibility
-from abogen.pyqt.conversion import (  # noqa: F401
-    ConversionThread,
-    VoicePreviewThread,
-    PlayAudioThread,
-)
-
-__all__ = ["ConversionThread", "VoicePreviewThread", "PlayAudioThread"]

abogen/main.py

@@ -2,13 +2,14 @@
 
 from __future__ import annotations
 
-import atexit
 import os
 import platform
-import signal
-import sys
 
-from abogen.utils import load_config, prevent_sleep_end
+# Initialise global shutdown handling (atexit, signals, Qt) as early as possible.
+from abogen import shutdown  # noqa: F401
+shutdown.register_shutdown()
+
+from abogen.utils import load_config
 from abogen.webui.app import main as _run_web_ui
 
 # Configure Hugging Face Hub behaviour (mirrors legacy GUI defaults).
@@ -27,17 +28,6 @@ os.environ.setdefault("MIOPEN_CONV_PRECISE_ROCM_TUNING", "0")
 if platform.system() == "Darwin" and platform.processor() == "arm":
     os.environ.setdefault("PYTORCH_ENABLE_MPS_FALLBACK", "1")
 
-atexit.register(prevent_sleep_end)
-
-
-def _cleanup_sleep(signum, _frame):
-    prevent_sleep_end()
-    sys.exit(0)
-
-
-signal.signal(signal.SIGINT, _cleanup_sleep)
-signal.signal(signal.SIGTERM, _cleanup_sleep)
-
 
 def main() -> None:
     """Launch the Flask-based web UI."""

abogen/predownload_gui.py

@@ -21,7 +21,8 @@ from PyQt6.QtWidgets import (
 )
 from PyQt6.QtCore import QThread, pyqtSignal
 
-from abogen.constants import COLORS, VOICES_INTERNAL
+from abogen.constants import COLORS
+from abogen.tts_plugin.utils import get_voices
 from abogen.spacy_utils import SPACY_MODELS
 import abogen.hf_tracker
 
@@ -114,7 +115,7 @@ class PreDownloadWorker(QThread):
             self._voices_success = False
             return
 
-        voice_list = VOICES_INTERNAL
+        voice_list = get_voices("kokoro")
         for idx, voice in enumerate(voice_list, start=1):
             if self._cancelled:
                 self._voices_success = False
@@ -462,14 +463,14 @@ class PreDownloadDialog(QDialog):
         try:
             from huggingface_hub import try_to_load_from_cache
 
-            for voice in VOICES_INTERNAL:
+            for voice in get_voices("kokoro"):
                 if not try_to_load_from_cache(
                     repo_id="hexgrad/Kokoro-82M", filename=f"voices/{voice}.pt"
                 ):
                     missing.append(voice)
         except Exception:
             # If HF missing, report all as missing
-            return False, list(VOICES_INTERNAL)
+            return False, list(get_voices("kokoro"))
         return (len(missing) == 0), missing
 
     def _check_kokoro_model(self) -> bool:

abogen/pyqt/conversion.py

@@ -5,6 +5,7 @@ import hashlib  # For generating unique cache filenames
 from platformdirs import user_desktop_dir
 from PyQt6.QtCore import QThread, pyqtSignal, Qt, QTimer
 from PyQt6.QtWidgets import QCheckBox, QVBoxLayout, QDialog, QLabel, QDialogButtonBox
+import numpy as np
 import soundfile as sf
 from abogen.utils import (
     create_process,
@@ -42,6 +43,10 @@ from abogen.subtitle_utils import (
     get_sample_voice_text,
     sanitize_name_for_os,
     _CHAPTER_MARKER_SEARCH_PATTERN,
+    _VOICE_MARKER_PATTERN,
+    _VOICE_MARKER_SEARCH_PATTERN,
+    split_text_by_voice_markers,
+    validate_voice_name,
 )
 
 class CountdownDialog(QDialog):
@@ -255,8 +260,7 @@ class ConversionThread(QThread):
         output_folder,
         subtitle_mode,
         output_format,
-        np_module,
-        kpipeline_class,
+        backend,
         start_time,
         total_char_count,
         use_gpu=True,
@@ -266,8 +270,7 @@ class ConversionThread(QThread):
         super().__init__()
         self._chapter_options_event = threading.Event()
         self._timestamp_response_event = threading.Event()
-        self.np = np_module
-        self.KPipeline = kpipeline_class
+        self.backend = backend
         self.file_name = file_name
         self.lang_code = lang_code
         self.speed = speed
@@ -296,6 +299,31 @@ class ConversionThread(QThread):
         self.use_spacy_segmentation = True  # Default, will be overridden from GUI
         # Set split pattern based on language and subtitle mode
         self.split_pattern = self._get_split_pattern(lang_code, subtitle_mode)
+        self.voice_cache = {}  # Cache for loaded voices
+
+    def load_voice_cached(self, voice_name, tts):
+        """Load voice with caching to avoid reloading same voice.
+
+        Args:
+            voice_name: Voice name or formula string
+            tts: TTS pipeline instance
+
+        Returns:
+            Loaded voice tensor or voice name string
+        """
+        # Check cache first
+        if voice_name in self.voice_cache:
+            return self.voice_cache[voice_name]
+
+        # Load voice
+        if "*" in voice_name:
+            loaded_voice = get_new_voice(tts, voice_name, self.use_gpu)
+        else:
+            loaded_voice = voice_name
+
+        # Cache it
+        self.voice_cache[voice_name] = loaded_voice
+        return loaded_voice
 
     def _stream_audio_in_chunks(
         self, segments, process_func, progress_prefix="Processing"
@@ -461,19 +489,6 @@ class ConversionThread(QThread):
 
             self.log_updated.emit(("\nInitializing TTS pipeline...", "grey"))
 
-            # Set device based on use_gpu setting and platform
-            if self.use_gpu:
-                if platform.system() == "Darwin" and platform.processor() == "arm":
-                    device = "mps"  # Use MPS for Apple Silicon
-                else:
-                    device = "cuda"  # Use CUDA for other platforms
-            else:
-                device = "cpu"
-
-            tts = self.KPipeline(
-                lang_code=self.lang_code, repo_id="hexgrad/Kokoro-82M", device=device
-            )
-
             # Check if the input is a subtitle file or timestamp text file
             is_subtitle_file = False
             is_timestamp_text = False
@@ -509,7 +524,7 @@ class ConversionThread(QThread):
 
             # Process subtitle files separately
             if is_subtitle_file or is_timestamp_text:
-                self._process_subtitle_file(tts, base_path, is_timestamp_text)
+                self._process_subtitle_file(self.backend, base_path, is_timestamp_text)
                 return
 
             if self.is_direct_text:
@@ -524,6 +539,26 @@ class ConversionThread(QThread):
             # Clean up text using utility function
             text = clean_text(text)
 
+            # Apply word substitutions if enabled
+            if getattr(self, "word_substitutions_enabled", False):
+                from abogen.word_substitution import apply_word_substitutions
+
+                self.log_updated.emit("Applying word substitutions...")
+
+                substitutions_list = getattr(self, "word_substitutions_list", "")
+                case_sensitive = getattr(self, "case_sensitive_substitutions", False)
+                replace_caps = getattr(self, "replace_all_caps", False)
+                replace_nums = getattr(self, "replace_numerals", False)
+                fix_punct = getattr(self, "fix_nonstandard_punctuation", False)
+
+                text = apply_word_substitutions(
+                    text,
+                    substitutions_list,
+                    case_sensitive,
+                    replace_caps,
+                    replace_nums,
+                    fix_punct,
+                )
 
             # --- Chapter splitting logic ---
             # Use pre-compiled pattern for better performance
@@ -550,6 +585,42 @@ class ConversionThread(QThread):
                 chapters = [("text", text)]
             total_chapters = len(chapters)
 
+            # --- Voice marker splitting logic ---
+            # Split each chapter by voice markers, preserving voice state across chapters
+            chapters_with_voices = []
+            current_voice = self.voice  # Start with default voice
+            total_valid_markers = 0
+            total_invalid_markers = 0
+
+            for chapter_name, chapter_text in chapters:
+                # Use current_voice as the starting voice for this chapter
+                voice_segments, last_voice, valid_count, invalid_count = split_text_by_voice_markers(chapter_text, current_voice)
+                chapters_with_voices.append((chapter_name, voice_segments))
+
+                # Update current_voice so next chapter continues with this voice
+                current_voice = last_voice
+
+                # Track total valid/invalid markers
+                total_valid_markers += valid_count
+                total_invalid_markers += invalid_count
+
+            # Log voice marker information with accurate counts
+            total_markers = total_valid_markers + total_invalid_markers
+            if total_markers > 0:
+                if total_invalid_markers == 0:
+                    # All markers were valid
+                    self.log_updated.emit(
+                        (f"\nDetected {total_markers} voice marker(s) - all valid", "grey")
+                    )
+                else:
+                    # Some markers were invalid
+                    self.log_updated.emit(
+                        (f"\nDetected {total_markers} voice marker(s) - {total_valid_markers} valid, {total_invalid_markers} invalid (using previous voice)", "orange")
+                    )
+
+            # Replace chapters with the new structure
+            chapters = chapters_with_voices
+
             # For text files with chapters, prompt user for options if not already set
             is_txt_file = not self.is_direct_text and (
                 self.file_name.lower().endswith(".txt")
@@ -842,7 +913,7 @@ class ConversionThread(QThread):
                 ]
                 srt_index = 1  # SRT numbering fix for chapter-only mode
             # Instead of processing the whole text, process by chapter
-            for chapter_idx, (chapter_name, chapter_text) in enumerate(chapters, 1):
+            for chapter_idx, (chapter_name, voice_segments) in enumerate(chapters, 1):
                 chapter_out_path = None
                 chapter_out_file = None
                 chapter_ffmpeg_proc = None
@@ -862,11 +933,6 @@ class ConversionThread(QThread):
                 if merge_chapters_at_end:
                     chapter_time["start"] = current_time
 
-                # Check if the voice is a formula and load it if necessary
-                if "*" in self.voice:
-                    loaded_voice = get_new_voice(tts, self.voice, self.use_gpu)
-                else:
-                    loaded_voice = self.voice
                 # Prepare per-chapter output file if needed
                 if save_chapters_separately and total_chapters > 1:
                     # First pass: keep alphanumeric, spaces, hyphens, and underscores
@@ -986,6 +1052,22 @@ class ConversionThread(QThread):
                     chapter_subtitle_path = None
                     chapter_subtitle_file = None
 
+
+                # Process each voice segment within the chapter
+                for segment_idx, (voice_name, segment_text) in enumerate(voice_segments):
+                    # Load voice for this segment (with caching)
+                    try:
+                        loaded_voice = self.load_voice_cached(voice_name, self.backend)
+                        if segment_idx > 0:
+                            voice_display = voice_name if len(voice_name) < 50 else voice_name[:47] + "..."
+                            self.log_updated.emit((f"  → Voice: {voice_display}", "grey"))
+                    except Exception:
+                        self.log_updated.emit(
+                            (f"⚠ Voice loading error for '{voice_name}', continuing with previous", "orange")
+                        )
+                        if segment_idx == 0:
+                            loaded_voice = self.load_voice_cached(self.voice, self.backend)
+
                     # Determine if spaCy segmentation should be used for PRE-TTS segmentation
                     # Only non-English languages use spaCy for pre-segmentation
                     # English uses spaCy only for subtitle generation (post-TTS)
@@ -1034,7 +1116,7 @@ class ConversionThread(QThread):
                         from abogen.spacy_utils import segment_sentences
 
                         spacy_sentences = segment_sentences(
-                        chapter_text,
+                            segment_text,
                             self.lang_code,
                             log_callback=lambda msg: self.log_updated.emit(msg),
                         )
@@ -1060,7 +1142,7 @@ class ConversionThread(QThread):
                             )
 
                     # Process text - either as spaCy sentences or as single text
-                text_segments = spacy_sentences if spacy_sentences else [chapter_text]
+                    text_segments = spacy_sentences if spacy_sentences else [segment_text]
 
                     # Print active split pattern used by the TTS engine once for this batch
                     try:
@@ -1070,7 +1152,7 @@ class ConversionThread(QThread):
                         print("Using split pattern: (unprintable)")
 
                     for text_segment in text_segments:
-                    for result in tts(
+                        for result in self.backend(
                             text_segment,
                             voice=loaded_voice,
                             speed=self.speed,
@@ -1272,7 +1354,7 @@ class ConversionThread(QThread):
                     silence_samples = int(
                         self.silence_duration * 24000
                     )  # Silence duration at 24,000 Hz
-                    silence_audio = self.np.zeros(silence_samples, dtype="float32")
+                    silence_audio = np.zeros(silence_samples, dtype="float32")
                     silence_bytes = silence_audio.tobytes()
 
                     if merged_out_file:
@@ -1611,7 +1693,7 @@ class ConversionThread(QThread):
             max_end_time = max(
                 (end for _, end, _ in subtitles if end is not None), default=0
             )
-            audio_buffer = self.np.zeros(
+            audio_buffer = np.zeros(
                 int(max_end_time * rate) + rate, dtype="float32"
             )
 
@@ -1675,7 +1757,7 @@ class ConversionThread(QThread):
                 # Generate TTS audio
                 tts_results = [
                     r
-                    for r in tts(
+                    for r in self.backend(
                         processed_text,
                         voice=loaded_voice,
                         speed=self.speed,
@@ -1693,11 +1775,11 @@ class ConversionThread(QThread):
 
                 # Concatenate audio and determine duration
                 full_audio = (
-                    self.np.concatenate(
+                    np.concatenate(
                         [a.numpy() if hasattr(a, "numpy") else a for a in audio_chunks]
                     )
                     if audio_chunks
-                    else self.np.zeros(
+                    else np.zeros(
                         int((subtitle_duration or 0) * rate), dtype="float32"
                     )
                 )
@@ -1731,8 +1813,8 @@ class ConversionThread(QThread):
                         num_stages = max(
                             1,
                             int(
-                                self.np.ceil(
-                                    self.np.log(speed_factor) / self.np.log(2.0)
+                                np.ceil(
+                                    np.log(speed_factor) / np.log(2.0)
                                 )
                             ),
                         )
@@ -1765,7 +1847,7 @@ class ConversionThread(QThread):
                             stdout=subprocess.PIPE,
                             stderr=subprocess.PIPE,
                         )
-                        full_audio = self.np.frombuffer(
+                        full_audio = np.frombuffer(
                             speed_proc.communicate(input=full_audio.tobytes())[0],
                             dtype="float32",
                         )
@@ -1779,7 +1861,7 @@ class ConversionThread(QThread):
 
                         tts_results = [
                             r
-                            for r in tts(
+                            for r in self.backend(
                                 processed_text,
                                 voice=loaded_voice,
                                 speed=new_speed,
@@ -1790,14 +1872,14 @@ class ConversionThread(QThread):
                         audio_chunks = [r.audio for r in tts_results]
 
                         full_audio = (
-                            self.np.concatenate(
+                            np.concatenate(
                                 [
                                     a.numpy() if hasattr(a, "numpy") else a
                                     for a in audio_chunks
                                 ]
                             )
                             if audio_chunks
-                            else self.np.zeros(
+                            else np.zeros(
                                 int(subtitle_duration * rate), dtype="float32"
                             )
                         )
@@ -1814,10 +1896,10 @@ class ConversionThread(QThread):
                 # Pad or trim to subtitle duration
                 target_samples = int(subtitle_duration * rate)
                 if len(full_audio) < target_samples:
-                    full_audio = self.np.concatenate(
+                    full_audio = np.concatenate(
                         [
                             full_audio,
-                            self.np.zeros(
+                            np.zeros(
                                 target_samples - len(full_audio), dtype="float32"
                             ),
                         ]
@@ -1830,10 +1912,10 @@ class ConversionThread(QThread):
                 end_sample = start_sample + len(full_audio)
                 if end_sample > len(audio_buffer):
                     # Extend buffer if needed
-                    audio_buffer = self.np.concatenate(
+                    audio_buffer = np.concatenate(
                         [
                             audio_buffer,
-                            self.np.zeros(
+                            np.zeros(
                                 end_sample - len(audio_buffer), dtype="float32"
                             ),
                         ]
@@ -1875,7 +1957,7 @@ class ConversionThread(QThread):
                 self.progress_updated.emit(percent, etr_str)
 
             # Normalize audio buffer to prevent clipping from mixed overlaps
-            max_amplitude = self.np.abs(audio_buffer).max()
+            max_amplitude = np.abs(audio_buffer).max()
             if max_amplitude > 1.0:
                 self.log_updated.emit(
                     f"\n  -> Normalizing audio (peak: {max_amplitude:.2f})"
@@ -2315,6 +2397,10 @@ class ConversionThread(QThread):
         self.cancel_requested = True
         self.should_cancel = True
         self.waiting_for_user_input = False
+        # Clear voice cache (instance and module-level)
+        self.voice_cache.clear()
+        from abogen.voice_cache import clear_voice_cache
+        clear_voice_cache()
         # Terminate subprocess if running
         if self.process:
             try:
@@ -2344,8 +2430,7 @@ class VoicePreviewThread(QThread):
 
     def __init__(
         self,
-        np_module,
-        kpipeline_class,
+        backend,
         lang_code,
         voice,
         speed,
@@ -2353,8 +2438,7 @@ class VoicePreviewThread(QThread):
         parent=None,
     ):
         super().__init__(parent)
-        self.np_module = np_module
-        self.kpipeline_class = kpipeline_class
+        self.backend = backend
         self.lang_code = lang_code
         self.voice = voice
         self.speed = speed
@@ -2388,31 +2472,19 @@ class VoicePreviewThread(QThread):
         # Generate the preview and save to cache
         try:
 
-            # Set device based on use_gpu setting and platform
-            if self.use_gpu:
-                if platform.system() == "Darwin" and platform.processor() == "arm":
-                    device = "mps"  # Use MPS for Apple Silicon
-                else:
-                    device = "cuda"  # Use CUDA for other platforms
-            else:
-                device = "cpu"
-
-            tts = self.kpipeline_class(
-                lang_code=self.lang_code, repo_id="hexgrad/Kokoro-82M", device=device
-            )
             # Enable voice formula support for preview
             if "*" in self.voice:
-                loaded_voice = get_new_voice(tts, self.voice, self.use_gpu)
+                loaded_voice = get_new_voice(self.backend, self.voice, self.use_gpu)
             else:
                 loaded_voice = self.voice
             sample_text = get_sample_voice_text(self.lang_code)
             audio_segments = []
-            for result in tts(
+            for result in self.backend(
                 sample_text, voice=loaded_voice, speed=self.speed, split_pattern=None
             ):
                 audio_segments.append(result.audio)
             if audio_segments:
-                audio = self.np_module.concatenate(audio_segments)
+                audio = np.concatenate(audio_segments)
                 # Save directly to the cache path
                 sf.write(self.cache_path, audio, 24000)
                 self.temp_wav = self.cache_path

abogen/pyqt/gui.py

@@ -74,7 +74,7 @@ from abogen.subtitle_utils import (
     calculate_text_length,
 )
 
-from abogen.conversion import ConversionThread, VoicePreviewThread, PlayAudioThread
+from abogen.pyqt.conversion import ConversionThread, VoicePreviewThread, PlayAudioThread, ChapterOptionsDialog, TimestampDetectionDialog
 from abogen.pyqt.book_handler import HandlerDialog
 from abogen.constants import (
     PROGRAM_NAME,
@@ -82,11 +82,11 @@ from abogen.constants import (
     GITHUB_URL,
     PROGRAM_DESCRIPTION,
     LANGUAGE_DESCRIPTIONS,
-    VOICES_INTERNAL,
     SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION,
     COLORS,
     SUBTITLE_FORMATS,
 )
+from abogen.tts_plugin.utils import get_voices
 import threading
 from abogen.pyqt.voice_formula_gui import VoiceFormulaDialog
 from abogen.voice_profiles import load_profiles
@@ -665,6 +665,11 @@ class TextboxDialog(QDialog):
         self.insert_chapter_btn.clicked.connect(self.insert_chapter_marker)
         button_layout.addWidget(self.insert_chapter_btn)
 
+        self.insert_voice_btn = QPushButton("Insert Voice Marker", self)
+        self.insert_voice_btn.setToolTip("Insert a voice change marker at the cursor position")
+        self.insert_voice_btn.clicked.connect(self.insert_voice_marker)
+        button_layout.addWidget(self.insert_voice_btn)
+
         self.cancel_button = QPushButton("Cancel", self)
         self.cancel_button.clicked.connect(self.reject)
 
@@ -767,6 +772,23 @@ class TextboxDialog(QDialog):
         self.update_char_count()
         self.text_edit.setFocus()
 
+    def insert_voice_marker(self):
+        """Insert a voice marker template at cursor position."""
+        cursor = self.text_edit.textCursor()
+        # Use the currently selected voice as the default
+        try:
+            parent_window = self.parent()
+            if parent_window and hasattr(parent_window, 'selected_voice'):
+                default_voice = parent_window.selected_voice or "af_heart"
+            else:
+                default_voice = "af_heart"
+        except Exception:
+            default_voice = "af_heart"
+        cursor.insertText(f"\n<<VOICE:{default_voice}>>\n")
+        self.text_edit.setTextCursor(cursor)
+        self.update_char_count()
+        self.text_edit.setFocus()
+
 
 def migrate_subtitle_format(config):
     """Convert old subtitle_format values to new internal keys."""
@@ -783,6 +805,108 @@ def migrate_subtitle_format(config):
         save_config(config)
 
 
+class WordSubstitutionsDialog(QDialog):
+    """Dialog for configuring word substitutions and text preprocessing options."""
+
+    def __init__(
+        self,
+        parent=None,
+        initial_list="",
+        initial_case_sensitive=False,
+        initial_caps=False,
+        initial_numerals=False,
+        initial_punctuation=False,
+    ):
+        super().__init__(parent)
+        self.setWindowTitle("Word Substitutions Settings")
+        self.setWindowFlags(
+            Qt.WindowType.Window
+            | Qt.WindowType.WindowCloseButtonHint
+            | Qt.WindowType.WindowMaximizeButtonHint
+        )
+        self.resize(600, 500)
+
+        layout = QVBoxLayout(self)
+
+        # Instructions
+        instructions = QLabel(
+            "Enter word substitutions (one per line) in format: Word|NewWord\n"
+            "  - If nothing after |, the word will be erased completely\n"
+            "  - Substitutions match whole words only (e.g., \"tree\" won't match \"trees\" but will match \"tree's\")\n"
+            "  - By default, matching is case-insensitive (e.g., \"gonna\" matches \"Gonna\", \"GONNA\", etc.)",
+            self,
+        )
+        instructions.setStyleSheet(
+            "padding: 10px; background-color: #f0f0f0; border-radius: 5px;"
+        )
+        instructions.setWordWrap(True)
+        layout.addWidget(instructions)
+
+        # Text edit area
+        self.text_edit = QTextEdit(self)
+        self.text_edit.setAcceptRichText(False)
+        self.text_edit.setPlaceholderText("Word|NewWord")
+        self.text_edit.setPlainText(initial_list)
+        layout.addWidget(self.text_edit)
+
+        # Checkboxes
+        self.case_sensitive_checkbox = QCheckBox(
+            "Case-sensitive word matching", self
+        )
+        self.case_sensitive_checkbox.setChecked(initial_case_sensitive)
+        layout.addWidget(self.case_sensitive_checkbox)
+
+        self.caps_checkbox = QCheckBox("Replace ALL CAPS with lowercase", self)
+        self.caps_checkbox.setChecked(initial_caps)
+        layout.addWidget(self.caps_checkbox)
+
+        self.numerals_checkbox = QCheckBox(
+            "Replace Numerals with Words (e.g., 309 \u2192 three hundred and nine)", self
+        )
+        self.numerals_checkbox.setChecked(initial_numerals)
+        layout.addWidget(self.numerals_checkbox)
+
+        self.punctuation_checkbox = QCheckBox(
+            "Fix Nonstandard Punctuation (curly quotes and other Unicode punctuation that may affect how words sound)",
+            self,
+        )
+        self.punctuation_checkbox.setChecked(initial_punctuation)
+        layout.addWidget(self.punctuation_checkbox)
+
+        # Buttons
+        button_layout = QHBoxLayout()
+        self.cancel_button = QPushButton("Cancel", self)
+        self.cancel_button.clicked.connect(self.reject)
+        self.ok_button = QPushButton("OK", self)
+        self.ok_button.setDefault(True)
+        self.ok_button.clicked.connect(self.accept)
+
+        button_layout.addStretch()
+        button_layout.addWidget(self.cancel_button)
+        button_layout.addWidget(self.ok_button)
+        layout.addLayout(button_layout)
+
+    def get_substitutions_list(self):
+        """Get the substitutions list as plain text."""
+        return self.text_edit.toPlainText()
+
+    def get_case_sensitive(self):
+        """Get whether case-sensitive matching is enabled."""
+        return self.case_sensitive_checkbox.isChecked()
+
+    def get_replace_all_caps(self):
+        """Get whether ALL CAPS replacement is enabled."""
+        return self.caps_checkbox.isChecked()
+
+    def get_replace_numerals(self):
+        """Get whether numeral-to-word conversion is enabled."""
+        return self.numerals_checkbox.isChecked()
+
+    def get_fix_nonstandard_punctuation(self):
+        """Get whether nonstandard punctuation fixing is enabled."""
+        return self.punctuation_checkbox.isChecked()
+
+
 class abogen(QWidget):
     def __init__(self):
         super().__init__()
@@ -833,6 +957,19 @@ class abogen(QWidget):
         self.use_silent_gaps = self.config.get("use_silent_gaps", True)
         self.subtitle_speed_method = self.config.get("subtitle_speed_method", "tts")
         self.use_spacy_segmentation = self.config.get("use_spacy_segmentation", True)
+        # Word substitution settings
+        self.word_substitutions_enabled = self.config.get(
+            "word_substitutions_enabled", False
+        )
+        self.word_substitutions_list = self.config.get("word_substitutions_list", "")
+        self.case_sensitive_substitutions = self.config.get(
+            "case_sensitive_substitutions", False
+        )
+        self.replace_all_caps = self.config.get("replace_all_caps", False)
+        self.replace_numerals = self.config.get("replace_numerals", False)
+        self.fix_nonstandard_punctuation = self.config.get(
+            "fix_nonstandard_punctuation", False
+        )
         self._pending_close_event = None
         self.gpu_ok = False  # Initialize GPU availability status
 
@@ -1071,6 +1208,35 @@ class abogen(QWidget):
         subtitle_layout.addWidget(self.subtitle_combo)
         controls_layout.addLayout(subtitle_layout)
 
+        # Word Substitutions section
+        word_sub_layout = QHBoxLayout()
+        word_sub_layout.setSpacing(7)
+        word_sub_label = QLabel("Word Substitutions:", self)
+        word_sub_layout.addWidget(word_sub_label)
+
+        self.word_sub_combo = QComboBox(self)
+        self.word_sub_combo.addItems(["Disabled", "Enabled"])
+        self.word_sub_combo.setStyleSheet(
+            "QComboBox { min-height: 20px; padding: 6px 12px; }"
+        )
+        self.word_sub_combo.setSizePolicy(
+            QSizePolicy.Policy.Expanding, QSizePolicy.Policy.Fixed
+        )
+        self.word_sub_combo.setCurrentText(
+            "Enabled" if self.word_substitutions_enabled else "Disabled"
+        )
+        self.word_sub_combo.currentTextChanged.connect(self.on_word_sub_changed)
+        word_sub_layout.addWidget(self.word_sub_combo)
+
+        self.btn_word_sub_settings = QPushButton("Settings", self)
+        self.btn_word_sub_settings.setFixedSize(80, 36)
+        self.btn_word_sub_settings.setStyleSheet("QPushButton { padding: 6px 12px; }")
+        self.btn_word_sub_settings.clicked.connect(self.show_word_sub_dialog)
+        self.btn_word_sub_settings.setEnabled(self.word_substitutions_enabled)
+        word_sub_layout.addWidget(self.btn_word_sub_settings)
+
+        controls_layout.addLayout(word_sub_layout)
+
         # Output voice format
         format_layout = QHBoxLayout()
         format_layout.setSpacing(7)
@@ -1707,7 +1873,7 @@ class abogen(QWidget):
         for pname in load_profiles().keys():
             self.voice_combo.addItem(profile_icon, pname, f"profile:{pname}")
         # re-add voices
-        for v in VOICES_INTERNAL:
+        for v in get_voices("kokoro"):
             icon = QIcon()
             flag_path = get_resource_path("abogen.assets.flags", f"{v[0]}.png")
             if flag_path and os.path.exists(flag_path):
@@ -2015,6 +2181,21 @@ class abogen(QWidget):
                 self.subtitle_speed_method = getattr(
                     queued_item, "subtitle_speed_method", "tts"
                 )
+                # Word substitution settings
+                self.word_substitutions_enabled = getattr(
+                    queued_item, "word_substitutions_enabled", False
+                )
+                self.word_substitutions_list = getattr(
+                    queued_item, "word_substitutions_list", ""
+                )
+                self.case_sensitive_substitutions = getattr(
+                    queued_item, "case_sensitive_substitutions", False
+                )
+                self.replace_all_caps = getattr(queued_item, "replace_all_caps", False)
+                self.replace_numerals = getattr(queued_item, "replace_numerals", False)
+                self.fix_nonstandard_punctuation = getattr(
+                    queued_item, "fix_nonstandard_punctuation", False
+                )
 
                 # This ensures that if conversion.py (or utils) reads from config/disk
                 # instead of using passed arguments, it sees the correct queue values.
@@ -2023,6 +2204,13 @@ class abogen(QWidget):
                 self.config["selected_format"] = self.selected_format
                 self.config["use_silent_gaps"] = self.use_silent_gaps
                 self.config["subtitle_speed_method"] = self.subtitle_speed_method
+                # Word substitution settings
+                self.config["word_substitutions_enabled"] = self.word_substitutions_enabled
+                self.config["word_substitutions_list"] = self.word_substitutions_list
+                self.config["case_sensitive_substitutions"] = self.case_sensitive_substitutions
+                self.config["replace_all_caps"] = self.replace_all_caps
+                self.config["replace_numerals"] = self.replace_numerals
+                self.config["fix_nonstandard_punctuation"] = self.fix_nonstandard_punctuation
 
                 # Sync Voice/Profile in config
                 self.config["selected_voice"] = self.selected_voice
@@ -2128,9 +2316,9 @@ class abogen(QWidget):
             file_size_str = "Unknown"
 
         # pipeline_loaded_callback remains unchanged
-        def pipeline_loaded_callback(np_module, kpipeline_class, error):
+        def pipeline_loaded_callback(backend, error):
             if error:
-                self.update_log((f"Error loading numpy or KPipeline: {error}", "red"))
+                self.update_log((f"Error loading TTS backend: {error}", "red"))
                 prevent_sleep_end()
                 return
 
@@ -2153,8 +2341,7 @@ class abogen(QWidget):
                 self.selected_output_folder,
                 subtitle_mode=actual_subtitle_mode,
                 output_format=self.selected_format,
-                np_module=np_module,
-                kpipeline_class=kpipeline_class,
+                backend=backend,
                 start_time=self.start_time,
                 total_char_count=self.char_count,
                 use_gpu=self.gpu_ok,
@@ -2179,6 +2366,21 @@ class abogen(QWidget):
             self.conversion_thread.subtitle_speed_method = self.subtitle_speed_method
             # Pass use_spacy_segmentation setting
             self.conversion_thread.use_spacy_segmentation = self.use_spacy_segmentation
+            # Pass word substitution settings
+            self.conversion_thread.word_substitutions_enabled = (
+                self.word_substitutions_enabled
+            )
+            self.conversion_thread.word_substitutions_list = (
+                self.word_substitutions_list
+            )
+            self.conversion_thread.case_sensitive_substitutions = (
+                self.case_sensitive_substitutions
+            )
+            self.conversion_thread.replace_all_caps = self.replace_all_caps
+            self.conversion_thread.replace_numerals = self.replace_numerals
+            self.conversion_thread.fix_nonstandard_punctuation = (
+                self.fix_nonstandard_punctuation
+            )
             # Pass separate_chapters_format setting
             self.conversion_thread.separate_chapters_format = (
                 self.separate_chapters_format
@@ -2223,7 +2425,20 @@ class abogen(QWidget):
             self.gpu_ok = gpu_ok
             self.update_log((gpu_msg, gpu_ok))
             self.update_log("Loading modules...")
-            load_thread = LoadPipelineThread(pipeline_loaded_callback)
+
+            # Determine device based on GPU availability
+            if gpu_ok:
+                if platform.system() == "Darwin" and platform.processor() == "arm":
+                    device = "mps"
+                else:
+                    device = "cuda"
+            else:
+                device = "cpu"
+
+            lang_code = self.selected_lang or "a"
+            load_thread = LoadPipelineThread(
+                pipeline_loaded_callback, lang_code=lang_code, device=device
+            )
             load_thread.start()
 
         threading.Thread(target=gpu_and_load, daemon=True).start()
@@ -2660,18 +2875,27 @@ class abogen(QWidget):
             )
             self.loading_movie.start()
 
-        def pipeline_loaded_callback(np_module, kpipeline_class, error):
-            self._on_pipeline_loaded_for_preview(np_module, kpipeline_class, error)
+        # Determine device based on GPU availability
+        if self.gpu_ok:
+            if platform.system() == "Darwin" and platform.processor() == "arm":
+                device = "mps"
+            else:
+                device = "cuda"
+        else:
+            device = "cpu"
 
-        load_thread = LoadPipelineThread(pipeline_loaded_callback)
+        lang = self.selected_lang or "a"
+        load_thread = LoadPipelineThread(
+            self._on_pipeline_loaded_for_preview, lang_code=lang, device=device
+        )
         load_thread.start()
 
-    def _on_pipeline_loaded_for_preview(self, np_module, kpipeline_class, error):
+    def _on_pipeline_loaded_for_preview(self, backend, error):
         # stop loading animation and restore icon on error
         if error:
             self.loading_movie.stop()
             self._show_error_message_box(
-                "Loading Error", f"Error loading numpy or KPipeline: {error}"
+                "Loading Error", f"Error loading TTS backend: {error}"
             )
             self.btn_preview.setIcon(self.play_icon)
             self.btn_preview.setEnabled(True)
@@ -2709,7 +2933,7 @@ class abogen(QWidget):
         gpu_msg, gpu_ok = get_gpu_acceleration(self.use_gpu)
 
         self.preview_thread = VoicePreviewThread(
-            np_module, kpipeline_class, lang, voice, speed, gpu_ok
+            backend, lang, voice, speed, gpu_ok
         )
         self.preview_thread.finished.connect(self._play_preview_audio)
         self.preview_thread.error.connect(self._preview_error)
@@ -2927,6 +3151,41 @@ class abogen(QWidget):
         self.config["use_gpu"] = self.use_gpu
         save_config(self.config)
 
+    def on_word_sub_changed(self, text):
+        """Handle word substitution dropdown change."""
+        self.word_substitutions_enabled = text == "Enabled"
+        self.btn_word_sub_settings.setEnabled(self.word_substitutions_enabled)
+
+        # Save to config
+        self.config["word_substitutions_enabled"] = self.word_substitutions_enabled
+        save_config(self.config)
+
+    def show_word_sub_dialog(self):
+        """Show word substitutions settings dialog."""
+        dialog = WordSubstitutionsDialog(
+            self,
+            initial_list=self.word_substitutions_list,
+            initial_case_sensitive=self.case_sensitive_substitutions,
+            initial_caps=self.replace_all_caps,
+            initial_numerals=self.replace_numerals,
+            initial_punctuation=self.fix_nonstandard_punctuation,
+        )
+
+        if dialog.exec() == QDialog.DialogCode.Accepted:
+            self.word_substitutions_list = dialog.get_substitutions_list()
+            self.case_sensitive_substitutions = dialog.get_case_sensitive()
+            self.replace_all_caps = dialog.get_replace_all_caps()
+            self.replace_numerals = dialog.get_replace_numerals()
+            self.fix_nonstandard_punctuation = dialog.get_fix_nonstandard_punctuation()
+
+            # Save all settings to config
+            self.config["word_substitutions_list"] = self.word_substitutions_list
+            self.config["case_sensitive_substitutions"] = self.case_sensitive_substitutions
+            self.config["replace_all_caps"] = self.replace_all_caps
+            self.config["replace_numerals"] = self.replace_numerals
+            self.config["fix_nonstandard_punctuation"] = self.fix_nonstandard_punctuation
+            save_config(self.config)
+
     def cleanup_conversion_thread(self):
         # Stop conversion thread
         if (
@@ -2977,12 +3236,16 @@ class abogen(QWidget):
             )
             box.setDefaultButton(QMessageBox.StandardButton.No)
             if box.exec() == QMessageBox.StandardButton.Yes:
+                from abogen import shutdown
+                shutdown.request_shutdown()
                 self.cleanup_conversion_thread()
                 self.cleanup_preview_threads()
                 event.accept()
             else:
                 event.ignore()
         else:
+            from abogen import shutdown
+            shutdown.request_shutdown()
             self.cleanup_conversion_thread()
             self.cleanup_preview_threads()
             event.accept()
@@ -2991,8 +3254,6 @@ class abogen(QWidget):
         """Show dialog to ask user about chapter processing options when chapters are detected in a .txt file"""
         # Check if this is a timestamp detection (-1) or chapter detection
         if chapter_count == -1:
-            from abogen.conversion import TimestampDetectionDialog
-
             dialog = TimestampDetectionDialog(parent=self)
             dialog.setWindowModality(Qt.WindowModality.ApplicationModal)
 
@@ -3007,8 +3268,6 @@ class abogen(QWidget):
             return
 
         # Normal chapter detection
-        from abogen.conversion import ChapterOptionsDialog
-
         dialog = ChapterOptionsDialog(chapter_count, parent=self)
         dialog.setWindowModality(Qt.WindowModality.ApplicationModal)
 

abogen/pyqt/main.py

@@ -1,10 +1,10 @@
 import os
 import sys
 import platform
-import atexit
-import signal
-from abogen.utils import get_resource_path, load_config, prevent_sleep_end
 
+# Initialise global shutdown handling (atexit, signals, Qt) as early as possible.
+from abogen import shutdown  # noqa: F401
+shutdown.register_shutdown()
 
 # Fix PyTorch DLL loading issue ([WinError 1114]) on Windows before importing PyQt6
 if platform.system() == "Windows":
@@ -94,6 +94,7 @@ os.environ["HF_HUB_DISABLE_TELEMETRY"] = "1"  # Disable Hugging Face telemetry
 os.environ["HF_HUB_ETAG_TIMEOUT"] = "10"  # Metadata request timeout (seconds)
 os.environ["HF_HUB_DOWNLOAD_TIMEOUT"] = "10"  # File download timeout (seconds)
 os.environ["HF_HUB_DISABLE_SYMLINKS_WARNING"] = "1"  # Disable symlinks warning
+from abogen.utils import load_config
 if load_config().get("disable_kokoro_internet", False):
     print("INFO: Kokoro's internet access is disabled.")
     os.environ["HF_HUB_OFFLINE"] = "1"  # Disable Hugging Face Hub internet access
@@ -105,25 +106,6 @@ from abogen.constants import PROGRAM_NAME, VERSION
 os.environ["MIOPEN_FIND_MODE"] = "FAST"
 os.environ["MIOPEN_CONV_PRECISE_ROCM_TUNING"] = "0"
 
-# Reset sleep states
-atexit.register(prevent_sleep_end)
-
-
-# Also handle signals (Ctrl+C, kill, etc.)
-def _cleanup_sleep(signum, frame):
-    prevent_sleep_end()
-    sys.exit(0)
-
-
-signal.signal(signal.SIGINT, _cleanup_sleep)
-signal.signal(signal.SIGTERM, _cleanup_sleep)
-
-# Ensure sys.stdout and sys.stderr are valid in GUI mode
-if sys.stdout is None:
-    sys.stdout = open(os.devnull, "w")
-if sys.stderr is None:
-    sys.stderr = open(os.devnull, "w")
-
 # Enable MPS GPU acceleration on Mac Apple Silicon
 if platform.system() == "Darwin" and platform.processor() == "arm":
     os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"

abogen/pyqt/predownload_gui.py

@@ -21,7 +21,8 @@ from PyQt6.QtWidgets import (
 )
 from PyQt6.QtCore import QThread, pyqtSignal
 
-from abogen.constants import COLORS, VOICES_INTERNAL
+from abogen.constants import COLORS
+from abogen.tts_plugin.utils import get_voices
 from abogen.spacy_utils import SPACY_MODELS
 import abogen.hf_tracker
 
@@ -114,7 +115,7 @@ class PreDownloadWorker(QThread):
             self._voices_success = False
             return
 
-        voice_list = VOICES_INTERNAL
+        voice_list = get_voices("kokoro")
         for idx, voice in enumerate(voice_list, start=1):
             if self._cancelled:
                 self._voices_success = False
@@ -462,14 +463,14 @@ class PreDownloadDialog(QDialog):
         try:
             from huggingface_hub import try_to_load_from_cache
 
-            for voice in VOICES_INTERNAL:
+            for voice in get_voices("kokoro"):
                 if not try_to_load_from_cache(
                     repo_id="hexgrad/Kokoro-82M", filename=f"voices/{voice}.pt"
                 ):
                     missing.append(voice)
         except Exception:
             # If HF missing, report all as missing
-            return False, list(VOICES_INTERNAL)
+            return False, list(get_voices("kokoro"))
         return (len(missing) == 0), missing
 
     def _check_kokoro_model(self) -> bool:

abogen/pyqt/queue_manager_gui.py

@@ -35,6 +35,12 @@ OVERRIDE_FIELDS = [
     "replace_single_newlines",
     "use_silent_gaps",
     "subtitle_speed_method",
+    "word_substitutions_enabled",
+    "word_substitutions_list",
+    "case_sensitive_substitutions",
+    "replace_all_caps",
+    "replace_numerals",
+    "fix_nonstandard_punctuation",
 ]
 
 
@@ -474,6 +480,21 @@ class QueueManager(QDialog):
             attrs["subtitle_speed_method"] = getattr(
                 parent, "subtitle_speed_method", "tts"
             )
+            # word substitutions
+            attrs["word_substitutions_enabled"] = getattr(
+                parent, "word_substitutions_enabled", False
+            )
+            attrs["word_substitutions_list"] = getattr(
+                parent, "word_substitutions_list", ""
+            )
+            attrs["case_sensitive_substitutions"] = getattr(
+                parent, "case_sensitive_substitutions", False
+            )
+            attrs["replace_all_caps"] = getattr(parent, "replace_all_caps", False)
+            attrs["replace_numerals"] = getattr(parent, "replace_numerals", False)
+            attrs["fix_nonstandard_punctuation"] = getattr(
+                parent, "fix_nonstandard_punctuation", False
+            )
             # book handler options
             attrs["save_chapters_separately"] = getattr(
                 parent, "save_chapters_separately", None

abogen/pyqt/queued_item.py

@@ -19,3 +19,10 @@ class QueuedItem:
     save_base_path: str = None
     save_chapters_separately: bool = None
     merge_chapters_at_end: bool = None
+    # Word Substitution fields
+    word_substitutions_enabled: bool = False
+    word_substitutions_list: str = ""
+    case_sensitive_substitutions: bool = False
+    replace_all_caps: bool = False
+    replace_numerals: bool = False
+    fix_nonstandard_punctuation: bool = False

abogen/pyqt/voice_formula_gui.py

@@ -28,11 +28,11 @@ from PyQt6.QtWidgets import (
 from PyQt6.QtCore import Qt, QTimer, QPoint, QRect, QSize
 from PyQt6.QtGui import QPixmap, QIcon, QAction
 from abogen.constants import (
-    VOICES_INTERNAL,
     SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION,
     LANGUAGE_DESCRIPTIONS,
     COLORS,
 )
+from abogen.tts_plugin.utils import get_voices
 import re
 import platform
 from abogen.utils import get_resource_path
@@ -179,7 +179,7 @@ class VoiceMixer(QWidget):
         layout.addWidget(QLabel(name), alignment=Qt.AlignmentFlag.AlignCenter)
 
         # Voice name label with gender icon
-        is_female = self.voice_name in VOICES_INTERNAL and self.voice_name[1] == "f"
+        is_female = self.voice_name in get_voices("kokoro") and self.voice_name[1] == "f"
 
         # Icons layout (flag and gender)
         icons_layout = QHBoxLayout()
@@ -772,7 +772,7 @@ class VoiceFormulaDialog(QDialog):
 
     def add_voices(self, initial_state):
         first_enabled_voice = None
-        for voice in VOICES_INTERNAL:
+        for voice in get_voices("kokoro"):
             language_code = voice[0]  # First character is the language code
             matching_voice = next(
                 (item for item in initial_state if item[0] == voice), None

abogen/shutdown.py

@@ -0,0 +1,160 @@
+"""Graceful shutdown - single module, no over-engineering."""
+from __future__ import annotations
+
+import atexit
+import gc
+import signal
+import sys
+from typing import Callable
+
+_CLEANUP_FUNCS: list[Callable[[], None]] = []
+_EXECUTED = False
+
+
+def register_cleanup(fn: Callable[[], None]) -> None:
+    """Register a cleanup function to run on shutdown."""
+    _CLEANUP_FUNCS.append(fn)
+
+
+def _run_cleanups() -> None:
+    global _EXECUTED
+    if _EXECUTED:
+        return
+    _EXECUTED = True
+    for fn in _CLEANUP_FUNCS:
+        try:
+            fn()
+        except Exception:
+            pass
+
+
+# ---- Register built-in cleanup functions ----
+
+# 1. Restore sleep prevention
+def _restore_sleep() -> None:
+    try:
+        from abogen.utils import prevent_sleep_end
+        prevent_sleep_end()
+    except Exception:
+        pass
+
+register_cleanup(_restore_sleep)
+
+# 2. Shutdown web UI ConversionService
+def _shutdown_conversion_service() -> None:
+    try:
+        from abogen.webui.service import get_service
+        svc = get_service()
+        if svc is not None:
+            svc.shutdown()
+    except Exception:
+        pass
+
+register_cleanup(_shutdown_conversion_service)
+
+# 3. Clear TTS pipelines and GPU memory
+def _cleanup_tts_pipelines() -> None:
+    # Clear web UI pipeline cache
+    try:
+        from abogen.webui.conversion_runner import _PIPELINES
+        _PIPELINES.clear()
+    except Exception:
+        pass
+
+    # Clear PyQt conversion thread voice cache
+    try:
+        from abogen.pyqt.conversion import ConversionThread
+        if hasattr(ConversionThread, "voice_cache"):
+            ConversionThread.voice_cache.clear()
+    except Exception:
+        pass
+
+    gc.collect()
+
+    # Release CUDA cache
+    try:
+        import torch
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+    except Exception:
+        pass
+
+register_cleanup(_cleanup_tts_pipelines)
+
+# 4. Clear global voice cache
+def _clear_voice_cache() -> None:
+    try:
+        from abogen.voice_cache import clear_voice_cache
+        clear_voice_cache()
+    except Exception:
+        pass
+
+register_cleanup(_clear_voice_cache)
+
+# 5. Terminate child processes (ffmpeg, etc.)
+def _terminate_subprocesses() -> None:
+    try:
+        import psutil
+    except Exception:
+        return
+
+    try:
+        current = psutil.Process()
+        for child in current.children(recursive=True):
+            try:
+                child.terminate()
+            except Exception:
+                pass
+        gone, alive = psutil.wait_procs(current.children(recursive=True), timeout=3)
+        for proc in alive:
+            try:
+                proc.kill()
+            except Exception:
+                pass
+    except Exception:
+        pass
+
+register_cleanup(_terminate_subprocesses)
+
+
+def register_shutdown() -> None:
+    """Install process-wide shutdown hooks (atexit, signals, Qt)."""
+    if register_shutdown._registered:
+        return
+    register_shutdown._registered = True
+
+    atexit.register(_run_cleanups)
+
+    # POSIX signals
+    for sig in (signal.SIGINT, signal.SIGTERM):
+        try:
+            signal.signal(sig, _on_signal)
+        except Exception:
+            pass
+
+    # Qt hook
+    try:
+        from PyQt6.QtWidgets import QApplication
+
+        app = QApplication.instance()
+        if app is not None:
+            app.aboutToQuit.connect(_run_cleanups)
+    except Exception:
+        pass
+
+
+register_shutdown._registered = False
+
+
+def _on_signal(signum: int, _frame) -> None:
+    _run_cleanups()
+    sys.exit(0)
+
+
+def request_shutdown() -> None:
+    """Programmatically trigger cleanup (e.g., from GUI closeEvent)."""
+    _run_cleanups()
+
+
+__all__ = ["register_shutdown", "request_shutdown", "register_cleanup"]

abogen/subtitle_utils.py

@@ -15,6 +15,8 @@ _ASS_STYLING_PATTERN = re.compile(r"\{[^}]+\}")
 _ASS_NEWLINE_N_PATTERN = re.compile(r"\\N")
 _ASS_NEWLINE_LOWER_N_PATTERN = re.compile(r"\\n")
 _CHAPTER_MARKER_SEARCH_PATTERN = re.compile(r"<<CHAPTER_MARKER:(.*?)>>")
+_VOICE_MARKER_PATTERN = re.compile(r"<<VOICE:[^>]*>>")
+_VOICE_MARKER_SEARCH_PATTERN = re.compile(r"<<VOICE:(.*?)>>")
 _WEBVTT_HEADER_PATTERN = re.compile(r"^WEBVTT.*?\n", re.MULTILINE)
 _VTT_STYLE_PATTERN = re.compile(r"STYLE\s*\n.*?(?=\n\n|$)", re.DOTALL)
 _VTT_NOTE_PATTERN = re.compile(r"NOTE\s*\n.*?(?=\n\n|$)", re.DOTALL)
@@ -31,17 +33,19 @@ _LINUX_ILLEGAL_CHARS_PATTERN = re.compile(r"[/\x00]")
 
 
 def clean_subtitle_text(text):
-    """Remove chapter markers and metadata tags from subtitle text."""
+    """Remove chapter markers, voice markers, and metadata tags from subtitle text."""
     # Use pre-compiled patterns for better performance
     text = _METADATA_TAG_PATTERN.sub("", text)
     text = _CHAPTER_MARKER_PATTERN.sub("", text)
+    text = _VOICE_MARKER_PATTERN.sub("", text)
     return text.strip()
 
 
 def calculate_text_length(text):
     # Use pre-compiled patterns for better performance
-    # Ignore chapter markers and metadata patterns in a single pass
+    # Ignore chapter markers, voice markers, and metadata patterns in a single pass
     text = _CHAPTER_MARKER_PATTERN.sub("", text)
+    text = _VOICE_MARKER_PATTERN.sub("", text)
     text = _METADATA_TAG_PATTERN.sub("", text)
     # Ignore newlines and leading/trailing spaces
     text = text.replace("\n", "").strip()
@@ -459,3 +463,122 @@ def sanitize_name_for_os(name, is_folder=True):
         sanitized = sanitized[:255].rstrip(". ")
 
     return sanitized
+
+
+def validate_voice_name(voice_name):
+    """Validate voice name against available voices (case-insensitive).
+    Handles both single voices and formulas like 'af_heart*0.5 + am_echo*0.5'.
+
+    Args:
+        voice_name: Voice name or formula string to validate
+
+    Returns:
+        Tuple of (is_valid, invalid_voice_name):
+            - is_valid: True if all voices in the name/formula are valid
+            - invalid_voice_name: The first invalid voice found, or None if all valid
+    """
+    from abogen.tts_plugin.utils import get_voices
+
+    # Create case-insensitive lookup set (done once per call)
+    voice_lookup_lower = {v.lower() for v in get_voices("kokoro")}
+    voice_name = voice_name.strip()
+
+    # Check if it's a formula (contains *)
+    if "*" in voice_name:
+        # Extract voice names from formula
+        voices = voice_name.split("+")
+        for term in voices:
+            if "*" in term:
+                base_voice = term.split("*")[0].strip()
+                # Case-insensitive comparison
+                if base_voice.lower() not in voice_lookup_lower:
+                    return False, base_voice
+        return True, None
+    else:
+        # Single voice - case-insensitive comparison
+        if voice_name.lower() not in voice_lookup_lower:
+            return False, voice_name
+        return True, None
+
+
+def split_text_by_voice_markers(text, default_voice):
+    """Split text by voice markers, returning list of (voice, text) tuples.
+
+    IMPORTANT: Returns the last voice used so it can persist across chapters.
+    Voice names are normalized to lowercase to match canonical voice names.
+
+    Args:
+        text: Text potentially containing <<VOICE:name>> markers
+        default_voice: Voice to use if no markers found or before first marker
+
+    Returns:
+        Tuple of (segments_list, last_voice_used, valid_count, invalid_count):
+            - segments_list: List of (voice_name, segment_text) tuples
+            - last_voice_used: The voice that should continue into next chapter
+            - valid_count: Number of valid voice markers processed
+            - invalid_count: Number of invalid voice markers skipped
+    """
+    from abogen.tts_plugin.utils import get_voices
+
+    voice_splits = list(_VOICE_MARKER_SEARCH_PATTERN.finditer(text))
+
+    if not voice_splits:
+        # No voice markers, return entire text with default voice
+        return [(default_voice, text)], default_voice, 0, 0
+
+    segments = []
+    current_voice = default_voice
+    valid_markers = 0
+    invalid_markers = 0
+
+    # Text before first marker uses default voice
+    first_start = voice_splits[0].start()
+    if first_start > 0:
+        intro_text = text[:first_start].strip()
+        if intro_text:
+            segments.append((current_voice, intro_text))
+
+    # Process each voice marker
+    for idx, match in enumerate(voice_splits):
+        voice_name = match.group(1).strip()
+        start = match.end()
+        end = voice_splits[idx + 1].start() if idx + 1 < len(voice_splits) else len(text)
+        segment_text = text[start:end].strip()
+
+        # Validate voice name
+        is_valid, invalid_voice = validate_voice_name(voice_name)
+        if is_valid:
+            # Normalize to lowercase to match canonical form
+            # Handle both single voices and formulas
+            if "*" in voice_name:
+                # Normalize each voice in the formula
+                normalized_parts = []
+                for part in voice_name.split("+"):
+                    part = part.strip()
+                    if "*" in part:
+                        voice_part, weight = part.split("*", 1)
+                        # Find the canonical (lowercase) voice name
+                        voice_part_lower = voice_part.strip().lower()
+                        canonical_voice = next(
+                            (v for v in get_voices("kokoro") if v.lower() == voice_part_lower),
+                            voice_part.strip()
+                        )
+                        normalized_parts.append(f"{canonical_voice}*{weight.strip()}")
+                current_voice = " + ".join(normalized_parts)
+            else:
+                # Find the canonical (lowercase) voice name
+                voice_name_lower = voice_name.lower()
+                current_voice = next(
+                    (v for v in get_voices("kokoro") if v.lower() == voice_name_lower),
+                    voice_name
+                )
+            valid_markers += 1
+        else:
+            # Invalid voice - stay with previous voice
+            invalid_markers += 1
+
+        if segment_text:
+            segments.append((current_voice, segment_text))
+
+    # Return segments, last voice, and counts
+    return segments, current_voice, valid_markers, invalid_markers

abogen/text_extractor.py

@@ -1023,8 +1023,13 @@ class EpubExtractor:
         if not html:
             return ""
         soup = BeautifulSoup(html, "html.parser")
-        for tag in soup.find_all(["p", "div"]):
+
+        # Add line breaks after block-level elements to ensure pauses in speech
+        for tag in soup.find_all(
+            ["p", "div", "h1", "h2", "h3", "h4", "h5", "h6", "li", "blockquote"]
+        ):
             tag.append("\n\n")
+
         for ol in soup.find_all("ol"):
             start_attr = ol.get("start")
             try:

abogen/tts_plugin/__init__.py

@@ -0,0 +1,170 @@
+"""TTS Plugin Architecture - Public API.
+
+This package defines the frozen Plugin API for the TTS Plugin Architecture.
+All public interfaces are fully defined but contain no business logic.
+
+Public modules:
+- types: Core domain value objects (AudioFormat, Duration, VoiceSelection, etc.)
+- errors: Error hierarchy (EngineError and subtypes)
+- manifest: Plugin manifest types (PluginManifest, EngineManifest, etc.)
+- engine: Engine and EngineSession protocols
+- capabilities: Optional capability interfaces (VoiceLister, PreviewGenerator, etc.)
+- host_context: HostContext dataclass
+- plugin: Plugin contract (create_engine function signature)
+- loader: Plugin discovery and loading
+- plugin_manager: Plugin management and engine creation
+- utils: Direct utility functions (get_voices, create_pipeline, etc.)
+
+Usage:
+    from abogen.tts_plugin import (
+        # Types
+        AudioFormat,
+        Duration,
+        VoiceSelection,
+        ParameterValues,
+        SynthesisRequest,
+        SynthesizedAudio,
+        EngineConfig,
+        # Errors
+        EngineError,
+        ModelNotFoundError,
+        ModelLoadError,
+        NetworkError,
+        InvalidInputError,
+        ConfigurationError,
+        CancelledError,
+        InternalError,
+        # Manifest
+        PluginManifest,
+        EngineManifest,
+        VoiceSourceManifest,
+        VoiceManifest,
+        ParameterManifest,
+        AudioFormatManifest,
+        EnumOption,
+        RequirementManifest,
+        GpuRequirement,
+        ModelManifest,
+        # Engine
+        Engine,
+        EngineSession,
+        # Capabilities
+        VoiceLister,
+        PreviewGenerator,
+        StreamingSynthesizer,
+        CancelableSession,
+        # Host Context
+        HostContext,
+        HttpClient,
+        # Plugin Manager
+        get_plugin_manager,
+        reset_plugin_manager,
+        # Utils
+        get_voices,
+        get_default_voice,
+        is_plugin_registered,
+        resolve_voice_to_plugin,
+        create_pipeline,
+    )
+"""
+
+from abogen.tts_plugin.capabilities import (
+    CancelableSession,
+    PreviewGenerator,
+    StreamingSynthesizer,
+    VoiceLister,
+)
+from abogen.tts_plugin.engine import Engine, EngineSession
+from abogen.tts_plugin.errors import (
+    CancelledError,
+    ConfigurationError,
+    EngineError,
+    InternalError,
+    InvalidInputError,
+    ModelLoadError,
+    ModelNotFoundError,
+    NetworkError,
+)
+from abogen.tts_plugin.host_context import HttpClient, HostContext
+from abogen.tts_plugin.manifest import (
+    AudioFormatManifest,
+    EngineManifest,
+    EnumOption,
+    GpuRequirement,
+    ModelManifest,
+    ParameterManifest,
+    PluginManifest,
+    RequirementManifest,
+    VoiceManifest,
+    VoiceSourceManifest,
+)
+from abogen.tts_plugin.types import (
+    AudioFormat,
+    Duration,
+    EngineConfig,
+    ParameterValues,
+    SynthesisRequest,
+    SynthesizedAudio,
+    VoiceSelection,
+)
+
+# Plugin Manager and Utils
+from abogen.tts_plugin.plugin_manager import get_plugin_manager, reset_plugin_manager
+from abogen.tts_plugin.utils import (
+    create_pipeline,
+    get_default_voice,
+    get_voices,
+    is_plugin_registered,
+    resolve_voice_to_plugin,
+)
+
+__all__ = [
+    # Types
+    "AudioFormat",
+    "Duration",
+    "VoiceSelection",
+    "ParameterValues",
+    "SynthesisRequest",
+    "SynthesizedAudio",
+    "EngineConfig",
+    # Errors
+    "EngineError",
+    "ModelNotFoundError",
+    "ModelLoadError",
+    "NetworkError",
+    "InvalidInputError",
+    "ConfigurationError",
+    "CancelledError",
+    "InternalError",
+    # Manifest
+    "PluginManifest",
+    "EngineManifest",
+    "VoiceSourceManifest",
+    "VoiceManifest",
+    "ParameterManifest",
+    "AudioFormatManifest",
+    "EnumOption",
+    "RequirementManifest",
+    "GpuRequirement",
+    "ModelManifest",
+    # Engine
+    "Engine",
+    "EngineSession",
+    # Capabilities
+    "VoiceLister",
+    "PreviewGenerator",
+    "StreamingSynthesizer",
+    "CancelableSession",
+    # Host Context
+    "HostContext",
+    "HttpClient",
+    # Plugin Manager
+    "get_plugin_manager",
+    "reset_plugin_manager",
+    # Utils
+    "get_voices",
+    "get_default_voice",
+    "is_plugin_registered",
+    "resolve_voice_to_plugin",
+    "create_pipeline",
+]

abogen/tts_plugin/capabilities.py

@@ -0,0 +1,103 @@
+"""Capability interfaces for the TTS Plugin Architecture.
+
+This module defines optional capability interfaces that engines can implement.
+Capabilities are additive; implementing new capabilities doesn't break old plugins.
+"""
+
+from __future__ import annotations
+
+from typing import Iterator, Protocol, runtime_checkable
+
+from abogen.tts_plugin.manifest import VoiceManifest
+from abogen.tts_plugin.types import SynthesisRequest, SynthesizedAudio, VoiceSelection
+
+
+@runtime_checkable
+class VoiceLister(Protocol):
+    """Protocol for listing available voices.
+
+    Engines that support voice listing should implement this interface.
+    """
+
+    def listVoices(self, sourceId: str) -> list[VoiceManifest]:
+        """List available voices for a given source.
+
+        Args:
+            sourceId: The voice source identifier.
+
+        Returns:
+            List of VoiceManifest describing available voices.
+
+        Raises:
+            EngineError: On failure.
+        """
+        ...
+
+
+@runtime_checkable
+class PreviewGenerator(Protocol):
+    """Protocol for generating voice previews.
+
+    Engines that support voice preview should implement this interface.
+    """
+
+    def generatePreview(self, voice: VoiceSelection, text: str) -> SynthesizedAudio:
+        """Generate a preview audio for a voice.
+
+        Args:
+            voice: Voice selection for the preview.
+            text: Text to use for the preview.
+
+        Returns:
+            SynthesizedAudio with the preview audio data.
+
+        Raises:
+            EngineError: On failure.
+        """
+        ...
+
+
+@runtime_checkable
+class StreamingSynthesizer(Protocol):
+    """Protocol for streaming synthesis.
+
+    Optional capability of EngineSession, not Engine.
+    Engines that support streaming synthesis should implement this interface.
+    """
+
+    def synthesizeStream(self, request: SynthesisRequest) -> Iterator[bytes]:
+        """Synthesize audio in streaming mode.
+
+        Args:
+            request: The synthesis request.
+
+        Yields:
+            Audio chunks as they become available.
+
+        Raises:
+            CancelledError: If cancel() is called during iteration.
+            EngineError: On synthesis failure.
+        """
+        ...
+        # This is a generator function; implementation will use yield
+        yield b""  # pragma: no cover
+
+
+@runtime_checkable
+class CancelableSession(Protocol):
+    """Protocol for cancellation support.
+
+    Optional capability for engines that support cancellation.
+    cancel() causes synthesize() to raise CancelledError.
+    """
+
+    def cancel(self) -> None:
+        """Cancel in-progress synthesis.
+
+        After cancellation, synthesize() raises CancelledError.
+        The session remains usable after cancellation.
+
+        Raises:
+            EngineError: If called after dispose().
+        """
+        ...

abogen/tts_plugin/engine.py

@@ -0,0 +1,95 @@
+"""Engine interfaces for the TTS Plugin Architecture.
+
+This module defines the core Engine and EngineSession protocols.
+These are the primary interfaces that plugin implementations must satisfy.
+"""
+
+from __future__ import annotations
+
+from typing import Protocol, runtime_checkable
+
+from abogen.tts_plugin.types import SynthesisRequest, SynthesizedAudio
+
+
+@runtime_checkable
+class EngineSession(Protocol):
+    """Protocol for a session that owns mutable execution state.
+
+    An EngineSession is created by Engine.createSession() and owns
+    mutable execution state isolated from other concurrent work.
+    It is NOT thread-safe.
+
+    Lifecycle:
+        1. Created by Engine.createSession()
+        2. Used for synthesis via synthesize()
+        3. Disposed via dispose()
+
+    After dispose(), all methods except dispose() raise EngineError.
+    """
+
+    def synthesize(self, request: SynthesisRequest) -> SynthesizedAudio:
+        """Synthesize audio from text.
+
+        Args:
+            request: The synthesis request containing text, voice, parameters, and format.
+
+        Returns:
+            SynthesizedAudio with the synthesized audio data.
+
+        Raises:
+            EngineError: On synthesis failure. Session remains usable after error.
+            EngineError: If called after dispose().
+        """
+        ...
+
+    def dispose(self) -> None:
+        """Release session resources.
+
+        This method is idempotent and safe to call multiple times.
+        It never raises exceptions (catches and logs internally).
+        After dispose(), all methods except dispose() raise EngineError.
+        """
+        ...
+
+
+@runtime_checkable
+class Engine(Protocol):
+    """Protocol for a TTS engine that creates sessions.
+
+    An Engine is a factory for EngineSession instances. It is stateless
+    and thread-safe for createSession().
+
+    Lifecycle:
+        1. Created via create_engine() (plugin contract)
+        2. Sessions created via createSession()
+        3. Disposed via dispose()
+
+    Thread Safety:
+        - createSession() is thread-safe and can be called from any thread.
+        - dispose() must be called after all sessions are disposed.
+        - Disposing engine while sessions are alive violates API contract.
+    """
+
+    def createSession(self) -> EngineSession:
+        """Create a new session for synthesis.
+
+        Returns:
+            A new EngineSession instance. Ownership transfers to caller.
+
+        Raises:
+            EngineError: On failure. No partially initialized session is returned.
+        """
+        ...
+
+    def dispose(self) -> None:
+        """Release engine resources.
+
+        Caller must ensure all sessions created by this engine are disposed
+        before calling dispose(). Disposing an engine while any session is
+        still alive violates the API contract; behavior is undefined.
+
+        This method is idempotent and safe to call multiple times.
+        It never raises exceptions (catches and logs internally).
+        After dispose(), all methods except dispose() raise EngineError.
+        """
+        ...

abogen/tts_plugin/errors.py

@@ -0,0 +1,62 @@
+"""Error hierarchy for the TTS Plugin Architecture.
+
+This module defines typed exceptions that engines raise.
+Engines should never raise raw exceptions; they must use EngineError or its subtypes.
+"""
+
+from __future__ import annotations
+
+
+class EngineError(Exception):
+    """Base exception for all engine errors.
+
+    All engine operations that can fail should raise EngineError or one of its subtypes.
+    After dispose(), all methods except dispose() raise EngineError.
+    """
+
+    pass
+
+
+class ModelNotFoundError(EngineError):
+    """Raised when a required model is not found."""
+
+    pass
+
+
+class ModelLoadError(EngineError):
+    """Raised when a model fails to load."""
+
+    pass
+
+
+class NetworkError(EngineError):
+    """Raised when a network operation fails."""
+
+    pass
+
+
+class InvalidInputError(EngineError):
+    """Raised when invalid input is provided to the engine."""
+
+    pass
+
+
+class ConfigurationError(EngineError):
+    """Raised when there is a configuration error."""
+
+    pass
+
+
+class CancelledError(EngineError):
+    """Raised when an operation is cancelled.
+
+    This is raised by synthesize() when cancel() is called during synthesis.
+    """
+
+    pass
+
+
+class InternalError(EngineError):
+    """Raised when an internal engine error occurs."""
+
+    pass

abogen/tts_plugin/host_context.py

@@ -0,0 +1,46 @@
+"""Host context for the TTS Plugin Architecture.
+
+This module defines the HostContext dataclass that provides minimal
+host services to plugins. It is the only interface through which
+plugins can access host functionality.
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Protocol, runtime_checkable
+
+
+@runtime_checkable
+class HttpClient(Protocol):
+    """Protocol for HTTP client provided by host.
+
+    Plugins can use this for network requests (e.g., API-based engines).
+    """
+
+    def get(self, url: str, **kwargs: object) -> object:
+        """Perform an HTTP GET request."""
+        ...
+
+    def post(self, url: str, **kwargs: object) -> object:
+        """Perform an HTTP POST request."""
+        ...
+
+
+@dataclass(frozen=True)
+class HostContext:
+    """Minimal host context provided to plugins.
+
+    Contains only essential host services. No business logic.
+
+    Attributes:
+        config_dir: Directory for API keys, preferences, and configuration.
+        logger: Logger for plugin logging.
+        http_client: HTTP client for network requests.
+    """
+
+    config_dir: Path
+    logger: logging.Logger
+    http_client: HttpClient

abogen/tts_plugin/loader.py

@@ -0,0 +1,365 @@
+"""Plugin loader infrastructure for the TTS Plugin Architecture.
+
+This module provides functionality to discover, import, validate, and load
+TTS plugins. It handles both valid and invalid plugins, providing diagnostic
+messages for errors.
+
+The loader does NOT:
+- Create Engine instances (that's the plugin's create_engine() responsibility)
+- Manage plugin lifecycle (that's the Plugin Manager's responsibility)
+- Implement any TTS engine functionality
+"""
+
+from __future__ import annotations
+
+import importlib
+import re
+import sys
+import types
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Callable
+
+from abogen.tts_plugin.manifest import ModelManifest, PluginManifest
+
+
+# Host API version for compatibility checking
+HOST_API_VERSION = "1.0"
+
+
+@dataclass(frozen=True)
+class PluginLoadError:
+    """Diagnostic information for a failed plugin load.
+
+    Attributes:
+        plugin_id: Plugin identifier if available, otherwise directory name.
+        path: Path to the plugin directory.
+        errors: List of error messages describing what went wrong.
+    """
+
+    plugin_id: str
+    path: Path
+    errors: tuple[str, ...] = field(default_factory=tuple)
+
+
+@dataclass(frozen=True)
+class PluginLoadResult:
+    """Result of loading a plugin.
+
+    Attributes:
+        success: Whether the plugin loaded successfully.
+        manifest: The plugin manifest if successful.
+        model_requirements: Model requirements if successful.
+        create_engine: The create_engine function if successful.
+        module: The plugin module if successful.
+        error: Error information if failed.
+    """
+
+    success: bool
+    manifest: PluginManifest | None = None
+    model_requirements: tuple[ModelManifest, ...] | None = None
+    create_engine: Callable[..., Any] | None = None
+    module: types.ModuleType | None = None
+    error: PluginLoadError | None = None
+
+
+def _parse_api_version(version: str) -> tuple[int, int] | None:
+    """Parse an api_version string into (major, minor) tuple.
+
+    Args:
+        version: Version string in format "MAJOR.MINOR".
+
+    Returns:
+        Tuple of (major, minor) or None if invalid format.
+    """
+    match = re.match(r"^(\d+)\.(\d+)$", version)
+    if match:
+        return int(match.group(1)), int(match.group(2))
+    return None
+
+
+def _check_api_version_compatibility(plugin_version: str) -> str | None:
+    """Check if plugin api_version is compatible with host.
+
+    Architecture spec:
+    - Format: semver (MAJOR.MINOR)
+    - Compatibility: Host rejects plugin if major version differs
+    - Minor version: backward compatible, Host accepts higher minor
+
+    Args:
+        plugin_version: Plugin's api_version string.
+
+    Returns:
+        Error message if incompatible, None if compatible.
+    """
+    plugin_ver = _parse_api_version(plugin_version)
+    if plugin_ver is None:
+        return f"Invalid api_version format: '{plugin_version}'. Expected format: MAJOR.MINOR"
+
+    host_ver = _parse_api_version(HOST_API_VERSION)
+    if host_ver is None:
+        return f"Invalid host api_version format: '{HOST_API_VERSION}'"
+
+    if plugin_ver[0] != host_ver[0]:
+        return (
+            f"api_version major mismatch: plugin={plugin_ver[0]}, host={host_ver[0]}. "
+            f"Major version must match for compatibility."
+        )
+
+    return None
+
+
+def _validate_manifest(module: types.ModuleType, plugin_dir: Path) -> list[str]:
+    """Validate that a plugin module has required exports.
+
+    Args:
+        module: The imported plugin module.
+        plugin_dir: Path to the plugin directory.
+
+    Returns:
+        List of error messages (empty if valid).
+    """
+    errors: list[str] = []
+
+    # Check PLUGIN_MANIFEST
+    manifest = getattr(module, "PLUGIN_MANIFEST", None)
+    if manifest is None:
+        errors.append("Missing PLUGIN_MANIFEST export")
+    elif not isinstance(manifest, PluginManifest):
+        errors.append(
+            f"PLUGIN_MANIFEST must be a PluginManifest instance, "
+            f"got {type(manifest).__name__}"
+        )
+
+    # Check MODEL_REQUIREMENTS
+    model_reqs = getattr(module, "MODEL_REQUIREMENTS", None)
+    if model_reqs is None:
+        errors.append("Missing MODEL_REQUIREMENTS export")
+    elif not isinstance(model_reqs, list):
+        errors.append(
+            f"MODEL_REQUIREMENTS must be a list, got {type(model_reqs).__name__}"
+        )
+    else:
+        for i, req in enumerate(model_reqs):
+            if not isinstance(req, ModelManifest):
+                errors.append(
+                    f"MODEL_REQUIREMENTS[{i}] must be a ModelManifest instance, "
+                    f"got {type(req).__name__}"
+                )
+
+    # Check create_engine
+    create_engine = getattr(module, "create_engine", None)
+    if create_engine is None:
+        errors.append("Missing create_engine export")
+    elif not callable(create_engine):
+        errors.append(
+            f"create_engine must be callable, got {type(create_engine).__name__}"
+        )
+
+    return errors
+
+
+def _validate_capabilities(manifest: PluginManifest) -> list[str]:
+    """Validate plugin capabilities.
+
+    Args:
+        manifest: The plugin manifest to validate.
+
+    Returns:
+        List of error messages (empty if valid).
+    """
+    errors: list[str] = []
+
+    # Known capabilities (can be extended)
+    known_capabilities = frozenset({
+        "voice_list",
+        "preview",
+        "voice_clone",
+        "voice_blend",
+        "streaming",
+        "cancel",
+    })
+
+    for cap in manifest.capabilities:
+        if cap not in known_capabilities:
+            errors.append(f"Unknown capability: '{cap}'")
+
+    return errors
+
+
+def _validate_api_version(manifest: PluginManifest) -> list[str]:
+    """Validate api_version compatibility.
+
+    Args:
+        manifest: The plugin manifest to validate.
+
+    Returns:
+        List of error messages (empty if valid).
+    """
+    errors: list[str] = []
+    error = _check_api_version_compatibility(manifest.api_version)
+    if error:
+        errors.append(error)
+    return errors
+
+
+def load_plugin_from_dir(plugin_dir: Path) -> PluginLoadResult:
+    """Load and validate a plugin from a directory.
+
+    The plugin directory must contain an __init__.py that exports:
+    - PLUGIN_MANIFEST: PluginManifest
+    - MODEL_REQUIREMENTS: list[ModelManifest]
+    - create_engine: Callable
+
+    Args:
+        plugin_dir: Path to the plugin directory.
+
+    Returns:
+        PluginLoadResult with success status and either plugin data or error info.
+    """
+    plugin_id = plugin_dir.name
+    errors: list[str] = []
+
+    # Check if directory exists
+    if not plugin_dir.exists():
+        return PluginLoadResult(
+            success=False,
+            error=PluginLoadError(
+                plugin_id=plugin_id,
+                path=plugin_dir,
+                errors=(f"Plugin directory does not exist: {plugin_dir}",),
+            ),
+        )
+
+    # Check for __init__.py
+    init_file = plugin_dir / "__init__.py"
+    if not init_file.exists():
+        return PluginLoadResult(
+            success=False,
+            error=PluginLoadError(
+                plugin_id=plugin_id,
+                path=plugin_dir,
+                errors=("Missing __init__.py in plugin directory",),
+            ),
+        )
+
+    # Import the module
+    module_name = f"abogen.tts_plugin._loaded.{plugin_id}"
+    try:
+        # Remove from cache if already imported (for testing)
+        if module_name in sys.modules:
+            del sys.modules[module_name]
+
+        spec = importlib.util.spec_from_file_location(
+            module_name, init_file, submodule_search_locations=[]
+        )
+        if spec is None or spec.loader is None:
+            return PluginLoadResult(
+                success=False,
+                error=PluginLoadError(
+                    plugin_id=plugin_id,
+                    path=plugin_dir,
+                    errors=(f"Failed to create module spec for {init_file}",),
+                ),
+            )
+
+        module = importlib.util.module_from_spec(spec)
+        sys.modules[module_name] = module
+        spec.loader.exec_module(module)
+    except Exception as e:
+        # Clean up module from sys.modules on import failure
+        if module_name in sys.modules:
+            del sys.modules[module_name]
+        return PluginLoadResult(
+            success=False,
+            error=PluginLoadError(
+                plugin_id=plugin_id,
+                path=plugin_dir,
+                errors=(f"Failed to import plugin module: {e}",),
+            ),
+        )
+
+    # Validate manifest
+    manifest_errors = _validate_manifest(module, plugin_dir)
+    errors.extend(manifest_errors)
+
+    # If manifest is valid, perform additional validation
+    manifest = getattr(module, "PLUGIN_MANIFEST", None)
+    if isinstance(manifest, PluginManifest):
+        # Validate api_version
+        api_errors = _validate_api_version(manifest)
+        errors.extend(api_errors)
+
+        # Validate capabilities
+        cap_errors = _validate_capabilities(manifest)
+        errors.extend(cap_errors)
+
+        # Use manifest id if available
+        plugin_id = manifest.id
+
+    # Check if any errors occurred
+    if errors:
+        # Clean up module from sys.modules
+        if module_name in sys.modules:
+            del sys.modules[module_name]
+
+        return PluginLoadResult(
+            success=False,
+            error=PluginLoadError(
+                plugin_id=plugin_id,
+                path=plugin_dir,
+                errors=tuple(errors),
+            ),
+        )
+
+    # Get MODEL_REQUIREMENTS
+    model_requirements = tuple(getattr(module, "MODEL_REQUIREMENTS", []))
+    create_engine = getattr(module, "create_engine", None)
+
+    return PluginLoadResult(
+        success=True,
+        manifest=manifest,
+        model_requirements=model_requirements,
+        create_engine=create_engine,
+        module=module,
+    )
+
+
+def discover_plugins(plugin_dirs: list[Path]) -> list[PluginLoadResult]:
+    """Discover and load plugins from multiple directories.
+
+    Args:
+        plugin_dirs: List of directories to scan for plugins.
+
+    Returns:
+        List of PluginLoadResult, one per plugin directory found.
+    """
+    results: list[PluginLoadResult] = []
+
+    for plugin_dir in plugin_dirs:
+        if not plugin_dir.exists():
+            continue
+
+        # Scan for subdirectories (each is a potential plugin)
+        for item in sorted(plugin_dir.iterdir()):
+            if item.is_dir() and not item.name.startswith("."):
+                result = load_plugin_from_dir(item)
+                results.append(result)
+
+    return results
+
+
+def load_plugin(
+    plugin_dir: Path,
+) -> PluginLoadResult:
+    """Load a single plugin from a directory.
+
+    This is the main entry point for loading a plugin.
+
+    Args:
+        plugin_dir: Path to the plugin directory.
+
+    Returns:
+        PluginLoadResult with success status and either plugin data or error info.
+    """
+    return load_plugin_from_dir(plugin_dir)

abogen/tts_plugin/manifest.py

@@ -0,0 +1,189 @@
+"""Plugin manifest types for the TTS Plugin Architecture.
+
+This module contains static metadata types that describe plugins.
+These types have no dependencies and are immutable.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any
+
+
+@dataclass(frozen=True)
+class AudioFormatManifest:
+    """Manifest describing an audio format.
+
+    Attributes:
+        mime: MIME type of the audio.
+        extension: File extension.
+    """
+
+    mime: str
+    extension: str
+
+
+@dataclass(frozen=True)
+class EnumOption:
+    """Manifest describing an enum option for a parameter.
+
+    Attributes:
+        value: The enum value.
+        label: Human-readable label.
+    """
+
+    value: str
+    label: str
+
+
+@dataclass(frozen=True)
+class ParameterManifest:
+    """Manifest describing a synthesis parameter.
+
+    Attributes:
+        id: Parameter identifier.
+        name: Human-readable name.
+        description: Parameter description.
+        type: Parameter type ("float", "int", "string", "boolean", "enum").
+        default: Default value.
+        min: Minimum value (optional, for numeric types).
+        max: Maximum value (optional, for numeric types).
+        step: Step size (optional, for numeric types).
+        options: Available options (optional, for enum type).
+        unit: Unit of measurement (optional).
+        group: Parameter group (optional).
+    """
+
+    id: str
+    name: str
+    description: str
+    type: str
+    default: Any
+    min: float | None = None
+    max: float | None = None
+    step: float | None = None
+    options: tuple[EnumOption, ...] = field(default_factory=tuple)
+    unit: str | None = None
+    group: str | None = None
+
+
+@dataclass(frozen=True)
+class VoiceManifest:
+    """Manifest describing a voice.
+
+    Attributes:
+        id: Voice identifier.
+        name: Human-readable name.
+        tags: Voice tags (e.g., language, style).
+    """
+
+    id: str
+    name: str
+    tags: tuple[str, ...] = field(default_factory=tuple)
+
+
+@dataclass(frozen=True)
+class VoiceSourceManifest:
+    """Manifest describing a voice source.
+
+    Attributes:
+        id: Voice source identifier.
+        name: Human-readable name.
+        type: Source type ("list", "speaker_id", "clone", "blend", "generate", "none").
+        config: Source-specific configuration.
+    """
+
+    id: str
+    name: str
+    type: str
+    config: Any = None
+
+
+@dataclass(frozen=True)
+class EngineManifest:
+    """Manifest describing engine capabilities.
+
+    Attributes:
+        voiceSources: Available voice sources.
+        parameters: Available synthesis parameters.
+        audioFormats: Supported audio formats.
+    """
+
+    voiceSources: tuple[VoiceSourceManifest, ...] = field(default_factory=tuple)
+    parameters: tuple[ParameterManifest, ...] = field(default_factory=tuple)
+    audioFormats: tuple[AudioFormatManifest, ...] = field(default_factory=tuple)
+
+
+@dataclass(frozen=True)
+class GpuRequirement:
+    """Manifest describing GPU requirements.
+
+    Attributes:
+        required: Whether GPU is required.
+        type: GPU type (e.g., "cuda", "rocm").
+        memory: Required GPU memory in GB.
+    """
+
+    required: bool = False
+    type: str | None = None
+    memory: float | None = None
+
+
+@dataclass(frozen=True)
+class RequirementManifest:
+    """Manifest describing plugin requirements.
+
+    Attributes:
+        gpu: GPU requirements (optional).
+        memory: Required RAM in GB (optional).
+        internet: Whether internet is required (optional).
+    """
+
+    gpu: GpuRequirement | None = None
+    memory: float | None = None
+    internet: bool | None = None
+
+
+@dataclass(frozen=True)
+class ModelManifest:
+    """Manifest describing a model requirement.
+
+    Attributes:
+        id: Model identifier.
+        name: Human-readable name.
+        size: Model size as string (e.g., "100MB", "2GB").
+    """
+
+    id: str
+    name: str
+    size: str
+
+
+@dataclass(frozen=True)
+class PluginManifest:
+    """Main manifest for a TTS plugin.
+
+    Attributes:
+        id: Plugin identifier (unique).
+        name: Human-readable name.
+        version: Plugin version.
+        api_version: API version (semver format: MAJOR.MINOR).
+        description: Plugin description.
+        author: Plugin author.
+        capabilities: List of capability identifiers.
+        requires: Plugin requirements.
+        engine: Engine manifest.
+        voices: Optional static voice catalog. None = not declared (use VoiceLister),
+            empty tuple = explicitly no static voices, non-empty = static catalog.
+    """
+
+    id: str
+    name: str
+    version: str
+    api_version: str
+    description: str
+    author: str
+    capabilities: tuple[str, ...] = field(default_factory=tuple)
+    requires: RequirementManifest = field(default_factory=RequirementManifest)
+    engine: EngineManifest = field(default_factory=EngineManifest)
+    voices: tuple[VoiceManifest, ...] | None = None

abogen/tts_plugin/plugin.py

@@ -0,0 +1,55 @@
+"""Plugin contract for the TTS Plugin Architecture.
+
+This module defines the plugin contract that all TTS plugins must implement.
+Each plugin must export:
+- PLUGIN_MANIFEST: PluginManifest instance
+- MODEL_REQUIREMENTS: list of ModelManifest instances
+- create_engine(): Factory function that creates an Engine
+
+The create_engine() function is the entry point for plugin activation.
+It must be atomic: succeed fully or raise and clean up.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Protocol, runtime_checkable
+
+from abogen.tts_plugin.engine import Engine
+from abogen.tts_plugin.host_context import HostContext
+from abogen.tts_plugin.types import EngineConfig
+
+
+@runtime_checkable
+class Plugin(Protocol):
+    """Protocol defining the plugin contract.
+
+    Every TTS plugin must implement this protocol by exporting:
+    - PLUGIN_MANIFEST: PluginManifest
+    - MODEL_REQUIREMENTS: list[ModelManifest]
+    - create_engine: Callable[[HostContext, Path | None, EngineConfig], Engine]
+    """
+
+    def create_engine(
+        self,
+        context: HostContext,
+        model_path: Path | None,
+        config: EngineConfig,
+    ) -> Engine:
+        """Create an engine instance.
+
+        This is the factory function that creates an Engine from a plugin.
+        It must be atomic: succeed fully or raise EngineError and clean up.
+
+        Args:
+            context: Host services (config dir, logger, http client).
+            model_path: Resolved model path, or None for cloud/no-model engines.
+            config: Engine initialization settings.
+
+        Returns:
+            A fully initialized Engine instance.
+
+        Raises:
+            EngineError: On failure. Cleans up partially created resources.
+        """
+        ...

abogen/tts_plugin/plugin_manager.py

@@ -0,0 +1,153 @@
+"""Plugin Manager
+
+Provides a simple interface for consumers to access TTS engines via the
+new Plugin Architecture. Discovers, loads, and manages plugins from the
+plugins directory.
+
+Usage:
+    from abogen.tts_plugin.plugin_manager import get_plugin_manager
+
+    manager = get_plugin_manager()
+    engine = manager.create_engine("kokoro", lang_code="a", device="cpu")
+    session = engine.create_session()
+    try:
+        result = session.synthesize("Hello world")
+    finally:
+        session.dispose()
+"""
+
+from typing import Any, Dict, List, Optional, Type
+
+from abogen.tts_plugin.engine import Engine, EngineSession
+from abogen.tts_plugin.manifest import PluginManifest
+from abogen.tts_plugin.types import AudioFormat
+
+
+class PluginManager:
+    """Manages TTS plugins and provides a simple interface for consumers."""
+
+    def __init__(self) -> None:
+        self._plugins: Dict[str, dict] = {}
+        self._engines: Dict[str, Engine] = {}
+        self._loaded = False
+
+    def discover(self, plugins_dir: str = "plugins") -> None:
+        """Discover and load all plugins from the given directory."""
+        import os
+        from pathlib import Path
+        from abogen.tts_plugin.loader import load_plugin_from_dir
+
+        self._plugins.clear()
+        self._engines.clear()
+
+        plugins_path = Path(plugins_dir)
+        if not plugins_path.exists():
+            self._loaded = True
+            return
+
+        for entry in plugins_path.iterdir():
+            if entry.is_dir() and (entry / "__init__.py").exists():
+                try:
+                    result = load_plugin_from_dir(entry)
+                    if result.success and result.manifest is not None:
+                        self._plugins[result.manifest.id] = {
+                            "manifest": result.manifest,
+                            "create_engine": result.create_engine,
+                            "module": result.module,
+                        }
+                except Exception as e:
+                    # Log error but continue with other plugins
+                    print(f"Warning: Failed to load plugin from {entry}: {e}")
+
+        self._loaded = True
+
+    def _ensure_loaded(self) -> None:
+        """Ensure plugins have been discovered."""
+        if not self._loaded:
+            self.discover()
+
+    def list_plugins(self) -> List[PluginManifest]:
+        """Return manifests for all loaded plugins."""
+        self._ensure_loaded()
+        return [info["manifest"] for info in self._plugins.values()]
+
+    def get_plugin(self, plugin_id: str) -> Optional[dict]:
+        """Get plugin info by ID."""
+        self._ensure_loaded()
+        return self._plugins.get(plugin_id)
+
+    def has_plugin(self, plugin_id: str) -> bool:
+        """Check if a plugin is loaded."""
+        self._ensure_loaded()
+        return plugin_id in self._plugins
+
+    def create_engine(self, plugin_id: str, **kwargs: Any) -> Engine:
+        """Create an engine instance for the given plugin.
+
+        Args:
+            plugin_id: The plugin identifier (e.g., "kokoro")
+            **kwargs: Arguments passed to the engine constructor
+
+        Returns:
+            An Engine instance
+
+        Raises:
+            KeyError: If plugin_id is not found
+            Exception: If engine creation fails
+        """
+        self._ensure_loaded()
+
+        if plugin_id not in self._plugins:
+            raise KeyError(f"Plugin not found: {plugin_id}")
+
+        plugin_info = self._plugins[plugin_id]
+        create_engine_func = plugin_info["create_engine"]
+
+        # Create engine using the plugin's factory
+        engine = create_engine_func(**kwargs)
+        return engine
+
+    def get_or_create_engine(self, plugin_id: str, **kwargs: Any) -> Engine:
+        """Get an existing engine or create a new one.
+
+        Engines are cached by plugin_id. If you need multiple instances
+        with different parameters, use create_engine() directly.
+        """
+        self._ensure_loaded()
+
+        cache_key = plugin_id
+        if cache_key in self._engines:
+            return self._engines[cache_key]
+
+        engine = self.create_engine(plugin_id, **kwargs)
+        self._engines[cache_key] = engine
+        return engine
+
+    def dispose_all(self) -> None:
+        """Dispose all cached engines."""
+        for engine in self._engines.values():
+            try:
+                engine.dispose()
+            except Exception:
+                pass  # dispose() should never raise
+        self._engines.clear()
+
+
+# Global singleton
+_manager: Optional[PluginManager] = None
+
+
+def get_plugin_manager() -> PluginManager:
+    """Get the global PluginManager instance."""
+    global _manager
+    if _manager is None:
+        _manager = PluginManager()
+    return _manager
+
+
+def reset_plugin_manager() -> None:
+    """Reset the global PluginManager (for testing)."""
+    global _manager
+    if _manager is not None:
+        _manager.dispose_all()
+    _manager = None

abogen/tts_plugin/types.py

@@ -0,0 +1,111 @@
+"""Core domain types for the TTS Plugin Architecture.
+
+This module contains immutable value objects that form the core domain.
+These types have zero dependencies and are used across the plugin system.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any, Mapping
+
+
+@dataclass(frozen=True)
+class AudioFormat:
+    """Immutable value object representing an audio format.
+
+    Attributes:
+        mime: MIME type of the audio (e.g., "audio/wav", "audio/mpeg").
+        extension: File extension (e.g., "wav", "mp3").
+    """
+
+    mime: str
+    extension: str
+
+
+@dataclass(frozen=True)
+class Duration:
+    """Immutable value object representing a time duration.
+
+    Attributes:
+        seconds: Duration in seconds.
+    """
+
+    seconds: float
+
+
+@dataclass(frozen=True)
+class VoiceSelection:
+    """Immutable value object for voice selection. Opaque to engine.
+
+    Attributes:
+        source: Voice source identifier (e.g., "builtin", "clone").
+        key: Voice key within the source.
+        payload: Optional payload for clone/blend sources.
+    """
+
+    source: str
+    key: str
+    payload: Any = None
+
+
+@dataclass(frozen=True)
+class ParameterValues:
+    """Immutable value object for synthesis parameters. Behaves like Mapping[str, Any].
+
+    Attributes:
+        values: Mapping of parameter names to their values.
+    """
+
+    values: Mapping[str, Any] = field(default_factory=dict)
+
+
+@dataclass(frozen=True)
+class SynthesisRequest:
+    """Immutable value object for a synthesis request.
+
+    Attributes:
+        text: Text to synthesize.
+        voice: Voice selection.
+        parameters: Synthesis parameters.
+        format: Desired audio output format.
+    """
+
+    text: str
+    voice: VoiceSelection
+    parameters: ParameterValues
+    format: AudioFormat
+
+
+@dataclass(frozen=True)
+class SynthesizedAudio:
+    """Immutable value object for synthesized audio result.
+
+    Attributes:
+        data: Raw audio bytes.
+        format: Audio format of the result.
+        duration: Duration of the audio.
+    """
+
+    data: bytes
+    format: AudioFormat
+    duration: Duration
+
+
+@dataclass(frozen=True)
+class EngineConfig:
+    """Immutable configuration of an Engine instance.
+
+    Contains parameters that define how a particular Engine instance is
+    created and that remain constant throughout the lifetime of that Engine.
+
+    Plugin implementations may ignore fields that are not applicable to them.
+
+    Attributes:
+        device: Device to use (e.g., "cpu", "cuda:0").
+        lang_code: Language code for the engine (e.g., "a" for Kokoro English).
+            Plugins that do not require a language code ignore this field.
+    """
+
+    device: str = "cpu"
+    lang_code: str = "a"

abogen/tts_plugin/utils.py

@@ -0,0 +1,235 @@
+"""TTS Plugin Architecture — direct utility functions.
+
+Provides helpers that replace the former compatibility adapter by
+calling the Plugin Manager directly.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Iterator
+
+import numpy as np
+
+from abogen.tts_plugin.plugin_manager import get_plugin_manager
+
+
+def get_voices(plugin_id: str) -> tuple[str, ...]:
+    """Return the voice-id tuple for *plugin_id*.
+
+    Uses the official Plugin Architecture: PluginManager → Engine → VoiceLister.
+    First checks plugin manifest for static voice catalog.
+    """
+    import logging
+    import tempfile
+    from pathlib import Path
+
+    from abogen.tts_plugin.host_context import HostContext
+    from abogen.tts_plugin.types import EngineConfig
+
+    manager = get_plugin_manager()
+    if not manager.has_plugin(plugin_id):
+        return ()
+
+    # Check manifest for static voice catalog
+    plugin_info = manager.get_plugin(plugin_id)
+    if plugin_info is not None:
+        manifest = plugin_info.get("manifest")
+        if manifest is not None and manifest.voices is not None:
+            return tuple(v.id for v in manifest.voices)
+
+    ctx = HostContext(
+        config_dir=Path(tempfile.gettempdir()),
+        logger=logging.getLogger(f"abogen.utils.{plugin_id}"),
+        http_client=type("_StubHttpClient", (), {
+            "get": staticmethod(lambda url, **kw: None),
+            "post": staticmethod(lambda url, **kw: None),
+        })(),
+    )
+
+    try:
+        engine = manager.create_engine(
+            plugin_id,
+            context=ctx,
+            model_path=None,
+            config=EngineConfig(device="cpu"),
+        )
+    except Exception:
+        return ()
+
+    try:
+        from abogen.tts_plugin.capabilities import VoiceLister
+
+        if isinstance(engine, VoiceLister):
+            manifests = engine.listVoices("builtin")
+            return tuple(v.id for v in manifests)
+        return ()
+    except Exception:
+        return ()
+    finally:
+        engine.dispose()
+
+
+def get_default_voice(plugin_id: str, fallback: str = "") -> str:
+    """Return the first voice of *plugin_id*, or *fallback*."""
+    voices = get_voices(plugin_id)
+    return voices[0] if voices else fallback
+
+
+def is_plugin_registered(plugin_id: str) -> bool:
+    """Check whether *plugin_id* is loaded by the Plugin Manager."""
+    return get_plugin_manager().has_plugin(plugin_id)
+
+
+def resolve_voice_to_plugin(spec: str, fallback: str = "kokoro") -> str:
+    """Determine which plugin owns the given voice specification.
+
+    Resolution rules:
+    1. Empty spec -> fallback
+    2. Kokoro formula (contains '*' or '+') -> "kokoro"
+    3. Exact voice-id match against loaded plugins -> plugin id
+    4. Unknown voice -> fallback
+    """
+    raw = str(spec or "").strip()
+    if not raw:
+        return fallback
+
+    if "*" in raw or "+" in raw:
+        return "kokoro"
+
+    upper = raw.upper()
+    manager = get_plugin_manager()
+
+    for manifest in manager.list_plugins():
+        for voice_source in manifest.engine.voiceSources:
+            if voice_source.type == "list" and isinstance(voice_source.config, dict):
+                try:
+                    engine = manager.create_engine(manifest.id)
+                    try:
+                        if hasattr(engine, "listVoices"):
+                            voice_manifests = engine.listVoices(voice_source.id)
+                            voice_ids = [v.id.upper() for v in voice_manifests]
+                            if upper in voice_ids:
+                                return manifest.id
+                    finally:
+                        engine.dispose()
+                except Exception:
+                    continue
+
+    return fallback
+
+
+class Pipeline:
+    """Callable wrapper around Engine / EngineSession.
+
+    Presents the same interface that old callers expect::
+
+        pipeline = create_pipeline("kokoro", lang_code="a", device="cpu")
+        for segment in pipeline(text, voice="af_nova", speed=1.0):
+            audio = segment.audio
+    """
+
+    def __init__(self, engine: Any, **engine_kwargs: Any) -> None:
+        self._engine = engine
+        self._engine_kwargs = engine_kwargs
+        self._session: Any = None
+
+    def _ensure_session(self) -> Any:
+        if self._session is None:
+            self._session = self._engine.createSession()
+        return self._session
+
+    def __call__(
+        self,
+        text: str,
+        voice: str = "default",
+        speed: float = 1.0,
+        split_pattern: str | None = None,
+        **kwargs: Any,
+    ) -> Iterator[Any]:
+        from abogen.tts_plugin.types import (
+            AudioFormat,
+            ParameterValues,
+            SynthesisRequest,
+            VoiceSelection,
+        )
+
+        session = self._ensure_session()
+
+        params: dict[str, Any] = {"speed": speed}
+        if split_pattern is not None:
+            params["split_pattern"] = split_pattern
+        params.update(kwargs)
+
+        request = SynthesisRequest(
+            text=text,
+            voice=VoiceSelection(source="builtin", key=voice),
+            parameters=ParameterValues(values=params),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+
+        result = session.synthesize(request)
+        audio_array = np.frombuffer(result.data, dtype=np.float32)
+
+        from dataclasses import dataclass
+
+        @dataclass
+        class Segment:
+            graphemes: str
+            audio: np.ndarray
+
+        yield Segment(graphemes=text, audio=audio_array)
+
+    def dispose(self) -> None:
+        if self._session is not None:
+            try:
+                self._session.dispose()
+            except Exception:
+                pass
+            self._session = None
+
+    def __del__(self) -> None:
+        self.dispose()
+
+
+def create_pipeline(
+    plugin_id: str,
+    *,
+    lang_code: str = "a",
+    device: str = "cpu",
+) -> Pipeline:
+    """Create a callable TTS pipeline via the Plugin Architecture.
+
+    Builds a proper HostContext and EngineConfig, then delegates to the
+    PluginManager to create the engine. Returns a :class:`Pipeline` whose
+    ``__call__`` interface matches the callable protocol used by consumers.
+
+    Args:
+        plugin_id: Plugin identifier (e.g., "kokoro", "supertonic").
+        lang_code: Language code for the engine.
+        device: Device to use (e.g., "cpu", "cuda:0").
+
+    Returns:
+        A callable Pipeline instance.
+    """
+    import logging
+    import tempfile
+    from pathlib import Path
+
+    from abogen.tts_plugin.host_context import HostContext
+    from abogen.tts_plugin.types import EngineConfig
+
+    manager = get_plugin_manager()
+
+    ctx = HostContext(
+        config_dir=Path(tempfile.gettempdir()),
+        logger=logging.getLogger(f"abogen.pipeline.{plugin_id}"),
+        http_client=type("_StubHttpClient", (), {
+            "get": staticmethod(lambda url, **kw: None),
+            "post": staticmethod(lambda url, **kw: None),
+        })(),
+    )
+
+    config = EngineConfig(device=device, lang_code=lang_code)
+
+    engine = manager.create_engine(plugin_id, context=ctx, model_path=None, config=config)
+    return Pipeline(engine)

abogen/utils.py

@@ -529,21 +529,20 @@ def prevent_sleep_end():
                 _sleep_procs[system] = None
 
 
-def load_numpy_kpipeline():
-    import numpy as np
-    from kokoro import KPipeline  # type: ignore[import-not-found]
-
-    return np, KPipeline
-
-
 class LoadPipelineThread(Thread):
-    def __init__(self, callback):
+    def __init__(self, callback, lang_code="a", device="cpu"):
         super().__init__()
         self.callback = callback
+        self.lang_code = lang_code
+        self.device = device
 
     def run(self):
         try:
-            np_module, kpipeline_class = load_numpy_kpipeline()
-            self.callback(np_module, kpipeline_class, None)
+            from abogen.tts_plugin.utils import create_pipeline
+
+            backend = create_pipeline(
+                "kokoro", lang_code=self.lang_code, device=self.device
+            )
+            self.callback(backend, None)
         except Exception as e:
-            self.callback(None, None, str(e))
+            self.callback(None, str(e))

abogen/voice_cache.py

@@ -17,7 +17,7 @@ if LocalEntryNotFoundError is None:  # pragma: no cover - fallback for tests
         pass
 
 
-from abogen.constants import VOICES_INTERNAL
+from abogen.tts_plugin.utils import get_voices
 
 _CACHE_LOCK = threading.Lock()
 _CACHED_VOICES: Set[str] = set()
@@ -26,8 +26,9 @@ _BOOTSTRAPPED = False
 
 
 def _normalize_targets(voices: Optional[Iterable[str]]) -> Set[str]:
+    kokoro_voices = get_voices("kokoro")
     if not voices:
-        return set(VOICES_INTERNAL)
+        return set(kokoro_voices)
     normalized: Set[str] = set()
     for voice in voices:
         if not voice:
@@ -35,7 +36,7 @@ def _normalize_targets(voices: Optional[Iterable[str]]) -> Set[str]:
         voice_id = str(voice).strip()
         if not voice_id:
             continue
-        if voice_id in VOICES_INTERNAL:
+        if voice_id in kokoro_voices:
             normalized.add(voice_id)
     return normalized
 
@@ -143,3 +144,11 @@ def _ensure_single_voice_asset(
 
     hf_hub_download(resume_download=True, **common_kwargs)
     return True
+
+
+def clear_voice_cache() -> None:
+    """Clear the in‑process voice cache (used during shutdown)."""
+    with _CACHE_LOCK:
+        _CACHED_VOICES.clear()
+    global _BOOTSTRAPPED
+    _BOOTSTRAPPED = False

abogen/voice_formulas.py

@@ -1,7 +1,7 @@
 import re
 from typing import List, Tuple
 
-from abogen.constants import VOICES_INTERNAL
+from abogen.tts_plugin.utils import get_voices
 
 
 # Calls parsing and loads the voice to gpu or cpu
@@ -22,6 +22,7 @@ def parse_formula_terms(formula: str) -> List[Tuple[str, float]]:
         raise ValueError("Empty voice formula")
 
     terms: List[Tuple[str, float]] = []
+    kokoro_voices = get_voices("kokoro")
     for segment in formula.split("+"):
         part = segment.strip()
         if not part:
@@ -30,7 +31,7 @@ def parse_formula_terms(formula: str) -> List[Tuple[str, float]]:
             raise ValueError("Each component must be in the form voice*weight")
         voice_name, raw_weight = part.split("*", 1)
         voice_name = voice_name.strip()
-        if voice_name not in VOICES_INTERNAL:
+        if voice_name not in kokoro_voices:
             raise ValueError(f"Unknown voice: {voice_name}")
         try:
             weight = float(raw_weight.strip())

abogen/voice_metadata.py

@@ -0,0 +1,33 @@
+from dataclasses import dataclass
+
+
+@dataclass(frozen=True)
+class VoiceMetadata:
+    """
+    Immutable metadata describing a voice from a TTS backend.
+
+    This model describes a voice independently of any backend implementation.
+    Backends populate these objects; the application consumes them.
+
+    The ``backend_id`` field is set by the backend itself (via
+    ``self.metadata.id``) — the application never hardcodes it.
+    This ensures renaming a backend does not require touching voice definitions.
+    """
+
+    id: str
+    """Unique voice identifier within the backend (e.g. ``"af_alloy"``, ``"M1"``)."""
+
+    display_name: str
+    """Human-readable display name (e.g. ``"Alloy"``, ``"Male 1"``)."""
+
+    language: str
+    """Language code — backend-specific format is acceptable (e.g. ``"a"``, ``"en"``)."""
+
+    gender: str
+    """Gender category: ``"female"``, ``"male"``, or ``"unknown"``."""
+
+    backend_id: str
+    """Identifier of the backend that owns this voice (e.g. ``"kokoro"``).
+
+    Set automatically by the backend — never hardcoded in voice definitions.
+    """

abogen/voice_profiles.py

@@ -2,8 +2,7 @@ import json
 import os
 from typing import Any, Dict, Iterable, List, Tuple
 
-from abogen.constants import VOICES_INTERNAL
-from abogen.tts_supertonic import DEFAULT_SUPERTONIC_VOICES
+from abogen.tts_plugin.utils import get_voices, is_plugin_registered
 from abogen.utils import get_user_config_path
 
 
@@ -70,7 +69,8 @@ def serialize_profiles() -> Dict[str, Dict[str, Iterable[Tuple[str, float]]]]:
 
 def _normalize_supertonic_voice(value: Any) -> str:
     raw = str(value or "").strip().upper()
-    return raw if raw in DEFAULT_SUPERTONIC_VOICES else "M1"
+    supertonic_voices = get_voices("supertonic")
+    return raw if raw in supertonic_voices else "M1"
 
 
 def _coerce_supertonic_steps(value: Any) -> int:
@@ -101,7 +101,7 @@ def normalize_profile_entry(entry: Any) -> Dict[str, Any]:
         return {}
 
     provider = str(entry.get("provider") or "kokoro").strip().lower()
-    if provider not in {"kokoro", "supertonic"}:
+    if not is_plugin_registered(provider):
         provider = "kokoro"
 
     language = str(entry.get("language") or "a").strip().lower() or "a"
@@ -135,6 +135,7 @@ def normalize_profile_entry(entry: Any) -> Dict[str, Any]:
 
 def _normalize_voice_entries(entries: Iterable) -> List[Tuple[str, float]]:
     normalized: List[Tuple[str, float]] = []
+    kokoro_voices = get_voices("kokoro")
     for item in entries or []:
         if isinstance(item, dict):
             voice = item.get("id") or item.get("voice")
@@ -143,7 +144,7 @@ def _normalize_voice_entries(entries: Iterable) -> List[Tuple[str, float]]:
             voice, weight = item[0], item[1]
         else:
             continue
-        if voice not in VOICES_INTERNAL:
+        if voice not in kokoro_voices:
             continue
         if weight is None:
             continue

abogen/webui/Dockerfile

@@ -2,7 +2,6 @@ FROM nvidia/cuda:12.6.3-cudnn-runtime-ubuntu22.04
 
 ENV PYTHONDONTWRITEBYTECODE=1 \
   PYTHONUNBUFFERED=1 \
-  PIP_NO_CACHE_DIR=1 \
   VIRTUAL_ENV=/opt/venv \
   PATH=/opt/venv/bin:$PATH
 
@@ -27,22 +26,22 @@ RUN python3 -m venv "$VIRTUAL_ENV"
 WORKDIR /app
 
 COPY pyproject.toml README.md ./
-COPY abogen ./abogen
-
-RUN pip install --upgrade pip \
+RUN pip install uv \
   && if [ -n "$TORCH_VERSION" ]; then \
-       pip install torch=="$TORCH_VERSION" torchvision=="$TORCH_VERSION" torchaudio=="$TORCH_VERSION" --index-url "$TORCH_INDEX_URL"; \
+       uv pip install --system torch=="$TORCH_VERSION" torchvision=="$TORCH_VERSION" torchaudio=="$TORCH_VERSION" --index-url "$TORCH_INDEX_URL"; \
      else \
-       pip install torch torchvision torchaudio --index-url "$TORCH_INDEX_URL"; \
+       uv pip install --system torch torchvision torchaudio --index-url "$TORCH_INDEX_URL"; \
      fi \
-  && pip install --no-cache-dir . \
+  && uv pip install --system . \
     https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.8.0/en_core_web_sm-3.8.0-py3-none-any.whl \
-  && pip install --no-cache-dir "mutagen>=1.47.0"
+  && uv pip install --system "mutagen>=1.47.0"
+
+COPY abogen ./abogen
 
 # Install onnxruntime-gpu for CUDA acceleration (supertonic uses ONNX Runtime)
 # Set USE_GPU=false to skip this for CPU-only deployments
 RUN if [ "$USE_GPU" = "true" ]; then \
-      pip install --no-cache-dir onnxruntime-gpu; \
+      uv pip install --system onnxruntime-gpu; \
     fi
 
 ENV ABOGEN_HOST=0.0.0.0 \

abogen/webui/app.py

@@ -1,6 +1,5 @@
 from __future__ import annotations
 
-import atexit
 import logging
 import os
 from pathlib import Path
@@ -8,6 +7,8 @@ from typing import Any, Optional
 
 from flask import Flask
 
+from abogen import shutdown  # noqa: F401
+shutdown.register_shutdown()
 from abogen.utils import get_user_cache_path, get_user_output_path, get_user_settings_dir
 
 from .conversion_runner import run_conversion_job
@@ -113,8 +114,6 @@ def create_app(config: Optional[dict[str, Any]] = None) -> Flask:
     app.register_blueprint(books_bp, url_prefix="/find-books")
     app.register_blueprint(api_bp, url_prefix="/api")
 
-    atexit.register(service.shutdown)
-
     global _access_log_filter_attached
     if not _access_log_filter_attached:
         logging.getLogger("werkzeug").addFilter(_SuppressSuccessfulAccessFilter())

abogen/webui/conversion_runner.py

@@ -20,7 +20,7 @@ import numpy as np
 import soundfile as sf
 import static_ffmpeg
 
-from abogen.constants import VOICES_INTERNAL
+from abogen.tts_plugin.utils import get_voices, is_plugin_registered, resolve_voice_to_plugin
 from abogen.epub3.exporter import build_epub3_package
 from abogen.kokoro_text_normalization import ApostropheConfig, normalize_for_pipeline, HAS_NUM2WORDS
 from abogen.normalization_settings import (
@@ -39,14 +39,14 @@ from abogen.utils import (
     get_user_cache_path,
     get_user_output_path,
     load_config,
-    load_numpy_kpipeline,
 )
+from abogen.tts_plugin.utils import create_pipeline
 from abogen.voice_cache import ensure_voice_assets
 from abogen.voice_formulas import extract_voice_ids, get_new_voice
 from abogen.voice_profiles import load_profiles, normalize_profile_entry
 from abogen.pronunciation_store import increment_usage
 from abogen.llm_client import LLMClientError
-from abogen.tts_supertonic import DEFAULT_SUPERTONIC_VOICES, SupertonicPipeline
+
 
 from .service import Job, JobStatus
 
@@ -56,26 +56,27 @@ SAMPLE_RATE = 24000
 
 
 def _supertonic_voice_from_spec(spec: Any, fallback: str) -> str:
+    """Normalize a voice specification for Supertonic.
+
+    This function only performs Supertonic-specific normalization (uppercase conversion
+    and fallback handling). Backend resolution is handled by the registry.
+    """
     raw = str(spec or "").strip()
     fallback_raw = str(fallback or "").strip()
 
-    # SuperTonic voices are discrete IDs (M1/F3/...). If we see a Kokoro mix
-    # formula (contains '*' or '+'), ignore it and fall back to a safe voice.
-    if not raw or "*" in raw or "+" in raw:
-        raw = fallback_raw
-    if not raw or "*" in raw or "+" in raw:
-        raw = "M1"
+    # Normalize to uppercase for Supertonic voice IDs
+    upper = raw.upper() if raw else ""
+
+    # If empty or contains formula characters, use fallback
+    if not upper or "*" in upper or "+" in upper:
+        upper = fallback_raw.upper() if fallback_raw else ""
+
+    # If still empty, use default Supertonic voice
+    if not upper or "*" in upper or "+" in upper:
+        upper = "M1"
 
-    upper = raw.upper()
-    if upper in DEFAULT_SUPERTONIC_VOICES:
     return upper
 
-    fallback_upper = fallback_raw.upper() if fallback_raw else ""
-    if fallback_upper in DEFAULT_SUPERTONIC_VOICES:
-        return fallback_upper
-
-    return "M1"
-
 
 def _split_speaker_reference(value: Any) -> tuple[Optional[str], str]:
     raw = str(value or "").strip()
@@ -118,15 +119,7 @@ def _formula_from_kokoro_entry(entry: Mapping[str, Any]) -> str:
 
 
 def _infer_provider_from_spec(value: Any, fallback: str = "kokoro") -> str:
-    raw = str(value or "").strip()
-    if not raw:
-        return fallback
-    upper = raw.upper()
-    if upper in DEFAULT_SUPERTONIC_VOICES:
-        return "supertonic"
-    if "*" in raw or "+" in raw:
-        return "kokoro"
-    return fallback
+    return resolve_voice_to_plugin(str(value or ""), fallback=fallback)
 
 
 class _JobCancelled(Exception):
@@ -575,7 +568,7 @@ def _spec_to_voice_ids(spec: Any) -> Set[str]:
             return set(extract_voice_ids(text))
         except ValueError:
             return set()
-    if text in VOICES_INTERNAL:
+    if text in get_voices("kokoro"):
         return {text}
     return set()
 
@@ -639,7 +632,7 @@ def _collect_required_voice_ids(job: Job) -> Set[str]:
             for key in ("resolved_voice", "voice_formula", "voice"):
                 voices.update(_spec_to_voice_ids(payload.get(key)))
 
-    voices.update(VOICES_INTERNAL)
+    voices.update(get_voices("kokoro"))
     return voices
 
 
@@ -1573,7 +1566,7 @@ def run_conversion_job(job: Job) -> None:
         def get_pipeline(provider: str) -> Any:
             nonlocal kokoro_cache_ready
             provider_norm = str(provider or "kokoro").strip().lower() or "kokoro"
-            if provider_norm not in {"kokoro", "supertonic"}:
+            if not is_plugin_registered(provider_norm):
                 provider_norm = "kokoro"
 
             existing = pipelines.get(provider_norm)
@@ -1581,10 +1574,8 @@ def run_conversion_job(job: Job) -> None:
                 return existing
 
             if provider_norm == "supertonic":
-                pipelines[provider_norm] = SupertonicPipeline(
-                    sample_rate=SAMPLE_RATE,
-                    auto_download=True,
-                    total_steps=int(getattr(job, "supertonic_total_steps", 5) or 5),
+                pipelines[provider_norm] = create_pipeline(
+                    "supertonic",
                 )
                 return pipelines[provider_norm]
 
@@ -1594,16 +1585,12 @@ def run_conversion_job(job: Job) -> None:
             device = "cpu"
             if not disable_gpu:
                 device = _select_device()
-            _np, KPipeline = load_numpy_kpipeline()
-            # Try to initialize with the selected device; fall back to CPU if CUDA fails
-            try:
-                pipelines[provider_norm] = KPipeline(lang_code=job.language, repo_id="hexgrad/Kokoro-82M", device=device)
-            except RuntimeError as e:
-                if "CUDA" in str(e) and device != "cpu":
-                    job.add_log(f"CUDA initialization failed, falling back to CPU: {e}", level="warning")
-                    pipelines[provider_norm] = KPipeline(lang_code=job.language, repo_id="hexgrad/Kokoro-82M", device="cpu")
-                else:
-                    raise
+            # Create KPipeline instance directly (uses new Plugin Architecture)
+            pipelines[provider_norm] = create_pipeline(
+                "kokoro",
+                lang_code=job.language,
+                device=device
+            )
             if not kokoro_cache_ready:
                 _initialize_voice_cache(job)
                 kokoro_cache_ready = True
@@ -1644,8 +1631,8 @@ def run_conversion_job(job: Job) -> None:
                 return provider, resolved, cached, speed, steps
 
             if provider == "kokoro":
-                kokoro_pipeline = get_pipeline("kokoro")
-                choice = _resolve_voice(kokoro_pipeline, resolved, job.use_gpu)
+                kokoro_backend = get_pipeline("kokoro")
+                choice = _resolve_voice(kokoro_backend, resolved, job.use_gpu)
             else:
                 choice = resolved
 
@@ -1774,8 +1761,8 @@ def run_conversion_job(job: Job) -> None:
         voice_cache: Dict[str, Any] = {}
         base_provider, base_voice_resolved, _, _ = resolve_voice_target(base_voice_spec)
         if base_provider == "kokoro" and base_voice_resolved and "*" not in base_voice_resolved:
-            kokoro_pipeline = get_pipeline("kokoro")
-            voice_cache[f"kokoro:{base_voice_resolved}"] = _resolve_voice(kokoro_pipeline, base_voice_resolved, job.use_gpu)
+            kokoro_backend = get_pipeline("kokoro")
+            voice_cache[f"kokoro:{base_voice_resolved}"] = _resolve_voice(kokoro_backend, base_voice_resolved, job.use_gpu)
         processed_chars = 0
         subtitle_index = 1
         current_time = 0.0
@@ -1805,8 +1792,8 @@ def run_conversion_job(job: Job) -> None:
                     fallback_key = next(iter(voice_cache.keys()), "")
                     if fallback_key and fallback_key != "__custom_mix":
                         intro_voice_spec = fallback_key.split(":", 1)[-1]
-                if not intro_voice_spec and VOICES_INTERNAL:
-                    intro_voice_spec = VOICES_INTERNAL[0]
+                if not intro_voice_spec:
+                    intro_voice_spec = get_default_voice("kokoro")
 
                 if intro_voice_spec:
                     intro_provider, _, intro_voice_choice, intro_speed, intro_steps = resolve_voice_choice(
@@ -1860,14 +1847,15 @@ def run_conversion_job(job: Job) -> None:
                     total_steps=int(supertonic_steps_override if supertonic_steps_override is not None else getattr(job, "supertonic_total_steps", 5)),
                 )
             else:
-                kokoro_pipeline = get_pipeline("kokoro")
-                segment_iter = kokoro_pipeline(
+                kokoro_backend = get_pipeline("kokoro")
+                segment_iter = kokoro_backend(
                     normalized,
                     voice=voice_choice,
                     speed=float(speed_override if speed_override is not None else job.speed),
                     split_pattern=split_pattern,
                 )
 
+            try:
                 for segment in segment_iter:
                     canceller()
                     graphemes_raw = getattr(segment, "graphemes", "") or ""
@@ -1907,6 +1895,11 @@ def run_conversion_job(job: Job) -> None:
                     if audio_sink:
                         current_time += duration
 
+            except OverflowError as exc:
+                job.add_log(
+                    f"Skipped chunk — number too large for TTS conversion: {exc}",
+                    level="warning",
+                )
             return local_segments
 
         def append_silence(
@@ -1950,8 +1943,8 @@ def run_conversion_job(job: Job) -> None:
             if chapter_provider == "kokoro":
                 voice_choice = voice_cache.get(chapter_cache_key)
                 if voice_choice is None:
-                    kokoro_pipeline = get_pipeline("kokoro")
-                    voice_choice = _resolve_voice(kokoro_pipeline, chapter_voice_resolved, job.use_gpu)
+                    kokoro_backend = get_pipeline("kokoro")
+                    voice_choice = _resolve_voice(kokoro_backend, chapter_voice_resolved, job.use_gpu)
                     voice_cache[chapter_cache_key] = voice_choice
             else:
                 voice_choice = chapter_voice_resolved
@@ -2095,9 +2088,9 @@ def run_conversion_job(job: Job) -> None:
                         if chunk_provider == "kokoro":
                             chunk_voice_choice = voice_cache.get(chunk_cache_key)
                             if chunk_voice_choice is None:
-                                kokoro_pipeline = get_pipeline("kokoro")
+                                kokoro_backend = get_pipeline("kokoro")
                                 chunk_voice_choice = _resolve_voice(
-                                    kokoro_pipeline,
+                                    kokoro_backend,
                                     chunk_voice_resolved,
                                     job.use_gpu,
                                 )
@@ -2239,8 +2232,8 @@ def run_conversion_job(job: Job) -> None:
                 if fallback_key and fallback_key != "__custom_mix":
                     # `voice_cache` keys are internal and include provider prefixes.
                     outro_voice_spec = fallback_key.split(":", 1)[-1]
-            if not outro_voice_spec and VOICES_INTERNAL:
-                outro_voice_spec = VOICES_INTERNAL[0]
+            if not outro_voice_spec:
+                outro_voice_spec = get_default_voice("kokoro")
 
             if outro_text and outro_voice_spec:
                 outro_start_time = current_time
@@ -2412,6 +2405,11 @@ def run_conversion_job(job: Job) -> None:
 
         # Explicitly release the pipeline and force garbage collection to prevent
         # memory accumulation in the worker process, which can lead to host lockups.
+        for p in pipelines.values():
+            try:
+                p.dispose()
+            except Exception:
+                pass
         pipelines.clear()
         pipeline = None
         gc.collect()
@@ -2445,17 +2443,14 @@ def _load_pipeline(job: Job):
     disable_gpu = not job.use_gpu or not cfg.get("use_gpu", True)
     provider = str(getattr(job, "tts_provider", "kokoro") or "kokoro").strip().lower()
     if provider == "supertonic":
-        return SupertonicPipeline(
-            sample_rate=SAMPLE_RATE,
-            auto_download=True,
-            total_steps=int(getattr(job, "supertonic_total_steps", 5) or 5),
+        return create_pipeline(
+            "supertonic",
         )
 
     device = "cpu"
     if not disable_gpu:
         device = _select_device()
-    _np, KPipeline = load_numpy_kpipeline()
-    return KPipeline(lang_code=job.language, repo_id="hexgrad/Kokoro-82M", device=device)
+    return create_pipeline("kokoro", lang_code=job.language, device=device)
 
 
 def _select_device() -> str:

abogen/webui/debug_tts_runner.py

@@ -15,7 +15,7 @@ from abogen.normalization_settings import build_apostrophe_config
 from abogen.text_extractor import extract_from_path
 from abogen.voice_cache import ensure_voice_assets
 from abogen.webui.conversion_runner import SAMPLE_RATE, SPLIT_PATTERN, _select_device, _to_float32, _resolve_voice, _spec_to_voice_ids
-from abogen.utils import load_numpy_kpipeline
+from abogen.tts_plugin.utils import create_pipeline
 
 
 _MARKER_RE = re.compile(re.escape(MARKER_PREFIX) + r"(?P<code>[A-Z0-9_]+)" + re.escape(MARKER_SUFFIX))
@@ -45,8 +45,7 @@ def _load_pipeline(language: str, use_gpu: bool) -> Any:
     device = "cpu"
     if use_gpu:
         device = _select_device()
-    _np, KPipeline = load_numpy_kpipeline()
-    return KPipeline(lang_code=language, repo_id="hexgrad/Kokoro-82M", device=device)
+    return create_pipeline("kokoro", lang_code=language, device=device)
 
 
 def _extract_cases_from_text(text: str) -> List[Tuple[str, str]]:
@@ -248,4 +247,8 @@ def run_debug_tts_wavs(
         "sample_rate": SAMPLE_RATE,
     }
     (run_dir / "manifest.json").write_text(json.dumps(manifest, indent=2), encoding="utf-8")
+    try:
+        pipeline.dispose()
+    except Exception:
+        pass
     return manifest

abogen/webui/routes/api.py

@@ -34,6 +34,7 @@ from abogen.normalization_settings import (
 )
 from abogen.llm_client import list_models, LLMClientError
 from abogen.kokoro_text_normalization import normalize_for_pipeline
+from abogen.tts_plugin.utils import is_plugin_registered
 from abogen.integrations.audiobookshelf import AudiobookshelfClient, AudiobookshelfConfig
 from abogen.integrations.calibre_opds import (
     CalibreOPDSClient,
@@ -63,7 +64,7 @@ def api_save_voice_profile() -> ResponseReturnValue:
     if profile is None:
         # Speaker Studio payload format
         provider = str(payload.get("provider") or "kokoro").strip().lower()
-        if provider not in {"kokoro", "supertonic"}:
+        if not is_plugin_registered(provider):
             provider = "kokoro"
         if provider == "supertonic":
             profile = {
@@ -230,7 +231,7 @@ def api_speaker_preview() -> ResponseReturnValue:
     use_gpu = settings.get("use_gpu", False)
 
     base_spec, speaker_name = split_profile_spec(voice)
-    resolved_provider = tts_provider if tts_provider in {"kokoro", "supertonic"} else ""
+    resolved_provider = tts_provider if is_plugin_registered(tts_provider) else ""
 
     if speaker_name:
         entry = normalize_profile_entry(load_profiles().get(speaker_name))

abogen/webui/routes/utils/form.py

@@ -7,6 +7,7 @@ from flask.typing import ResponseReturnValue
 
 from abogen.webui.service import PendingJob, JobStatus
 from abogen.webui.routes.utils.service import get_service
+from abogen.tts_plugin.utils import is_plugin_registered
 from abogen.webui.routes.utils.settings import (
     load_settings,
     coerce_bool,
@@ -32,7 +33,7 @@ from abogen.webui.routes.utils.common import split_profile_spec
 from abogen.utils import calculate_text_length
 from abogen.voice_profiles import serialize_profiles, normalize_profile_entry
 from abogen.chunking import ChunkLevel, build_chunks_for_chapters
-from abogen.constants import VOICES_INTERNAL
+from abogen.tts_plugin.utils import get_default_voice
 from abogen.speaker_configs import get_config
 from abogen.kokoro_text_normalization import normalize_roman_numeral_titles
 from dataclasses import dataclass
@@ -579,7 +580,7 @@ def apply_book_step_form(
     # spec (e.g. "speaker:Name" for saved speakers, or a Kokoro mix formula).
     # This enables mixed-provider conversions (e.g. narrator=SuperTonic, characters=Kokoro).
     provider_value = str(form.get("tts_provider") or "").strip().lower()
-    if provider_value in {"kokoro", "supertonic"}:
+    if is_plugin_registered(provider_value):
         pending.tts_provider = provider_value
 
     # Determine the base speaker selection (saved speaker ref or raw voice).
@@ -616,8 +617,8 @@ def apply_book_step_form(
         custom_formula = ""
 
     base_voice_spec = resolved_default_voice or narrator_voice_raw
-    if not base_voice_spec and VOICES_INTERNAL:
-        base_voice_spec = VOICES_INTERNAL[0]
+    if not base_voice_spec:
+        base_voice_spec = get_default_voice("kokoro")
 
     voice_choice, resolved_language, selected_profile = resolve_voice_choice(
         pending.language,
@@ -796,8 +797,8 @@ def build_pending_job_from_extraction(
         profile_selection = inferred_profile
 
     base_voice = base_voice_input or resolved_default_voice or str(default_voice_setting).strip()
-    if not base_voice and VOICES_INTERNAL:
-        base_voice = VOICES_INTERNAL[0]
+    if not base_voice:
+        base_voice = get_default_voice("kokoro")
     selected_speaker_config = (form.get("speaker_config") or "").strip()
     speaker_config_payload = get_config(selected_speaker_config) if selected_speaker_config else None
 

abogen/webui/routes/utils/preview.py

@@ -14,13 +14,57 @@ _preview_pipelines: Dict[Tuple[str, str], Any] = {}
 _preview_pipeline_lock = threading.Lock()
 
 
+def clear_preview_pipelines() -> None:
+    """Dispose all cached preview pipelines and clear the cache."""
+    with _preview_pipeline_lock:
+        for pipeline in _preview_pipelines.values():
+            try:
+                pipeline.dispose()
+            except Exception:
+                pass
+        _preview_pipelines.clear()
+
+
 def _select_device() -> str:
     import platform
 
+    try:
+        import torch  # type: ignore[import-not-found]
+    except Exception:
+        return "cpu"
+
     system = platform.system()
     if system == "Darwin" and platform.processor() == "arm":
+        try:
+            if torch.backends.mps.is_available():
                 return "mps"
+        except Exception:
+            pass
+        return "cpu"
+
+    try:
+        if torch.cuda.is_available():
             return "cuda"
+    except Exception:
+        pass
+    return "cpu"
+
+
+def _resolve_pipeline(language: str, use_gpu: bool) -> Tuple[Any, bool]:
+    devices: List[str] = ["cpu"]
+    if use_gpu:
+        preferred = _select_device()
+        if preferred != "cpu":
+            devices.insert(0, preferred)
+
+    last_error: Optional[Exception] = None
+    for device in devices:
+        try:
+            return get_preview_pipeline(language, device), device != "cpu"
+        except Exception as exc:
+            last_error = exc
+
+    raise RuntimeError("Preview pipeline is unavailable") from last_error
 
 
 def _to_float32(audio_segment) -> np.ndarray:
@@ -45,10 +89,9 @@ def get_preview_pipeline(language: str, device: str) -> Any:
         pipeline = _preview_pipelines.get(key)
         if pipeline is not None:
             return pipeline
-        from abogen.utils import load_numpy_kpipeline
+        from abogen.tts_plugin.utils import create_pipeline
 
-        _, KPipeline = load_numpy_kpipeline()
-        pipeline = KPipeline(lang_code=language, repo_id="hexgrad/Kokoro-82M", device=device)
+        pipeline = create_pipeline("kokoro", lang_code=language, device=device)
         _preview_pipelines[key] = pipeline
         return pipeline
 
@@ -104,9 +147,9 @@ def generate_preview_audio(
             normalized_text = source_text
 
     if provider == "supertonic":
-        from abogen.tts_supertonic import SupertonicPipeline
+        from abogen.tts_plugin.utils import create_pipeline
 
-        pipeline = SupertonicPipeline(sample_rate=SAMPLE_RATE, auto_download=True, total_steps=supertonic_total_steps)
+        pipeline = create_pipeline("supertonic")
         segments = pipeline(
             normalized_text,
             voice=voice_spec,
@@ -115,15 +158,7 @@ def generate_preview_audio(
             total_steps=supertonic_total_steps,
         )
     else:
-        device = "cpu"
-        if use_gpu:
-            try:
-                device = _select_device()
-            except Exception:
-                device = "cpu"
-                use_gpu = False
-
-        pipeline = get_preview_pipeline(language, device)
+        pipeline, pipeline_uses_gpu = _resolve_pipeline(language, use_gpu)
         if pipeline is None:
             raise RuntimeError("Preview pipeline is unavailable")
 
@@ -131,7 +166,7 @@ def generate_preview_audio(
         if voice_spec and "*" in voice_spec:
             from abogen.voice_formulas import get_new_voice
 
-            voice_choice = get_new_voice(pipeline, voice_spec, use_gpu)
+            voice_choice = get_new_voice(pipeline, voice_spec, pipeline_uses_gpu)
 
         segments = pipeline(
             normalized_text,

abogen/webui/routes/utils/settings.py

@@ -6,8 +6,8 @@ from abogen.constants import (
     LANGUAGE_DESCRIPTIONS,
     SUBTITLE_FORMATS,
     SUPPORTED_SOUND_FORMATS,
-    VOICES_INTERNAL,
 )
+from abogen.tts_plugin.utils import get_default_voice
 from abogen.normalization_settings import (
     DEFAULT_LLM_PROMPT,
     environment_llm_defaults,
@@ -174,7 +174,7 @@ def settings_defaults() -> Dict[str, Any]:
         "subtitle_format": "srt",
         "save_mode": "default_output" if has_output_override() else "save_next_to_input",
         "default_speaker": "",
-        "default_voice": VOICES_INTERNAL[0] if VOICES_INTERNAL else "",
+        "default_voice": get_default_voice("kokoro"),
         "supertonic_total_steps": 5,
         "supertonic_speed": 1.0,
         "replace_single_newlines": False,

abogen/webui/routes/utils/voice.py

@@ -17,10 +17,10 @@ from abogen.constants import (
     SUPPORTED_SOUND_FORMATS,
     SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION,
     SAMPLE_VOICE_TEXTS,
-    VOICES_INTERNAL,
 )
+from abogen.tts_plugin.utils import get_voices
 from abogen.speaker_configs import list_configs
-from abogen.utils import load_numpy_kpipeline
+from abogen.tts_plugin.utils import create_pipeline
 from abogen.webui.conversion_runner import _select_device, _to_float32, SAMPLE_RATE, SPLIT_PATTERN
 
 _preview_pipeline_lock = threading.RLock()
@@ -285,7 +285,7 @@ def filter_voice_catalog(
 def build_voice_catalog() -> List[Dict[str, str]]:
     catalog: List[Dict[str, str]] = []
     gender_map = {"f": "Female", "m": "Male"}
-    for voice_id in VOICES_INTERNAL:
+    for voice_id in get_voices("kokoro"):
         prefix, _, rest = voice_id.partition("_")
         language_code = prefix[0] if prefix else "a"
         gender_code = prefix[1] if len(prefix) > 1 else ""
@@ -590,7 +590,7 @@ def template_options() -> Dict[str, Any]:
     voice_catalog = build_voice_catalog()
     return {
         "languages": LANGUAGE_DESCRIPTIONS,
-        "voices": VOICES_INTERNAL,
+        "voices": get_voices("kokoro"),
         "subtitle_formats": SUBTITLE_FORMATS,
         "supported_langs_for_subs": SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION,
         "output_formats": SUPPORTED_SOUND_FORMATS,
@@ -741,8 +741,7 @@ def get_preview_pipeline(language: str, device: str):
         pipeline = _preview_pipelines.get(key)
         if pipeline is not None:
             return pipeline
-        _, KPipeline = load_numpy_kpipeline()
-        pipeline = KPipeline(lang_code=language, repo_id="hexgrad/Kokoro-82M", device=device)
+        pipeline = create_pipeline("kokoro", lang_code=language, device=device)
         _preview_pipelines[key] = pipeline
         return pipeline
 

abogen/webui/routes/voices.py

@@ -17,7 +17,7 @@ from abogen.speaker_configs import (
     save_configs,
     delete_config,
 )
-from abogen.constants import VOICES_INTERNAL
+
 
 voices_bp = Blueprint("voices", __name__)
 

abogen/webui/service.py

@@ -1609,10 +1609,12 @@ def build_service(
     output_root: Optional[Path] = None,
     uploads_root: Optional[Path] = None,
 ) -> ConversionService:
+    global _service_instance
     output_root = output_root or default_storage_root()
     service = ConversionService(
         output_root=output_root,
         uploads_root=uploads_root,
         runner=runner,
     )
+    _service_instance = service
     return service

abogen/word_substitution.py

@@ -0,0 +1,254 @@
+"""
+Word substitution module for text-to-speech preprocessing.
+
+This module provides functionality to:
+- Replace words/phrases with custom text
+- Convert ALL CAPS to lowercase
+- Convert numerals to words
+- Fix nonstandard punctuation for TTS compatibility
+
+All substitutions preserve special markers (chapter, voice, metadata, timestamps).
+"""
+
+import re
+
+from abogen.subtitle_utils import (
+    _CHAPTER_MARKER_PATTERN,
+    _VOICE_MARKER_PATTERN,
+    _METADATA_TAG_PATTERN,
+    _TIMESTAMP_ONLY_PATTERN,
+)
+
+
+def apply_word_substitutions(
+    text,
+    substitutions_list_str,
+    case_sensitive=False,
+    replace_all_caps=False,
+    replace_numerals=False,
+    fix_nonstandard_punctuation=False,
+):
+    """
+    Apply word substitutions to text while preserving markers.
+
+    Args:
+        text: Input text
+        substitutions_list_str: Newline-separated "Word|NewWord" pairs
+        case_sensitive: If True, match words case-sensitively
+        replace_all_caps: Convert ALL CAPS words to lowercase
+        replace_numerals: Convert numbers to words
+        fix_nonstandard_punctuation: Fix curly quotes, em/en dashes, etc.
+
+    Returns:
+        Modified text
+    """
+    # Apply nonstandard punctuation fixes FIRST (if enabled)
+    if fix_nonstandard_punctuation:
+        text = fix_punctuation(text)
+
+    # Parse substitutions list
+    substitutions = parse_substitutions_list(substitutions_list_str)
+
+    # Split text into segments (markers vs content)
+    segments = split_text_preserving_markers(text)
+
+    # Process each segment
+    processed_segments = []
+    for segment_type, segment_text in segments:
+        if segment_type == "marker":
+            # Preserve markers unchanged
+            processed_segments.append(segment_text)
+        else:
+            # Apply substitutions to content
+            processed_text = segment_text
+
+            # Apply word substitutions
+            if substitutions:
+                processed_text = apply_word_replacements(
+                    processed_text, substitutions, case_sensitive
+                )
+
+            # Apply ALL CAPS conversion
+            if replace_all_caps:
+                processed_text = convert_all_caps_to_lowercase(processed_text)
+
+            # Apply numeral conversion
+            if replace_numerals:
+                processed_text = convert_numerals_to_words(processed_text)
+
+            processed_segments.append(processed_text)
+
+    return "".join(processed_segments)
+
+
+def parse_substitutions_list(substitutions_str):
+    """
+    Parse newline-separated "Word|NewWord" format.
+
+    Args:
+        substitutions_str: String with substitutions, one per line
+
+    Returns:
+        List of tuples: [(word, replacement), ...]
+    """
+    substitutions = []
+    for line in substitutions_str.strip().split("\n"):
+        line = line.strip()
+        if not line or "|" not in line:
+            continue
+
+        parts = line.split("|", 1)
+        if len(parts) == 2:
+            word = parts[0].strip()
+            replacement = parts[1].strip()
+            if word:  # Only add if word is not empty
+                substitutions.append((word, replacement))
+
+    return substitutions
+
+
+def split_text_preserving_markers(text):
+    """
+    Split text into segments alternating between markers and content.
+
+    Args:
+        text: Input text with potential markers
+
+    Returns:
+        List of tuples: [("marker"|"content", text), ...]
+    """
+    # Combined pattern for all markers and timestamps
+    marker_pattern = re.compile(
+        r"(<<CHAPTER_MARKER:[^>]*>>|<<VOICE:[^>]*>>|<<METADATA_[^:]+:[^>]*>>|\d{1,2}:\d{2}:\d{2}(?:[.,]\d{1,3})?)"
+    )
+
+    segments = []
+    last_end = 0
+
+    for match in marker_pattern.finditer(text):
+        # Content before marker
+        if match.start() > last_end:
+            segments.append(("content", text[last_end : match.start()]))
+
+        # Marker itself
+        segments.append(("marker", match.group(0)))
+        last_end = match.end()
+
+    # Remaining content after last marker
+    if last_end < len(text):
+        segments.append(("content", text[last_end:]))
+
+    return segments
+
+
+def apply_word_replacements(text, substitutions, case_sensitive=False):
+    """
+    Apply word substitutions using whole-word matching.
+
+    Args:
+        text: Input text
+        substitutions: List of (word, replacement) tuples
+        case_sensitive: If True, match case-sensitively
+
+    Returns:
+        Text with substitutions applied
+    """
+    for word, replacement in substitutions:
+        # Use word boundaries for exact matching
+        # Escape special regex characters
+        escaped_word = re.escape(word)
+        pattern = re.compile(
+            r"\b" + escaped_word + r"\b",
+            0 if case_sensitive else re.IGNORECASE,
+        )
+        text = pattern.sub(replacement, text)
+
+    return text
+
+
+def convert_all_caps_to_lowercase(text):
+    """
+    Convert ALL CAPS words to lowercase.
+
+    Args:
+        text: Input text
+
+    Returns:
+        Text with ALL CAPS converted to lowercase
+    """
+
+    def replace_caps(match):
+        word = match.group(0)
+        # Convert to lowercase
+        return word.lower()
+
+    # Match words that are ALL CAPS (2+ letters)
+    pattern = re.compile(r"\b[A-Z]{2,}\b")
+    return pattern.sub(replace_caps, text)
+
+
+def convert_numerals_to_words(text):
+    """
+    Convert numerals to words using num2words library.
+
+    Args:
+        text: Input text
+
+    Returns:
+        Text with numerals converted to words
+    """
+    try:
+        from num2words import num2words
+    except ImportError:
+        # If num2words not available, return unchanged
+        return text
+
+    def replace_number(match):
+        try:
+            number = int(match.group(0))
+            # Convert to words in English
+            return num2words(number)
+        except Exception:
+            # If conversion fails, return original
+            return match.group(0)
+
+    # Match integers (but not timestamps or other patterns)
+    # Negative lookbehind/ahead to avoid timestamps
+    pattern = re.compile(r"(?<!\d:)\b\d+\b(?!:\d)")
+    return pattern.sub(replace_number, text)
+
+
+def fix_punctuation(text):
+    """
+    Convert nonstandard punctuation to standard equivalents.
+
+    This helps TTS engines pronounce words correctly by converting:
+    - Curly quotes to straight quotes
+    - Ellipsis to three periods
+
+    Args:
+        text: Input text
+
+    Returns:
+        Text with nonstandard punctuation fixed
+    """
+    # Define replacements
+    replacements = {
+        # Curly double quotes
+        "\u201c": '"',  # Left double quotation mark
+        "\u201d": '"',  # Right double quotation mark
+        "\u201e": '"',  # Double low-9 quotation mark
+        # Curly single quotes
+        "\u2018": "'",  # Left single quotation mark
+        "\u2019": "'",  # Right single quotation mark
+        "\u201a": "'",  # Single low-9 quotation mark
+        "\u201b": "'",  # Single high-reversed-9 quotation mark
+        # Other punctuation
+        "\u2026": "...",  # Ellipsis
+    }
+
+    # Apply all replacements
+    for old_char, new_char in replacements.items():
+        text = text.replace(old_char, new_char)
+
+    return text

docs/contributing.md

@@ -0,0 +1,55 @@
+# Contributing to Abogen
+
+We welcome contributions to Abogen!
+
+## How to Contribute
+
+1. Fork the repository
+2. Create a branch for your feature
+3. Make your changes
+4. Write tests
+5. Submit a pull request
+
+## Code Standards
+
+- Follow PEP 8 for Python
+- Use TypeScript for JavaScript
+- Type hints required for new Python code
+- Document complex logic with comments
+
+## Plugin Architecture
+
+When contributing TTS engines, implement the **Plugin Architecture** contract.
+
+See [Developer Guide](developer-guide.md#5-adding-a-new-plugin) for:
+- Required exports (`PLUGIN_MANIFEST`, `MODEL_REQUIREMENTS`, `create_engine`)
+- Engine / EngineSession contracts
+- Capability interfaces (`VoiceLister`, `PreviewGenerator`, etc.)
+- Step-by-step plugin creation guide
+
+## Testing
+
+```bash
+# All tests
+pytest
+
+# Contract tests (architectural compliance)
+pytest tests/contracts/
+
+# Behavioral regression tests
+pytest tests/test_behavioral_regression.py
+```
+
+## Documentation
+
+- Update relevant docs in `docs/` when changing architecture or APIs
+- Add docstrings to all public functions/classes
+- Follow existing documentation style
+
+## Pull Request Checklist
+
+- [ ] Tests pass (`pytest`)
+- [ ] Code follows style guide (`ruff check`, `ruff format`)
+- [ ] Documentation updated
+- [ ] No legacy architecture references (`TTSBackend`, `register_backend`, `TTSBackendRegistry`)
+- [ ] Uses new Plugin Architecture patterns

docs/developer-guide.md

@@ -0,0 +1,270 @@
+# TTS Plugin Architecture — Architectural Reference
+
+This document describes the **stable architectural contracts** of the TTS Plugin Architecture. It documents invariants that only change when the architecture itself changes.
+
+---
+
+## 1. Architecture Overview
+
+```
+┌─────────────────────────────────────────────────────────────────┐
+│                        Host Application                         │
+│  ┌─────────────┐  ┌──────────────┐  ┌────────────────────────┐  │
+│  │ Plugin      │  │ HostContext  │  │ Plugin Discovery       │  │
+│  │ Manager     │──│ (config_dir, │  │ (plugin directories)   │  │
+│  │             │  │  logger,     │  │                        │  │
+│  │ - discover  │  │  http_client)│  └────────────────────────┘  │
+│  │ - validate  │  └──────────────┘            │                 │
+│  │ - activate  │                              ▼                 │
+│  │ - dispose   │  ┌─────────────────────────────────────────┐   │
+│  └──────┬──────┘  │              Plugin Package             │   │
+│         │         │  ┌──────────────┐ ┌─────────────────┐   │   │
+│         ▼         │  │ PLUGIN_      │ │ MODEL_          │   │   │
+│  ┌────────────┐   │  │ MANIFEST     │ │ REQUIREMENTS    │   │   │
+│  │   Engine   │◄──┤  │ create_engine│ │                 │   │   │
+│  └──────┬─────┘   │  └──────────────┘ └─────────────────┘   │   │
+│         │         └─────────────────────────────────────────┘   │
+│         │ createSession()                                       │
+│         ▼                                                       │
+│  ┌─────────────┐                                                │
+│  │EngineSession│                                                │
+│  └──────┬──────┘                                                │
+│         │ synthesize()                                          │
+│         ▼                                                       │
+│  ┌────────────────┐                                             │
+│  │SynthesizedAudio│                                             │
+│  └────────────────┘                                             │
+└─────────────────────────────────────────────────────────────────┘
+```
+
+### Core Components
+
+| Component | Responsibility |
+|-----------|----------------|
+| **PluginManifest** | Static metadata: id, name, version, api_version, capabilities, engine manifest |
+| **EngineManifest** | Voice sources, parameters, audio formats |
+| **HostContext** | Minimal host services: config_dir, logger, http_client |
+| **Engine** | Stateless factory for sessions; thread-safe `createSession()` |
+| **EngineSession** | Owns mutable execution state; not thread-safe |
+| **PluginManager** | Discovers, validates, and manages plugin lifecycle |
+| **Capabilities** | Optional interfaces: VoiceLister, PreviewGenerator, StreamingSynthesizer, CancelableSession |
+
+---
+
+## 2. Ownership Model
+
+### Engine Ownership
+```
+PluginManager.create_engine() → Engine
+```
+- **PluginManager** creates and caches engines
+- **Caller** receives `Engine` instance
+- **Caller** must dispose all sessions **before** disposing engine
+- **Engine.dispose()** releases engine resources
+- After `Engine.dispose()`: all methods except `dispose()` raise `EngineError`
+
+### Session Ownership
+```
+Engine.createSession() → EngineSession
+```
+- **Engine** creates session
+- **Ownership transfers to caller** immediately
+- **Caller** is responsible for `session.dispose()`
+- **Engine does NOT track sessions** — no registry, no callbacks
+- After `session.dispose()`: all methods except `dispose()` raise `EngineError`
+
+### Disposal Order (Invariant)
+```python
+# Correct
+engine = manager.create_engine("id")
+session = engine.createSession()
+try:
+    audio = session.synthesize(request)
+finally:
+    session.dispose()      # 1. Sessions FIRST
+engine.dispose()           # 2. Then engine
+
+# INCORRECT — violates contract (undefined behavior)
+engine.dispose()
+session.synthesize(request)  # EngineError
+```
+
+---
+
+## 3. Lifecycle State Machine
+
+```
+DISCOVERY
+  PluginManager.discover(plugin_dirs)
+  → Loads PLUGIN_MANIFEST, MODEL_REQUIREMENTS
+  → Validates api_version (major must match)
+  → Validates declared capabilities are implemented
+
+MODEL_DOWNLOAD (if MODEL_REQUIREMENTS non-empty)
+  Host reads MODEL_REQUIREMENTS
+  Downloads/caches models
+  Resolves model_path
+
+ACTIVATION
+  create_engine(context, model_path, config)
+  → Atomic: succeeds fully or raises EngineError
+  → Returns Engine
+
+SESSION_CREATION
+  engine.createSession() → EngineSession
+  → Ownership transfers to caller
+  → Raises EngineError on failure
+  → Never returns partial session
+
+SYNTHESIS
+  session.synthesize(request)
+  → Returns SynthesizedAudio
+  → Raises EngineError on failure
+  → Session remains usable after error
+
+SESSION_DISPOSAL
+  session.dispose()
+  → Idempotent, never raises
+  → After: all methods raise EngineError
+
+DEACTIVATION
+  engine.dispose()
+  → Caller MUST dispose all sessions first
+  → Idempotent, never raises
+  → After: all methods raise EngineError
+```
+
+---
+
+## 4. Protocol Contracts
+
+### Engine (Protocol)
+```python
+@runtime_checkable
+class Engine(Protocol):
+    def createSession(self) -> EngineSession:
+        """Create a new session. Thread-safe. Transfers ownership."""
+        ...
+
+    def dispose(self) -> None:
+        """Release engine resources.
+        Caller must dispose all sessions first.
+        Idempotent, never raises.
+        After: all methods except dispose() raise EngineError."""
+        ...
+```
+
+### EngineSession (Protocol)
+```python
+@runtime_checkable
+class EngineSession(Protocol):
+    def synthesize(self, request: SynthesisRequest) -> SynthesizedAudio:
+        """Synthesize audio.
+        Returns SynthesizedAudio or raises EngineError.
+        Session remains usable after error."""
+        ...
+
+    def dispose(self) -> None:
+        """Release session resources.
+        Idempotent, never raises.
+        After: all methods except dispose() raise EngineError."""
+        ...
+```
+
+### Capability Protocols (Optional)
+- **VoiceLister**: `listVoices(source_id: str) -> list[VoiceManifest]`
+- **PreviewGenerator**: `generatePreview(voice: VoiceSelection, text: str) -> SynthesizedAudio`
+- **StreamingSynthesizer**: `synthesizeStream(request: SynthesisRequest) -> Iterator[bytes]`
+- **CancelableSession**: `cancel() -> None` (causes in-flight synthesize to raise `CancelledError`)
+
+---
+
+## 5. Error Semantics
+
+```
+EngineError (base)
+├── ModelNotFoundError      # Required model not found
+├── ModelLoadError          # Model failed to load
+├── NetworkError            # Network operation failed
+├── InvalidInputError       # Request validation failed
+├── ConfigurationError      # Invalid configuration
+├── CancelledError          # Operation cancelled via CancelableSession
+└── InternalError           # Unexpected internal failure
+```
+
+### When Each Is Raised
+| Error | Raised By | Conditions |
+|-------|-----------|------------|
+| `ModelNotFoundError` | `create_engine()` | Required model not found at `model_path` |
+| `ModelLoadError` | `create_engine()` | Model exists but fails to load |
+| `NetworkError` | `synthesize()`, `create_engine()` | Network call fails (cloud engines) |
+| `InvalidInputError` | `synthesize()` | Request validation fails (empty text, invalid voice, etc.) |
+| `ConfigurationError` | `create_engine()` | Config values invalid for this engine |
+| `CancelledError` | `synthesize()`, `synthesizeStream()` | `CancelableSession.cancel()` called |
+| `InternalError` | Any | Unexpected internal failure (bug) |
+
+### Dispose Contract
+- `dispose()` is **idempotent** and **never raises**
+- After `dispose()`: all methods except `dispose()` raise `EngineError`
+- Engine: caller must dispose all sessions first; violating this is undefined behavior
+
+---
+
+## 6. Capabilities
+
+| Capability | Interface | Enables |
+|------------|-----------|---------|
+| `voice_list` | `VoiceLister` | `listVoices(source_id)` — enumerate available voices |
+| `preview` | `PreviewGenerator` | `generatePreview(voice, text)` — preview without session |
+| `streaming` | `StreamingSynthesizer` | `synthesizeStream(request)` — chunked audio output |
+| `cancel` | `CancelableSession` | `cancel()` — interrupt in-flight synthesis |
+
+Plugins declare capabilities in `PluginManifest.capabilities`. Host validates at load time.
+
+---
+
+## 7. Contract Tests
+
+**Location**: `tests/contracts/`
+
+**Purpose**: Verify every plugin satisfies the architectural contracts.
+
+**Guarantees**:
+- Required exports exist (`PLUGIN_MANIFEST`, `MODEL_REQUIREMENTS`, `create_engine`)
+- `create_engine` is atomic
+- `Engine.createSession()` transfers ownership, never returns partial
+- `dispose()` is idempotent on Engine and EngineSession
+- After `dispose()`, methods raise `EngineError`
+- `synthesize()` raises typed `EngineError` subtypes, session remains usable
+- Declared capabilities are actually implemented
+- Plugin loader validates manifest, api_version, capabilities
+
+**Run**: `pytest tests/contracts/ -v`
+
+---
+
+## 8. Behavioral Tests
+
+**Location**: `tests/test_behavioral_regression.py`
+
+**Purpose**: Verify user-facing behavior via public API only (`create_pipeline`, `Engine`, `EngineSession`, `PluginManager`).
+
+**Scope**:
+- Synthesis with various inputs (short, long, empty, unicode, mixed scripts)
+- Voice selection and listing
+- Parameter handling (speed, etc.)
+- Error scenarios (unknown plugin, disposal, etc.)
+- Resource cleanup (dispose idempotency, no leaks)
+- Pipeline utility (`create_pipeline`)
+
+**Run**: `pytest tests/test_behavioral_regression.py -v`
+
+---
+
+## 9. Reference
+
+- **Architecture Spec**: `docs/architecture-final-v2.md`
+- **Amendment (lang_code)**: `docs/architecture-amendment-001.md`
+- **Migration Roadmap**: `docs/migration-roadmap.md`
+- **Plugin Examples**: `plugins/kokoro/`, `plugins/supertonic/`
+- **Protocol Definitions**: `abogen/tts_plugin/engine.py`, `abogen/tts_plugin/capabilities.py`

docs/getting-started.md

@@ -0,0 +1,68 @@
+# Getting Started
+
+Quickstart for developers working on Abogen.
+
+## Prerequisites
+
+- Python 3.10+
+- Node.js 20+
+- npm 10+
+- Git
+- Docker (optional)
+
+## Installation
+
+```bash
+# Development install with all extras
+pip install -e .[dev]
+
+# Or with uv
+uv pip install -e .[dev]
+```
+
+## Running the Application
+
+```bash
+# Desktop GUI
+abogen
+
+# Web UI
+abogen-web
+
+# CLI
+abogen-cli
+```
+
+## Project Structure
+
+```
+abogen/
+├── pyqt/         - PyQt6 desktop GUI
+├── webui/        - Flask web UI
+├── tts_plugin/   - Plugin Architecture (Engine, EngineSession, Manifest)
+└── plugins/      - Built-in plugins (kokoro, supertonic)
+tests/
+├── contracts/    - Contract compliance tests
+└── ...
+```
+
+## Testing
+
+```bash
+# All tests
+pytest
+
+# Contract tests (architectural compliance)
+pytest tests/contracts/
+
+# Behavioral regression tests
+pytest tests/test_behavioral_regression.py
+```
+
+## Architecture
+
+See [Developer Guide](developer-guide.md) for Plugin Architecture details:
+- Engine / EngineSession lifecycle
+- Plugin contract (PLUGIN_MANIFEST, create_engine)
+- Adding new plugins
+- Capability interfaces

docs/testing.md

@@ -0,0 +1,269 @@
+# Testing Guide
+
+This document describes the testing strategy for Abogen's Plugin Architecture.
+
+## Test Categories
+
+### 0. Auto-Discovery Plugin Tests (`tests/plugins/`)
+
+**Purpose**: Automatically test every plugin in `plugins/` directory without manual test creation. These tests use discovery to find all plugins and run generic tests against each one.
+
+**What They Test**:
+- **Manifest structure**: Required fields, API version format, voices field
+- **Engine lifecycle**: `create_engine`, `dispose` idempotency, post-dispose behavior
+- **Capability implementation**: Declared capabilities are implemented (e.g., `voice_list` → `VoiceLister`)
+
+**How Auto-Discovery Works**:
+```python
+# tests/plugins/conftest.py
+@pytest.fixture(scope="module")
+def plugin_ids(plugins_dir: Path) -> list[str]:
+    """Discovers all plugin directories with __init__.py"""
+    return [item.name for item in plugins_dir.iterdir() 
+            if item.is_dir() and (item / "__init__.py").exists()]
+```
+
+**Test Structure**:
+```
+tests/plugins/
+├── conftest.py              # Fixtures: plugin_ids, loaded_plugin, host_context
+└── test_all_plugins.py      # Generic tests for every plugin
+    ├── TestAllPluginsManifest
+    ├── TestAllPluginsEngine
+    └── TestAllPluginsCapabilities
+```
+
+**Running Auto-Discovery Tests**:
+```bash
+# Test all plugins automatically
+pytest tests/plugins/ -v
+
+# Test specific plugin
+pytest tests/plugins/ -v -k "kokoro"
+
+# See which plugins were discovered
+pytest tests/plugins/ --collect-only
+```
+
+**Adding a New Plugin**:
+1. Create plugin directory: `plugins/my_plugin/`
+2. Add `__init__.py` with `PLUGIN_MANIFEST`, `MODEL_REQUIREMENTS`, `create_engine`
+3. Run `pytest tests/plugins/` — tests automatically discover and test your plugin!
+
+**When to Add Plugin-Specific Tests**:
+Auto-discovery tests cover generic contract validation. Create plugin-specific tests in `tests/test_<plugin>_plugin.py` for:
+- Integration with real dependencies (e.g., KPipeline for Kokoro)
+- Specific voice IDs and behavior
+- Plugin-specific parameters and features
+
+---
+
+### 1. Contract Tests (`tests/contracts/`)
+
+**Purpose**: Verify that every plugin satisfies the architectural contract. These tests ensure the Plugin Architecture's invariants are maintained.
+
+**What They Guarantee**:
+- Every plugin exports `PLUGIN_MANIFEST`, `MODEL_REQUIREMENTS`, `create_engine`
+- `create_engine` is atomic (succeeds fully or raises and cleans up)
+- `Engine.createSession()` returns valid `EngineSession`, transfers ownership
+- `Engine.dispose()` is idempotent, never raises
+- After `dispose()`, all methods raise `EngineError`
+- `EngineSession.synthesize()` returns `SynthesizedAudio` or raises `EngineError` (session remains usable)
+- `EngineSession.dispose()` is idempotent, never raises
+- Capability interfaces (`VoiceLister`, `PreviewGenerator`, etc.) are correctly implemented
+- Plugin Loader discovers, validates, and loads plugins correctly
+- Plugin Manager creates, caches, and disposes engines correctly
+- Value objects are immutable and have correct equality semantics
+- Error hierarchy is preserved (`EngineError` base with subtypes)
+
+**Why They Exist**:
+- Provide **compile-time-like guarantees** for a dynamic plugin system
+- Enable **safe plugin ecosystem** — host can trust any loaded plugin
+- Catch **architectural violations** early (missing dispose, wrong return types, etc.)
+- Document the **contract** in executable form
+
+**What Every New Plugin Must Pass**:
+```bash
+pytest tests/contracts/ -v
+# All tests must pass
+```
+
+**Running Contract Tests**:
+```bash
+# All contract tests
+pytest tests/contracts/
+
+# Specific contract
+pytest tests/contracts/test_engine_contract.py
+
+# With coverage
+pytest tests/contracts/ --cov=abogen.tts_plugin
+```
+
+---
+
+### 2. Behavioral Tests (`tests/test_behavioral_regression.py`)
+
+**Purpose**: Verify external user-facing behavior using only public API. These tests are **not coupled to internal implementation**.
+
+**What They Test**:
+- Synthesis with various inputs (short, long, empty, unicode, mixed scripts)
+- Voice selection and listing
+- Parameter handling (speed, etc.)
+- Error scenarios (unknown plugin, disposal, etc.)
+- Resource cleanup (dispose idempotency, no leaks)
+- Pipeline utility (`create_pipeline`)
+
+**Why They Test Public Behavior Only**:
+- **Refactoring safety**: Internal changes don't break tests
+- **Real-world usage**: Tests match how consumers actually use the API
+- **Plugin agnostic**: Parametrized across Kokoro, SuperTonic, and mock plugins
+- **Regression detection**: Catch behavioral regressions regardless of implementation
+
+**What They Don't Test**:
+- Internal class structure
+- Private methods
+- Implementation details (how audio is generated, model loading internals)
+
+**Running Behavioral Tests**:
+```bash
+# All behavioral tests
+pytest tests/test_behavioral_regression.py -v
+
+# With specific plugin (if installed)
+pytest tests/test_behavioral_regression.py -v -k "kokoro"
+```
+
+---
+
+### 4. Unit Tests (`tests/`)
+
+**Purpose**: Test individual modules in isolation.
+
+**Examples**:
+- `test_book_parser.py` — EPUB/PDF/text parsing
+- `test_text_normalization.py` — Text preprocessing
+- `test_chunk_helpers.py` — Text chunking logic
+- `test_voice_cache.py` — Voice caching
+
+---
+
+### 5. Integration Tests
+
+**Purpose**: Test cross-component interactions.
+
+**Examples**:
+- `test_kokoro_plugin.py` — Full Kokoro plugin integration
+- `test_supertonic_plugin.py` — Full SuperTonic plugin integration
+- `test_conversion_series.py` — End-to-end conversion pipeline
+
+---
+
+## Test Architecture
+
+```
+tests/
+├── contracts/           # Contract tests (architectural compliance)
+│   ├── conftest.py      # Shared fixtures (FakeEngine, FakeSession)
+│   ├── test_manifest_contract.py
+│   ├── test_plugin_contract.py
+│   ├── test_engine_contract.py
+│   ├── test_session_contract.py
+│   ├── test_capabilities_contract.py
+│   ├── test_loader_contract.py
+│   ├── test_plugin_manager_contract.py
+│   ├── test_types_contract.py
+│   ├── test_errors_contract.py
+│   ├── test_host_context_contract.py
+│   └── test_integration.py
+├── test_behavioral_regression.py  # Behavioral tests (public API)
+├── test_kokoro_plugin.py          # Kokoro integration
+├── test_supertonic_plugin.py      # SuperTonic integration
+└── ...                            # Other unit/integration tests
+```
+
+---
+
+## Adding Tests for a New Plugin
+
+### Auto-Discovery Tests (Automatic!)
+
+**No manual test creation required!** When you add a new plugin to `plugins/`:
+
+1. Create plugin directory: `plugins/my_plugin/`
+2. Add `__init__.py` with required exports:
+   ```python
+   PLUGIN_MANIFEST = PluginManifest(...)
+   MODEL_REQUIREMENTS = [...]
+   def create_engine(...): ...
+   ```
+3. Run `pytest tests/plugins/` — auto-discovery tests automatically find and test your plugin!
+
+**What's Tested Automatically**:
+- Manifest structure and required fields
+- API version compatibility
+- Engine creation and dispose contract
+- Capability implementation (if declared)
+
+### Plugin-Specific Tests (Optional)
+
+Create `tests/test_my_plugin_plugin.py` for:
+- Integration with real backend (e.g., KPipeline for Kokoro)
+- Specific voice IDs and behavior
+- Plugin-specific parameters and features
+
+### Contract Tests (Deprecated for New Plugins)
+
+**Note**: Auto-discovery tests (`tests/plugins/`) now cover contract validation for all plugins. Manual contract tests in `tests/contracts/` are only needed for testing internal architecture components.
+
+### Behavioral Tests (Recommended)
+
+Add parametrized tests to `tests/test_behavioral_regression.py`:
+
+```python
+# In _plugin_ids list, add your plugin
+_plugin_ids = ["kokoro", "supertonic", "my_plugin"]
+_plugin_engines["my_plugin"] = _YourMockEngine
+_plugin_default_voices["my_plugin"] = "voice1"
+_plugin_all_voices["my_plugin"] = ["voice1", "voice2"]
+```
+
+All existing behavioral tests will automatically run against your plugin.
+
+---
+
+## Continuous Integration
+
+```yaml
+# .github/workflows/test.yml
+- name: Contract Tests
+  run: pytest tests/contracts/ -v
+
+- name: Behavioral Tests
+  run: pytest tests/test_behavioral_regression.py -v
+
+- name: Unit & Integration Tests
+  run: pytest tests/ -v --ignore=tests/test_behavioral_regression.py
+```
+
+---
+
+## Test Design Principles
+
+### Contract Tests
+- **No mocks** for the system under test (test real plugin loading)
+- **Strict assertions** on types and behavior
+- **Document architecture** in test names and docstrings
+- **Fail fast** on architectural violations
+
+### Behavioral Tests
+- **Only public API** (`create_pipeline`, `Engine`, `EngineSession`, `PluginManager`)
+- **Parametrized** across plugins
+- **Realistic scenarios** (long text, unicode, mixed scripts)
+- **No implementation coupling** (test behavior, not internals)
+
+### General
+- **Fast**: Unit tests < 1s, Contract tests < 5s, Behavioral < 30s
+- **Isolated**: No shared state between tests
+- **Deterministic**: Same input → same output
+- **Descriptive names**: `test_<component>_<scenario>_<expected>`

docs/tts-plugin-architecture.md

@@ -0,0 +1,689 @@
+# TTS Plugin Architecture — Final Specification
+
+## 1. Core Domain
+
+Zero dependencies. Pure business logic.
+
+### 1.1 Engine
+
+Factory for sessions. Stateless. Thread-safe for createSession().
+
+```
+interface Engine:
+  createSession() -> EngineSession
+  dispose() -> void
+```
+
+**createSession() contract**:
+- Returns: EngineSession
+- Raises: EngineError on failure
+- Ownership: Transfers to caller
+- Thread-safe: Yes
+
+**dispose() contract**:
+- Releases engine resources
+- Caller must ensure all sessions created by this engine are disposed before calling dispose()
+- Disposing an engine while any session is still alive violates the API contract; behavior is undefined
+- Idempotent: Safe to call multiple times
+- Never raises: Catches and logs internally
+- After dispose(): All methods except dispose() raise EngineError
+
+### 1.2 EngineSession
+
+Owns mutable execution state isolated from other concurrent work. NOT thread-safe.
+
+```
+interface EngineSession:
+  synthesize(request: SynthesisRequest) -> SynthesizedAudio
+  dispose() -> void
+```
+
+**synthesize() contract**:
+- Returns: SynthesizedAudio
+- Raises: EngineError on failure (session remains usable)
+- Thread-safe: No
+
+**dispose() contract**:
+- Releases session resources
+- Idempotent: Safe to call multiple times
+- Never raises: Catches and logs internally
+- After dispose(): All methods except dispose() raise EngineError
+
+### 1.3 SynthesisRequest
+
+Immutable value object.
+
+```
+SynthesisRequest:
+  text: string
+  voice: VoiceSelection
+  parameters: ParameterValues
+  format: AudioFormat
+```
+
+### 1.4 SynthesizedAudio
+
+Immutable value object.
+
+```
+SynthesizedAudio:
+  data: bytes
+  format: AudioFormat
+  duration: Duration
+```
+
+### 1.5 VoiceSelection
+
+Immutable value object. Opaque to engine.
+
+```
+VoiceSelection:
+  source: string
+  key: string
+  payload: any = None  # Optional; required for clone/blend sources
+```
+
+### 1.6 ParameterValues
+
+Immutable value object. Behaves like Mapping[str, Any].
+
+```
+ParameterValues:
+  values: Mapping[str, Any]
+```
+
+### 1.7 AudioFormat
+
+Immutable value object.
+
+```
+AudioFormat:
+  mime: string
+  extension: string
+```
+
+### 1.8 Duration
+
+Immutable value object.
+
+```
+Duration:
+  seconds: number
+```
+
+### 1.9 EngineConfig
+
+Engine initialization settings only. No resource references.
+
+```
+EngineConfig:
+  device: string    # "cpu", "cuda:0", etc.
+  # Engine-specific settings (if any)
+  # Unknown keys are ignored (no error)
+```
+
+---
+
+## 2. Error Hierarchy
+
+Typed exceptions. Engines raise EngineError or subtypes. Never raw exceptions.
+
+```
+EngineError (base)
+├── ModelNotFoundError
+├── ModelLoadError
+├── NetworkError
+├── InvalidInputError
+├── ConfigurationError
+├── CancelledError
+└── InternalError
+```
+
+**Contract**:
+- synthesize() raises EngineError on failure, session remains usable
+- dispose() never raises (catches and logs internally)
+- create_engine() raises EngineError on failure, cleans up partially created resources
+- createSession() raises EngineError on failure, no partially initialized session returned
+- cancel() causes synthesize() to raise CancelledError
+
+---
+
+## 3. Capability Interfaces (Optional)
+
+Engines implement only what they support. Capabilities are additive.
+
+### 3.1 VoiceLister
+
+```
+interface VoiceLister:
+  listVoices(sourceId: string) -> list[VoiceManifest]
+```
+
+### 3.2 PreviewGenerator
+
+```
+interface PreviewGenerator:
+  generatePreview(voice: VoiceSelection, text: string) -> SynthesizedAudio
+```
+
+### 3.3 ModelRequirements
+
+Static at plugin level, not engine level. Host reads before creating engine.
+
+```
+MODEL_REQUIREMENTS = list[ModelManifest]
+```
+
+### 3.4 StreamingSynthesizer
+
+Optional capability of EngineSession, not Engine.
+
+```
+interface StreamingSynthesizer:
+  synthesizeStream(request: SynthesisRequest) -> Iterator[bytes]
+```
+
+**Iterator contract**:
+- Yields audio chunks as they become available
+- Raises CancelledError if cancel() is called during iteration
+- Raises EngineError on synthesis failure
+- Iterator exhaustion = synthesis complete
+- Session remains usable after iterator completes
+
+### 3.5 CancelableSession
+
+Optional capability for engines that support cancellation.
+
+```
+interface CancelableSession:
+  cancel() -> void
+```
+
+**cancel() contract**:
+- Cancels in-progress synthesize()
+- synthesize() raises CancelledError (subtype of EngineError)
+- EngineSession remains usable after cancellation (unless implementation documents otherwise)
+
+---
+
+## 4. Plugin Manifest
+
+Static metadata. Immutable. No dependencies.
+
+### 4.1 PluginManifest
+
+```
+PluginManifest:
+  id: string
+  name: string
+  version: string
+  api_version: string  # semver format: MAJOR.MINOR
+  description: string
+  author: string
+  capabilities: list[string]
+  requires: RequirementManifest
+  engine: EngineManifest
+```
+
+**api_version contract**:
+- Format: semver (MAJOR.MINOR)
+- Compatibility: Host rejects plugin if major version differs
+- Minor version: backward compatible, Host accepts higher minor
+
+### 4.2 EngineManifest
+
+```
+EngineManifest:
+  voiceSources: list[VoiceSourceManifest]
+  parameters: list[ParameterManifest]
+  audioFormats: list[AudioFormatManifest]
+```
+
+### 4.3 VoiceSourceManifest
+
+```
+VoiceSourceManifest:
+  id: string
+  name: string
+  type: string    # "list", "speaker_id", "clone", "blend", "generate", "none"
+  config: any
+```
+
+### 4.4 VoiceManifest
+
+```
+VoiceManifest:
+  id: string
+  name: string
+  tags: list[string]
+```
+
+### 4.5 ParameterManifest
+
+```
+ParameterManifest:
+  id: string
+  name: string
+  description: string
+  type: string    # "float", "int", "string", "boolean", "enum"
+  default: any
+  min: number (optional)
+  max: number (optional)
+  step: number (optional)
+  options: list[EnumOption] (optional)
+  unit: string (optional)
+  group: string (optional)
+```
+
+### 4.6 AudioFormatManifest
+
+```
+AudioFormatManifest:
+  mime: string
+  extension: string
+```
+
+### 4.7 EnumOption
+
+```
+EnumOption:
+  value: string
+  label: string
+```
+
+### 4.8 RequirementManifest
+
+```
+RequirementManifest:
+  gpu: GpuRequirement (optional)
+  memory: number (optional)
+  internet: boolean (optional)
+```
+
+### 4.9 GpuRequirement
+
+```
+GpuRequirement:
+  required: boolean
+  type: string (optional)
+  memory: number (optional)
+```
+
+### 4.10 ModelManifest
+
+```
+ModelManifest:
+  id: string
+  name: string
+  size: string
+```
+
+---
+
+## 5. Host Services
+
+### 5.1 HostContext
+
+Minimal. 3 fields maximum. No business logic.
+
+```
+HostContext:
+  config_dir: Path      # For API keys, preferences
+  logger: Logger        # For logging
+  http_client: HttpClient  # For network requests
+```
+
+---
+
+## 6. Plugin Contract
+
+### 6.1 Plugin Exports
+
+```python
+# plugins/kokoro/__init__.py
+
+PLUGIN_MANIFEST = PluginManifest(...)
+MODEL_REQUIREMENTS = [...]  # Static at plugin level
+
+def create_engine(
+  context: HostContext,
+  model_path: Path | None,
+  config: EngineConfig
+) -> Engine:
+  """Create engine. Atomic: succeeds fully or raises and cleans up."""
+  ...
+```
+
+### 6.2 create_engine() Contract
+
+- Parameters:
+  - context: HostContext (host services)
+  - model_path: Path | None (resolved model path, or None for cloud/no-model engines)
+  - config: EngineConfig (engine initialization settings)
+- Returns: Engine
+- Raises: EngineError on failure
+- Atomic: Succeeds fully or cleans up and raises
+- Thread-safe: Can be called from any thread
+
+---
+
+## 7. Object Lifecycle
+
+### 7.1 Engine Lifecycle
+
+```
+1. DISCOVERY
+   Host scans plugin directories
+   Loads PLUGIN_MANIFEST and MODEL_REQUIREMENTS
+
+2. MODEL DOWNLOAD (if MODEL_REQUIREMENTS non-empty)
+   Host reads MODEL_REQUIREMENTS
+   Downloads/caches required models
+   Resolves model_path for create_engine()
+
+3. ACTIVATION
+   Host calls create_engine(context, model_path, config)
+   Engine created, ready to use
+   Raises EngineError on failure
+
+4. SESSION CREATION
+   Client calls engine.createSession()
+   Returns EngineSession
+   Ownership transfers to caller
+   Raises EngineError on failure
+   No partially initialized session returned
+
+5. SYNTHESIS
+   Client calls session.synthesize(request)
+   Returns SynthesizedAudio
+   Raises EngineError on failure (session remains usable)
+
+6. SESSION DISPOSAL
+   Client calls session.dispose()
+   Releases session resources
+
+7. DEACTIVATION
+   Client calls engine.dispose()
+   Caller must ensure all sessions are disposed first
+   Disposing engine while sessions are alive is undefined behavior
+   Releases engine resources
+```
+
+### 7.2 EngineSession Lifecycle
+
+```
+1. CREATION
+   Created by Engine.createSession()
+   Ownership transfers to caller
+   Raises EngineError on failure
+
+2. USAGE
+   Client calls synthesize() one or more times
+   Each call returns SynthesizedAudio or raises EngineError
+   Session remains usable after synthesize() failure
+   If CancelableSession: cancel() causes synthesize() to raise CancelledError
+   If StreamingSynthesizer: iterator raises CancelledError on cancel(), EngineError on failure
+
+3. DISPOSAL
+   Client calls dispose()
+   Releases session resources
+   After dispose(), all methods except dispose() raise EngineError
+```
+
+### 7.3 Ownership Rules
+
+- Engine.createSession() transfers ownership of the returned session to the caller
+- Caller is responsible for disposing all sessions before disposing the engine
+- Engine does not track sessions; it has no lifecycle registry
+- Disposing an engine while any session is still alive violates the API contract; behavior is undefined
+- This design avoids coupling, synchronization overhead, and lifecycle registry complexity
+
+### 7.4 Concurrent Operations
+
+**Engine.dispose() concurrent with Engine.createSession()**:
+- createSession() must either succeed with fully initialized EngineSession or raise EngineError
+- Partially initialized EngineSession must never be returned
+- After dispose() completes, subsequent createSession() calls must raise EngineError
+
+**EngineSession.dispose() concurrent with EngineSession.synthesize()**:
+- Not thread-safe. Caller must ensure synthesize() completes before dispose().
+
+**EngineSession.dispose() concurrent with StreamingSynthesizer.synthesizeStream()**:
+- Not thread-safe. Caller must ensure stream iteration completes before dispose().
+
+---
+
+## 8. Thread Safety Contract
+
+| Component | Thread-safe | Notes |
+|-----------|-------------|-------|
+| Engine | Yes | createSession() can be called from any thread |
+| EngineSession | No | synthesize() must be called from one thread at a time |
+| HostContext | Yes | Provides shared services |
+| VoiceSelection | Yes | Immutable value object |
+| ParameterValues | Yes | Immutable value object |
+| AudioFormat | Yes | Immutable value object |
+| EngineConfig | Yes | Immutable value object |
+
+---
+
+## 9. dispose() Contract
+
+**General rules**:
+- Calling dispose() multiple times is safe (no-op on second call)
+- dispose() never raises exceptions (catches and logs internally)
+- After dispose(), all methods except dispose() raise EngineError
+
+**Engine.dispose()**:
+- Caller must ensure all sessions are disposed first
+- Disposing engine while sessions are alive violates API contract; behavior is undefined
+- Releases engine resources
+
+**EngineSession.dispose()**:
+- Releases session resources
+
+---
+
+## 10. Dependency Rules
+
+```
+Core Domain (Engine, EngineSession, Value Objects)
+  -> No dependencies
+
+Plugin Manifest (PluginManifest, ModelManifest, etc.)
+  -> No dependencies
+
+Host Context (HostContext)
+  -> Depends on: Core Domain (for types)
+
+Plugin Implementation
+  -> Depends on: Core Domain, Host Context
+
+Host
+  -> Depends on: Core Domain, Plugin Manifest
+```
+
+**Forbidden**:
+- Core Domain -> anything else
+- Plugin Manifest -> anything else
+- Plugin Implementation -> Host (only receives HostContext)
+- Host -> Plugin Implementation (only via create_engine function)
+
+---
+
+## 11. Architectural Invariants
+
+1. Core Domain has zero dependencies
+2. Plugins receive HostContext at creation, not via global state
+3. Model requirements are static (plugin level), not dynamic (engine level)
+4. Host validates capability implementation at load time: each capability declared in PluginManifest.capabilities must be implemented by the exported object via the corresponding interface
+5. synthesize() raises typed exceptions, not returns Result
+6. dispose() is idempotent and never raises
+7. No global state, no service locator
+8. VoiceSelection and ParameterValues are opaque to engine
+9. Display information comes from VoiceManifest
+10. HostContext is minimal (3 fields max)
+11. EngineConfig contains only engine settings, not resource references
+12. EngineSession owns mutable execution state isolated from other concurrent work
+13. Engine.createSession() transfers ownership to caller
+14. Caller must dispose all sessions before disposing engine
+15. After dispose(), all methods except dispose() raise EngineError
+16. create_engine() is atomic (all-or-nothing)
+17. Garbage collection without dispose() may leak (documented)
+18. Capabilities are additive (new capabilities don't break old plugins)
+19. api_version enables compatibility checking
+20. createSession() returns fully initialized session or raises, never partial
+21. cancel() causes synthesize() to raise CancelledError
+22. EngineSession remains usable after cancellation
+23. Engine does not track sessions; no lifecycle registry
+
+---
+
+## 12. Validation Examples
+
+### 12.1 Kokoro
+
+```python
+PLUGIN_MANIFEST = PluginManifest(
+  id="kokoro",
+  api_version="1.0",
+  capabilities=["voice_list", "preview", "voice_blend"],
+  engine=EngineManifest(
+    voiceSources=[
+      VoiceSourceManifest(id="builtin", type="list", config={"voices": [...]}),
+      VoiceSourceManifest(id="formula", type="blend", config={"syntax": "{a}*0.5+{b}*0.5"}),
+    ],
+    parameters=[ParameterManifest(id="speed", type="float", default=1.0, min=0.5, max=2.0)],
+    audioFormats=[AudioFormatManifest(mime="audio/wav", extension="wav")],
+  ),
+)
+
+MODEL_REQUIREMENTS = []
+
+def create_engine(context: HostContext, model_path: Path | None, config: EngineConfig) -> Engine:
+  model = load_kokoro(model_path)
+  return KokoroEngine(model, config.device)
+```
+
+### 12.2 SuperTonic
+
+```python
+PLUGIN_MANIFEST = PluginManifest(
+  id="supertonic",
+  api_version="1.0",
+  capabilities=["voice_list", "preview"],
+  engine=EngineManifest(
+    voiceSources=[VoiceSourceManifest(id="builtin", type="list", config={"voices": [...]})],
+    parameters=[
+      ParameterManifest(id="speed", type="float", default=1.0, min=0.5, max=2.0),
+      ParameterManifest(id="steps", type="int", default=20, min=5, max=50),
+    ],
+    audioFormats=[AudioFormatManifest(mime="audio/wav", extension="wav")],
+  ),
+)
+
+MODEL_REQUIREMENTS = []
+
+def create_engine(context: HostContext, model_path: Path | None, config: EngineConfig) -> Engine:
+  model = load_supertonic(model_path)
+  return SuperTonicEngine(model, config.device)
+```
+
+### 12.3 ElevenLabs
+
+```python
+PLUGIN_MANIFEST = PluginManifest(
+  id="elevenlabs",
+  api_version="1.0",
+  capabilities=["voice_list"],
+  requires=RequirementManifest(internet=True),
+  engine=EngineManifest(
+    voiceSources=[VoiceSourceManifest(id="cloud", type="list", config={"speakers": [...]})],
+    parameters=[ParameterManifest(id="stability", type="float", default=0.5, min=0.0, max=1.0)],
+    audioFormats=[AudioFormatManifest(mime="audio/mpeg", extension="mp3")],
+  ),
+)
+
+MODEL_REQUIREMENTS = []
+
+def create_engine(context: HostContext, model_path: Path | None, config: EngineConfig) -> Engine:
+  api_key = (context.config_dir / "elevenlabs_key").read_text()
+  return ElevenLabsEngine(api_key)
+```
+
+### 12.4 Piper
+
+```python
+PLUGIN_MANIFEST = PluginManifest(
+  id="piper",
+  api_version="1.0",
+  capabilities=[],
+  engine=EngineManifest(
+    voiceSources=[VoiceSourceManifest(id="downloadable", type="list", config={"models": [...]})],
+    parameters=[ParameterManifest(id="speed", type="float", default=1.0, min=0.5, max=2.0)],
+    audioFormats=[AudioFormatManifest(mime="audio/wav", extension="wav")],
+  ),
+)
+
+MODEL_REQUIREMENTS = [
+  ModelManifest(id="en_US-lessac-medium", name="English Lessac Medium", size="100MB"),
+]
+
+def create_engine(context: HostContext, model_path: Path | None, config: EngineConfig) -> Engine:
+  return PiperEngine(model_path, config.device)
+```
+
+### 12.5 XTTS (with streaming and cancellation)
+
+```python
+PLUGIN_MANIFEST = PluginManifest(
+  id="xtts",
+  api_version="1.0",
+  capabilities=["voice_list", "preview", "voice_clone", "streaming", "cancel"],
+  requires=RequirementManifest(gpu=GpuRequirement(required=True, type="cuda")),
+  engine=EngineManifest(
+    voiceSources=[
+      VoiceSourceManifest(id="speakers", type="speaker_id", config={"speakers": [...]}),
+      VoiceSourceManifest(id="clone", type="clone", config={"requiresAudio": True, "maxDuration": 30}),
+    ],
+    parameters=[ParameterManifest(id="temperature", type="float", default=0.7, min=0.1, max=1.0)],
+    audioFormats=[AudioFormatManifest(mime="audio/wav", extension="wav")],
+  ),
+)
+
+MODEL_REQUIREMENTS = [
+  ModelManifest(id="xtts_v2", name="XTTS v2", size="2GB"),
+]
+
+def create_engine(context: HostContext, model_path: Path | None, config: EngineConfig) -> Engine:
+  return XTTSEngine(model_path, config.device)
+```
+
+XTTS session implements: EngineSession, StreamingSynthesizer, CancelableSession.
+
+---
+
+## 13. Summary of All Decisions
+
+| Aspect | Decision |
+|--------|----------|
+| Engine | Factory, stateless, thread-safe for createSession() |
+| EngineSession | Owns mutable execution state, not thread-safe |
+| EngineSession ownership | Caller owns (transferred from createSession) |
+| Engine session tracking | None; engine does not track sessions |
+| StreamingSynthesizer | Optional capability of EngineSession |
+| CancelableSession | Optional capability, cancel() raises CancelledError |
+| dispose() | Idempotent, never raises |
+| Engine.dispose() | Caller must dispose sessions first; undefined if violated |
+| createSession() | Raises EngineError on failure, no partial sessions |
+| create_engine() | Atomic, takes context, model_path, config |
+| EngineConfig | Engine settings only, no resource references |
+| model_path | Separate argument, not in EngineConfig |
+| MODEL_REQUIREMENTS | Static at plugin level |
+| HostContext | Minimal (3 fields) |
+| Error handling | Typed exceptions (EngineError hierarchy) |
+| Thread safety | Documented per component |
+| Capabilities | Additive, optional interfaces |
+| API versioning | api_version in manifest |
+| Concurrent dispose/createSession | Fully initialized session or EngineError |
+| Concurrent dispose/synthesizeStream | Not thread-safe; caller must complete iteration first |

plugins/kokoro/__init__.py

@@ -0,0 +1,185 @@
+"""Kokoro TTS Plugin for the TTS Plugin Architecture.
+
+This plugin provides a Kokoro-based TTS engine that implements the
+Plugin API contract. It wraps the existing Kokoro backend in the
+new Engine/EngineSession architecture.
+
+Exports:
+    - PLUGIN_MANIFEST: PluginManifest
+    - MODEL_REQUIREMENTS: list[ModelManifest]
+    - create_engine: Factory function
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+from abogen.tts_plugin.engine import Engine
+from abogen.tts_plugin.host_context import HostContext
+from abogen.tts_plugin.manifest import (
+    AudioFormatManifest,
+    EngineManifest,
+    ModelManifest,
+    ParameterManifest,
+    PluginManifest,
+    RequirementManifest,
+    VoiceManifest,
+    VoiceSourceManifest,
+)
+from abogen.tts_plugin.types import EngineConfig
+
+from .engine import KokoroEngine
+
+
+def _load_kpipeline() -> Any:
+    """Lazy-load Kokoro dependencies."""
+    # Transformers 5.x moved AlbertModel out of top-level imports.
+    # Monkey-patch before kokoro imports it.
+    import transformers
+    if not hasattr(transformers, "AlbertModel"):
+        from transformers.models.albert import AlbertModel as _AlbertModel
+        transformers.AlbertModel = _AlbertModel
+
+    from kokoro import KPipeline  # type: ignore[import-not-found]
+    return KPipeline
+
+
+PLUGIN_MANIFEST = PluginManifest(
+    id="kokoro",
+    name="Kokoro",
+    version="0.9.4",
+    api_version="1.0",
+    description="Kokoro TTS engine - high quality multilingual text-to-speech",
+    author="Kokoro Team",
+    capabilities=("voice_list",),
+    requires=RequirementManifest(
+        internet=False,
+    ),
+    engine=EngineManifest(
+        voiceSources=(
+            VoiceSourceManifest(
+                id="builtin",
+                name="Built-in Voices",
+                type="list",
+                config={"voices": "See listVoices()"},
+            ),
+        ),
+        parameters=(
+            ParameterManifest(
+                id="speed",
+                name="Speed",
+                description="Speech speed multiplier",
+                type="float",
+                default=1.0,
+                min=0.5,
+                max=2.0,
+                step=0.1,
+            ),
+        ),
+        audioFormats=(
+            AudioFormatManifest(mime="audio/wav", extension="wav"),
+        ),
+    ),
+    voices=(
+        VoiceManifest(id="af_alloy", name="Alloy", tags=("en", "female")),
+        VoiceManifest(id="af_aoede", name="Aoede", tags=("en", "female")),
+        VoiceManifest(id="af_bella", name="Bella", tags=("en", "female")),
+        VoiceManifest(id="af_heart", name="Heart", tags=("en", "female")),
+        VoiceManifest(id="af_jessica", name="Jessica", tags=("en", "female")),
+        VoiceManifest(id="af_kore", name="Kore", tags=("en", "female")),
+        VoiceManifest(id="af_nicole", name="Nicole", tags=("en", "female")),
+        VoiceManifest(id="af_nova", name="Nova", tags=("en", "female")),
+        VoiceManifest(id="af_river", name="River", tags=("en", "female")),
+        VoiceManifest(id="af_sarah", name="Sarah", tags=("en", "female")),
+        VoiceManifest(id="af_sky", name="Sky", tags=("en", "female")),
+        VoiceManifest(id="am_adam", name="Adam", tags=("en", "male")),
+        VoiceManifest(id="am_echo", name="Echo", tags=("en", "male")),
+        VoiceManifest(id="am_eric", name="Eric", tags=("en", "male")),
+        VoiceManifest(id="am_fenrir", name="Fenrir", tags=("en", "male")),
+        VoiceManifest(id="am_liam", name="Liam", tags=("en", "male")),
+        VoiceManifest(id="am_michael", name="Michael", tags=("en", "male")),
+        VoiceManifest(id="am_onyx", name="Onyx", tags=("en", "male")),
+        VoiceManifest(id="am_puck", name="Puck", tags=("en", "male")),
+        VoiceManifest(id="am_santa", name="Santa", tags=("en", "male")),
+        VoiceManifest(id="bf_alice", name="Alice", tags=("en", "female")),
+        VoiceManifest(id="bf_emma", name="Emma", tags=("en", "female")),
+        VoiceManifest(id="bf_isabella", name="Isabella", tags=("en", "female")),
+        VoiceManifest(id="bf_lily", name="Lily", tags=("en", "female")),
+        VoiceManifest(id="bm_daniel", name="Daniel", tags=("en", "male")),
+        VoiceManifest(id="bm_fable", name="Fable", tags=("en", "male")),
+        VoiceManifest(id="bm_george", name="George", tags=("en", "male")),
+        VoiceManifest(id="bm_lewis", name="Lewis", tags=("en", "male")),
+        VoiceManifest(id="ef_dora", name="Dora", tags=("es", "female")),
+        VoiceManifest(id="em_alex", name="Alex", tags=("es", "male")),
+        VoiceManifest(id="em_santa", name="Santa", tags=("es", "male")),
+        VoiceManifest(id="ff_siwis", name="Siwis", tags=("fr", "female")),
+        VoiceManifest(id="hf_alpha", name="Alpha", tags=("hi", "female")),
+        VoiceManifest(id="hf_beta", name="Beta", tags=("hi", "female")),
+        VoiceManifest(id="hm_omega", name="Omega", tags=("hi", "male")),
+        VoiceManifest(id="hm_psi", name="Psi", tags=("hi", "male")),
+        VoiceManifest(id="if_sara", name="Sara", tags=("it", "female")),
+        VoiceManifest(id="im_nicola", name="Nicola", tags=("it", "male")),
+        VoiceManifest(id="jf_alpha", name="Alpha", tags=("ja", "female")),
+        VoiceManifest(id="jf_gongitsune", name="Gongitsune", tags=("ja", "female")),
+        VoiceManifest(id="jf_nezumi", name="Nezumi", tags=("ja", "female")),
+        VoiceManifest(id="jf_tebukuro", name="Tebukuro", tags=("ja", "female")),
+        VoiceManifest(id="jm_kumo", name="Kumo", tags=("ja", "male")),
+        VoiceManifest(id="pf_dora", name="Dora", tags=("pt", "female")),
+        VoiceManifest(id="pm_alex", name="Alex", tags=("pt", "male")),
+        VoiceManifest(id="pm_santa", name="Santa", tags=("pt", "male")),
+        VoiceManifest(id="zf_xiaobei", name="Xiaobei", tags=("zh", "female")),
+        VoiceManifest(id="zf_xiaoni", name="Xiaoni", tags=("zh", "female")),
+        VoiceManifest(id="zf_xiaoxiao", name="Xiaoxiao", tags=("zh", "female")),
+        VoiceManifest(id="zf_xiaoyi", name="Xiaoyi", tags=("zh", "female")),
+        VoiceManifest(id="zm_yunjian", name="Yunjian", tags=("zh", "male")),
+        VoiceManifest(id="zm_yunxi", name="Yunxi", tags=("zh", "male")),
+        VoiceManifest(id="zm_yunxia", name="Yunxia", tags=("zh", "female")),
+        VoiceManifest(id="zm_yunyang", name="Yunyang", tags=("zh", "male")),
+    ),
+)
+
+MODEL_REQUIREMENTS: list[ModelManifest] = []
+
+
+def create_engine(
+    context: HostContext,
+    model_path: Path | None,
+    config: EngineConfig,
+) -> Engine:
+    """Create a Kokoro engine instance.
+
+    This function is the plugin entry point. It must be atomic:
+    succeed fully or raise EngineError and clean up.
+
+    Args:
+        context: Host services (config dir, logger, http client).
+        model_path: Resolved model path, or None for default.
+        config: Engine initialization settings (device, etc.).
+
+    Returns:
+        A fully initialized KokoroEngine instance.
+
+    Raises:
+        EngineError: On failure. Cleans up partially created resources.
+    """
+    try:
+        KPipeline = _load_kpipeline()
+
+        # Determine repo_id from model_path or use default
+        repo_id = "hexgrad/Kokoro-82M"
+        if model_path is not None:
+            # If a specific model path is provided, use it as repo_id
+            repo_id = str(model_path)
+
+        pipeline = KPipeline(
+            lang_code=config.lang_code,
+            repo_id=repo_id,
+            device=config.device,
+        )
+
+        engine = KokoroEngine(pipeline)
+        return engine
+    except Exception as e:
+        from abogen.tts_plugin.errors import EngineError as EngineErrorClass
+        raise EngineErrorClass(f"Failed to create Kokoro engine: {e}") from e

plugins/kokoro/engine.py

@@ -0,0 +1,118 @@
+"""Kokoro Engine adapter for the TTS Plugin Architecture.
+
+This module adapts the existing Kokoro backend to the new Engine/EngineSession
+protocol. It wraps the KokoroBackend without modifying it.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Any
+
+import numpy as np
+
+from abogen.tts_plugin.capabilities import VoiceLister
+from abogen.tts_plugin.engine import Engine, EngineSession
+from abogen.tts_plugin.errors import EngineError
+from abogen.tts_plugin.manifest import VoiceManifest
+from abogen.tts_plugin.types import (
+    AudioFormat,
+    Duration,
+    SynthesisRequest,
+    SynthesizedAudio,
+)
+
+logger = logging.getLogger(__name__)
+
+# Sample rate for Kokoro audio
+_KOKORO_SAMPLE_RATE = 24000
+
+
+class KokoroSession:
+    """EngineSession implementation for Kokoro.
+
+    Owns mutable execution state for synthesis.
+    NOT thread-safe.
+    """
+
+    def __init__(self, pipeline: Any) -> None:
+        self._pipeline = pipeline
+        self._disposed = False
+
+    def synthesize(self, request: SynthesisRequest) -> SynthesizedAudio:
+        """Synthesize audio from text using Kokoro."""
+        if self._disposed:
+            raise EngineError("Session disposed")
+
+        try:
+            voice = request.voice.key
+            speed = request.parameters.values.get("speed", 1.0)
+            split_pattern = request.parameters.values.get("split_pattern", None)
+
+            audio_parts: list[np.ndarray] = []
+            for segment in self._pipeline(
+                request.text,
+                voice=voice,
+                speed=speed,
+                split_pattern=split_pattern,
+            ):
+                audio = segment.audio
+                if hasattr(audio, "numpy"):
+                    audio = audio.numpy()
+                audio_parts.append(np.asarray(audio, dtype="float32"))
+
+            if not audio_parts:
+                return SynthesizedAudio(
+                    data=b"",
+                    format=AudioFormat(mime="audio/wav", extension="wav"),
+                    duration=Duration(seconds=0.0),
+                )
+
+            combined = np.concatenate(audio_parts).astype("float32", copy=False)
+            audio_bytes = combined.tobytes()
+            duration_seconds = len(combined) / _KOKORO_SAMPLE_RATE
+
+            return SynthesizedAudio(
+                data=audio_bytes,
+                format=AudioFormat(mime="audio/wav", extension="wav"),
+                duration=Duration(seconds=duration_seconds),
+            )
+        except EngineError:
+            raise
+        except Exception as e:
+            raise EngineError(f"Synthesis failed: {e}") from e
+
+    def dispose(self) -> None:
+        """Release session resources. Idempotent."""
+        self._disposed = True
+
+
+class KokoroEngine:
+    """Engine implementation for Kokoro.
+
+    Factory for KokoroSession instances. Stateless and thread-safe.
+    """
+
+    def __init__(self, pipeline: Any) -> None:
+        self._pipeline = pipeline
+        self._disposed = False
+
+    def createSession(self) -> KokoroSession:
+        """Create a new KokoroSession."""
+        if self._disposed:
+            raise EngineError("Engine disposed")
+        return KokoroSession(self._pipeline)
+
+    def dispose(self) -> None:
+        """Release engine resources. Idempotent."""
+        self._disposed = True
+
+    def listVoices(self, sourceId: str) -> list[VoiceManifest]:
+        """List available Kokoro voices. Implements VoiceLister capability.
+
+        Note: Static voices are declared in the plugin manifest.
+        This method is a fallback for dynamic plugins.
+        """
+        if self._disposed:
+            raise EngineError("Engine disposed")
+        return []

plugins/supertonic/__init__.py

@@ -0,0 +1,136 @@
+"""SuperTonic TTS Plugin for the TTS Plugin Architecture.
+
+This plugin provides a SuperTonic-based TTS engine that implements the
+Plugin API contract. It wraps the existing SuperTonic backend in the
+new Engine/EngineSession architecture.
+
+Exports:
+    - PLUGIN_MANIFEST: PluginManifest
+    - MODEL_REQUIREMENTS: list[ModelManifest]
+    - create_engine: Factory function
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+from abogen.tts_plugin.engine import Engine
+from abogen.tts_plugin.host_context import HostContext
+from abogen.tts_plugin.manifest import (
+    AudioFormatManifest,
+    EngineManifest,
+    ModelManifest,
+    ParameterManifest,
+    PluginManifest,
+    RequirementManifest,
+    VoiceManifest,
+    VoiceSourceManifest,
+)
+from abogen.tts_plugin.types import EngineConfig
+
+from .engine import SuperTonicEngine
+
+
+def _load_supertonic_pipeline() -> Any:
+    """Lazy-load SuperTonic dependencies and create pipeline."""
+    from plugins.supertonic.pipeline import SupertonicPipeline
+
+    return SupertonicPipeline(
+        sample_rate=24000,
+        auto_download=True,
+        total_steps=5,
+    )
+
+
+PLUGIN_MANIFEST = PluginManifest(
+    id="supertonic",
+    name="SuperTonic",
+    version="0.1.0",
+    api_version="1.0",
+    description="SuperTonic TTS engine - fast high-quality text-to-speech",
+    author="SuperTonic Team",
+    capabilities=("voice_list",),
+    requires=RequirementManifest(
+        internet=False,
+    ),
+    engine=EngineManifest(
+        voiceSources=(
+            VoiceSourceManifest(
+                id="builtin",
+                name="Built-in Voices",
+                type="list",
+                config={"voices": "See listVoices()"},
+            ),
+        ),
+        parameters=(
+            ParameterManifest(
+                id="speed",
+                name="Speed",
+                description="Speech speed multiplier",
+                type="float",
+                default=1.0,
+                min=0.7,
+                max=2.0,
+                step=0.1,
+            ),
+            ParameterManifest(
+                id="total_steps",
+                name="Quality Steps",
+                description="Inference steps (higher = better quality, slower)",
+                type="int",
+                default=5,
+                min=2,
+                max=15,
+                step=1,
+            ),
+        ),
+        audioFormats=(
+            AudioFormatManifest(mime="audio/wav", extension="wav"),
+        ),
+    ),
+    voices=(
+        VoiceManifest(id="M1", name="Male 1", tags=("male",)),
+        VoiceManifest(id="M2", name="Male 2", tags=("male",)),
+        VoiceManifest(id="M3", name="Male 3", tags=("male",)),
+        VoiceManifest(id="M4", name="Male 4", tags=("male",)),
+        VoiceManifest(id="M5", name="Male 5", tags=("male",)),
+        VoiceManifest(id="F1", name="Female 1", tags=("female",)),
+        VoiceManifest(id="F2", name="Female 2", tags=("female",)),
+        VoiceManifest(id="F3", name="Female 3", tags=("female",)),
+        VoiceManifest(id="F4", name="Female 4", tags=("female",)),
+        VoiceManifest(id="F5", name="Female 5", tags=("female",)),
+    ),
+)
+
+MODEL_REQUIREMENTS: list[ModelManifest] = []
+
+
+def create_engine(
+    context: HostContext,
+    model_path: Path | None,
+    config: EngineConfig,
+) -> Engine:
+    """Create a SuperTonic engine instance.
+
+    This function is the plugin entry point. It must be atomic:
+    succeed fully or raise EngineError and clean up.
+
+    Args:
+        context: Host services (config dir, logger, http client).
+        model_path: Resolved model path, or None for default.
+        config: Engine initialization settings (device, etc.).
+
+    Returns:
+        A fully initialized SuperTonicEngine instance.
+
+    Raises:
+        EngineError: On failure. Cleans up partially created resources.
+    """
+    try:
+        pipeline = _load_supertonic_pipeline()
+        engine = SuperTonicEngine(pipeline)
+        return engine
+    except Exception as e:
+        from abogen.tts_plugin.errors import EngineError as EngineErrorClass
+        raise EngineErrorClass(f"Failed to create SuperTonic engine: {e}") from e

plugins/supertonic/engine.py

@@ -0,0 +1,125 @@
+"""SuperTonic Engine adapter for the TTS Plugin Architecture.
+
+This module adapts the existing SuperTonic backend to the new Engine/EngineSession
+protocol. It wraps the SupertonicPipeline without modifying it.
+"""
+
+from __future__ import annotations
+
+import io
+import logging
+from typing import Any
+
+import numpy as np
+
+from abogen.tts_plugin.capabilities import VoiceLister
+from abogen.tts_plugin.engine import Engine, EngineSession
+from abogen.tts_plugin.errors import EngineError
+from abogen.tts_plugin.manifest import VoiceManifest
+from abogen.tts_plugin.types import (
+    AudioFormat,
+    Duration,
+    SynthesisRequest,
+    SynthesizedAudio,
+)
+
+logger = logging.getLogger(__name__)
+
+# Sample rate for SuperTonic audio
+_SUPERTONIC_SAMPLE_RATE = 24000
+
+
+class SuperTonicSession:
+    """EngineSession implementation for SuperTonic.
+
+    Owns mutable execution state for synthesis.
+    NOT thread-safe.
+    """
+
+    def __init__(self, pipeline: Any) -> None:
+        self._pipeline = pipeline
+        self._disposed = False
+
+    def synthesize(self, request: SynthesisRequest) -> SynthesizedAudio:
+        """Synthesize audio from text using SuperTonic."""
+        if self._disposed:
+            raise EngineError("Session disposed")
+
+        try:
+            import soundfile as sf
+
+            voice = request.voice.key
+            speed = float(request.parameters.values.get("speed", 1.0))
+            total_steps = request.parameters.values.get("total_steps", None)
+            split_pattern = request.parameters.values.get("split_pattern", None)
+
+            if total_steps is not None:
+                total_steps = int(total_steps)
+
+            audio_parts: list[np.ndarray] = []
+            for segment in self._pipeline(
+                request.text,
+                voice=voice,
+                speed=speed,
+                split_pattern=split_pattern,
+                total_steps=total_steps,
+            ):
+                audio_parts.append(segment.audio)
+
+            if not audio_parts:
+                return SynthesizedAudio(
+                    data=b"",
+                    format=AudioFormat(mime="audio/wav", extension="wav"),
+                    duration=Duration(seconds=0.0),
+                )
+
+            combined = np.concatenate(audio_parts).astype("float32", copy=False)
+            buf = io.BytesIO()
+            sf.write(buf, combined, self._pipeline.sample_rate, format="WAV")
+            audio_bytes = buf.getvalue()
+            duration_seconds = len(combined) / self._pipeline.sample_rate
+
+            return SynthesizedAudio(
+                data=audio_bytes,
+                format=AudioFormat(mime="audio/wav", extension="wav"),
+                duration=Duration(seconds=duration_seconds),
+            )
+        except EngineError:
+            raise
+        except Exception as e:
+            raise EngineError(f"Synthesis failed: {e}") from e
+
+    def dispose(self) -> None:
+        """Release session resources. Idempotent."""
+        self._disposed = True
+
+
+class SuperTonicEngine:
+    """Engine implementation for SuperTonic.
+
+    Factory for SuperTonicSession instances. Stateless and thread-safe.
+    """
+
+    def __init__(self, pipeline: Any) -> None:
+        self._pipeline = pipeline
+        self._disposed = False
+
+    def createSession(self) -> SuperTonicSession:
+        """Create a new SuperTonicSession."""
+        if self._disposed:
+            raise EngineError("Engine disposed")
+        return SuperTonicSession(self._pipeline)
+
+    def dispose(self) -> None:
+        """Release engine resources. Idempotent."""
+        self._disposed = True
+
+    def listVoices(self, sourceId: str) -> list[VoiceManifest]:
+        """List available SuperTonic voices. Implements VoiceLister capability.
+        
+        Note: Static voice catalog is declared in plugin manifest.
+        This method is retained for VoiceLister interface compliance.
+        """
+        if self._disposed:
+            raise EngineError("Engine disposed")
+        return []

plugins/supertonic/pipeline.py

@@ -1,31 +1,25 @@
+"""SuperTonic Pipeline — self-contained TTS pipeline for the plugin.
+
+This module provides the SuperTonicPipeline class and supporting utilities
+used by the SuperTonic plugin. It is independent of the legacy
+abogen.tts_backends module.
+"""
+
 from __future__ import annotations
 
 import ast
-from dataclasses import dataclass
 import logging
-import math
 import re
 from typing import Any, Iterable, Iterator, Optional
 
 import numpy as np
 
-
 logger = logging.getLogger(__name__)
 
 
-DEFAULT_SUPERTONIC_VOICES = ("M1", "M2", "M3", "M4", "M5", "F1", "F2", "F3", "F4", "F5")
-
-
-@dataclass
-class SupertonicSegment:
-    graphemes: str
-    audio: np.ndarray
-
-
 def _ensure_float32_mono(wav: Any) -> np.ndarray:
     arr = np.asarray(wav, dtype="float32")
     if arr.ndim == 2:
-        # (n, 1) or (1, n) or (n, channels)
         if arr.shape[0] == 1 and arr.shape[1] > 1:
             arr = arr.reshape(-1)
         else:
@@ -62,7 +56,6 @@ def _split_text(
     else:
         parts = [stripped]
 
-    # Enforce max length by hard-splitting long parts.
     result: list[str] = []
     for part in parts:
         if len(part) <= max_chunk_length:
@@ -71,7 +64,6 @@ def _split_text(
         start = 0
         while start < len(part):
             end = min(len(part), start + max_chunk_length)
-            # Try to split at whitespace.
             if end < len(part):
                 ws = part.rfind(" ", start, end)
                 if ws > start + 40:
@@ -90,7 +82,6 @@ _UNSUPPORTED_CHARS_RE = re.compile(
 
 def _parse_unsupported_characters(error: BaseException) -> list[str]:
     """Best-effort extraction of unsupported characters from SuperTonic errors."""
-
     message = " ".join(
         str(part) for part in getattr(error, "args", ()) if part is not None
     ) or str(error)
@@ -136,16 +127,11 @@ def _configure_supertonic_gpu() -> None:
 
         available = ort.get_available_providers()
 
-        # Use CUDA if available, skip TensorRT (requires extra libs not always present)
-        # TensorrtExecutionProvider may be listed as available but fail at runtime
-        # if TensorRT libraries (libnvinfer.so) are not installed
         providers = []
         if "CUDAExecutionProvider" in available:
             providers.append("CUDAExecutionProvider")
         providers.append("CPUExecutionProvider")
 
-        # Patch supertonic's config and loader before TTS import
-        # We must patch both because loader imports the value at module load time
         import supertonic.config as supertonic_config
         import supertonic.loader as supertonic_loader
 
@@ -156,6 +142,16 @@ def _configure_supertonic_gpu() -> None:
         logger.warning("Could not configure supertonic GPU providers: %s", exc)
 
 
+class SupertonicSegment:
+    """A single synthesized audio segment."""
+
+    __slots__ = ("graphemes", "audio")
+
+    def __init__(self, graphemes: str, audio: np.ndarray) -> None:
+        self.graphemes = graphemes
+        self.audio = audio
+
+
 class SupertonicPipeline:
     """Minimal adapter that mimics Kokoro's pipeline iteration interface."""
 
@@ -171,7 +167,6 @@ class SupertonicPipeline:
         self.total_steps = int(total_steps)
         self.max_chunk_length = int(max_chunk_length)
 
-        # Configure GPU providers before importing TTS
         _configure_supertonic_gpu()
 
         try:
@@ -207,7 +202,6 @@ class SupertonicPipeline:
             removed: set[str] = set()
             last_exc: Exception | None = None
 
-            # SuperTonic can raise ValueError for unsupported characters; strip and retry.
             for attempt in range(3):
                 try:
                     wav, duration = self._tts.synthesize(
@@ -231,7 +225,6 @@ class SupertonicPipeline:
                         chunk_to_speak, unsupported
                     ).strip()
 
-                    # If we didn't change anything, don't loop forever.
                     if sanitized == chunk_to_speak.strip():
                         raise
 
@@ -249,7 +242,6 @@ class SupertonicPipeline:
                             sorted(removed),
                         )
             else:
-                # Exhausted retries.
                 assert last_exc is not None
                 raise last_exc
 
@@ -258,7 +250,6 @@ class SupertonicPipeline:
 
             audio = _ensure_float32_mono(wav)
 
-            # If duration is present, infer the source sample rate and resample if needed.
             src_rate = self.sample_rate
             try:
                 dur = float(duration)

pyproject.toml

@@ -96,8 +96,6 @@ exclude = [
 [tool.hatch.build.targets.wheel]
 packages = ["abogen"]
 
-[tool.hatch.build]
-include = ["abogen/webui/templates/**", "abogen/webui/static/**"]
 
 [tool.hatch.version]
 path = "abogen/VERSION"

tests/contracts/__init__.py

@@ -0,0 +1,5 @@
+"""Contract tests for the TTS Plugin API.
+
+This package contains reusable contract tests that any TTS plugin implementation
+must satisfy. Tests use only the public API and are engine-agnostic.
+"""

tests/contracts/conftest.py

@@ -0,0 +1,231 @@
+"""Shared fixtures and stubs for contract tests.
+
+This module provides minimal stub implementations that satisfy the public API
+for testing purposes. These stubs do NOT contain real business logic.
+"""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Iterator
+
+import pytest
+
+from abogen.tts_plugin.engine import Engine, EngineSession
+from abogen.tts_plugin.host_context import HostContext
+from abogen.tts_plugin.types import (
+    AudioFormat,
+    Duration,
+    EngineConfig,
+    ParameterValues,
+    SynthesisRequest,
+    SynthesizedAudio,
+    VoiceSelection,
+)
+
+
+class FakeHttpClient:
+    """Stub HTTP client that satisfies the HttpClient protocol."""
+
+    def get(self, url: str, **kwargs: object) -> object:
+        return None
+
+    def post(self, url: str, **kwargs: object) -> object:
+        return None
+
+
+class FakeEngineSession:
+    """Stub EngineSession for testing protocol compliance."""
+
+    def __init__(self) -> None:
+        self._disposed = False
+
+    def synthesize(self, request: SynthesisRequest) -> SynthesizedAudio:
+        if self._disposed:
+            from abogen.tts_plugin.errors import EngineError
+
+            raise EngineError("Session disposed")
+        return SynthesizedAudio(
+            data=b"\x00" * 100,
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+            duration=Duration(seconds=1.0),
+        )
+
+    def dispose(self) -> None:
+        self._disposed = True
+
+
+class FakeStreamingSession:
+    """Stub EngineSession with StreamingSynthesizer capability."""
+
+    def __init__(self) -> None:
+        self._disposed = False
+
+    def synthesize(self, request: SynthesisRequest) -> SynthesizedAudio:
+        if self._disposed:
+            from abogen.tts_plugin.errors import EngineError
+
+            raise EngineError("Session disposed")
+        return SynthesizedAudio(
+            data=b"\x00" * 100,
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+            duration=Duration(seconds=1.0),
+        )
+
+    def synthesizeStream(self, request: SynthesisRequest) -> Iterator[bytes]:
+        if self._disposed:
+            from abogen.tts_plugin.errors import EngineError
+
+            raise EngineError("Session disposed")
+        for i in range(3):
+            yield b"\x00" * 50
+
+    def dispose(self) -> None:
+        self._disposed = True
+
+
+class FakeCancelableSession:
+    """Stub EngineSession with CancelableSession capability."""
+
+    def __init__(self) -> None:
+        self._disposed = False
+        self._cancelled = False
+
+    def synthesize(self, request: SynthesisRequest) -> SynthesizedAudio:
+        if self._disposed:
+            from abogen.tts_plugin.errors import EngineError
+
+            raise EngineError("Session disposed")
+        if self._cancelled:
+            from abogen.tts_plugin.errors import CancelledError
+
+            raise CancelledError("Cancelled")
+        return SynthesizedAudio(
+            data=b"\x00" * 100,
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+            duration=Duration(seconds=1.0),
+        )
+
+    def cancel(self) -> None:
+        if self._disposed:
+            from abogen.tts_plugin.errors import EngineError
+
+            raise EngineError("Session disposed")
+        self._cancelled = True
+
+    def dispose(self) -> None:
+        self._disposed = True
+
+
+class FakeEngine:
+    """Stub Engine for testing protocol compliance."""
+
+    def __init__(self, session_class: type = FakeEngineSession) -> None:
+        self._disposed = False
+        self._session_class = session_class
+
+    def createSession(self) -> EngineSession:
+        if self._disposed:
+            from abogen.tts_plugin.errors import EngineError
+
+            raise EngineError("Engine disposed")
+        return self._session_class()
+
+    def dispose(self) -> None:
+        self._disposed = True
+
+
+class FakeVoiceListerEngine:
+    """Stub Engine that also implements VoiceLister."""
+
+    def __init__(self) -> None:
+        self._disposed = False
+
+    def createSession(self) -> EngineSession:
+        if self._disposed:
+            from abogen.tts_plugin.errors import EngineError
+
+            raise EngineError("Engine disposed")
+        return FakeEngineSession()
+
+    def listVoices(self, sourceId: str) -> list:
+        from abogen.tts_plugin.manifest import VoiceManifest
+
+        return [
+            VoiceManifest(id="voice1", name="Voice 1", tags=("en",)),
+            VoiceManifest(id="voice2", name="Voice 2", tags=("es",)),
+        ]
+
+    def dispose(self) -> None:
+        self._disposed = True
+
+
+class FakePreviewEngine:
+    """Stub Engine that also implements PreviewGenerator."""
+
+    def __init__(self) -> None:
+        self._disposed = False
+
+    def createSession(self) -> EngineSession:
+        if self._disposed:
+            from abogen.tts_plugin.errors import EngineError
+
+            raise EngineError("Engine disposed")
+        return FakeEngineSession()
+
+    def generatePreview(self, voice: VoiceSelection, text: str) -> SynthesizedAudio:
+        return SynthesizedAudio(
+            data=b"\x00" * 50,
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+            duration=Duration(seconds=0.5),
+        )
+
+    def dispose(self) -> None:
+        self._disposed = True
+
+
+@pytest.fixture
+def fake_http_client() -> FakeHttpClient:
+    return FakeHttpClient()
+
+
+@pytest.fixture
+def host_context(tmp_path: Path, fake_http_client: FakeHttpClient) -> HostContext:
+    return HostContext(
+        config_dir=tmp_path,
+        logger=logging.getLogger("test"),
+        http_client=fake_http_client,
+    )
+
+
+@pytest.fixture
+def fake_engine() -> FakeEngine:
+    return FakeEngine()
+
+
+@pytest.fixture
+def fake_session() -> FakeEngineSession:
+    return FakeEngineSession()
+
+
+@pytest.fixture
+def default_voice() -> VoiceSelection:
+    return VoiceSelection(source="builtin", key="af_nova")
+
+
+@pytest.fixture
+def default_format() -> AudioFormat:
+    return AudioFormat(mime="audio/wav", extension="wav")
+
+
+@pytest.fixture
+def default_request(
+    default_voice: VoiceSelection, default_format: AudioFormat
+) -> SynthesisRequest:
+    return SynthesisRequest(
+        text="Hello, world!",
+        voice=default_voice,
+        parameters=ParameterValues(values={}),
+        format=default_format,
+    )

tests/contracts/engine_contract.py

@@ -0,0 +1,120 @@
+"""Base contract tests for Engine implementations.
+
+Any new TTS plugin must inherit from these classes to verify
+it satisfies the Engine/EngineSession protocol.
+
+Usage:
+    from tests.contracts.engine_contract import EngineContractMixin
+
+    class TestMyEngine(EngineContractMixin):
+        @pytest.fixture
+        def engine(self):
+            return create_my_engine()
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from abogen.tts_plugin.engine import Engine, EngineSession
+from abogen.tts_plugin.errors import EngineError
+from abogen.tts_plugin.types import (
+    AudioFormat,
+    ParameterValues,
+    SynthesisRequest,
+    SynthesizedAudio,
+    VoiceSelection,
+)
+
+
+class EngineContractMixin:
+    """Base contract tests for Engine implementations.
+
+    Subclasses must define a module-level ``engine`` fixture returning
+    a fully initialized Engine instance. The tests below will use it
+    via pytest's standard fixture resolution.
+    """
+
+    def _req(self, text: str = "Hello", voice: str | None = None) -> SynthesisRequest:
+        return SynthesisRequest(
+            text=text,
+            voice=VoiceSelection(source="builtin", key=voice or "default"),
+            parameters=ParameterValues(values={}),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+
+    # ── Engine protocol ──────────────────────────────────────
+
+    def test_engine_satisfies_protocol(self, engine: Engine) -> None:
+        assert isinstance(engine, Engine)
+
+    def test_create_session_returns_session(self, engine: Engine) -> None:
+        session = engine.createSession()
+        assert isinstance(session, EngineSession)
+        session.dispose()
+
+    def test_create_session_returns_new_instances(self, engine: Engine) -> None:
+        s1 = engine.createSession()
+        s2 = engine.createSession()
+        assert s1 is not s2
+        s1.dispose()
+        s2.dispose()
+
+    def test_dispose_is_idempotent(self, engine: Engine) -> None:
+        engine.dispose()
+        engine.dispose()
+
+    def test_create_session_after_dispose_raises(self, engine: Engine) -> None:
+        engine.dispose()
+        with pytest.raises(EngineError):
+            engine.createSession()
+
+    # ── Session protocol ─────────────────────────────────────
+
+    def test_session_satisfies_protocol(self, engine: Engine) -> None:
+        session = engine.createSession()
+        assert isinstance(session, EngineSession)
+        session.dispose()
+        engine.dispose()
+
+    def test_session_synthesize_returns_audio(self, engine: Engine) -> None:
+        session = engine.createSession()
+        result = session.synthesize(self._req())
+        assert isinstance(result, SynthesizedAudio)
+        assert isinstance(result.data, bytes)
+        assert len(result.data) > 0
+        session.dispose()
+        engine.dispose()
+
+    def test_session_dispose_is_idempotent(self, engine: Engine) -> None:
+        session = engine.createSession()
+        session.dispose()
+        session.dispose()
+        engine.dispose()
+
+    def test_session_synthesize_after_dispose_raises(self, engine: Engine) -> None:
+        session = engine.createSession()
+        session.dispose()
+        with pytest.raises(EngineError):
+            session.synthesize(self._req())
+        engine.dispose()
+
+    def test_session_multiple_synthesize(self, engine: Engine) -> None:
+        session = engine.createSession()
+        r1 = session.synthesize(self._req())
+        r2 = session.synthesize(self._req())
+        assert isinstance(r1.data, bytes)
+        assert isinstance(r2.data, bytes)
+        session.dispose()
+        engine.dispose()
+
+    # ── Lifecycle ────────────────────────────────────────────
+
+    def test_full_lifecycle(self, engine: Engine) -> None:
+        s1 = engine.createSession()
+        s2 = engine.createSession()
+        s1.synthesize(self._req())
+        s2.synthesize(self._req())
+        s1.dispose()
+        s2.dispose()
+        engine.dispose()

tests/contracts/test_capabilities_contract.py

@@ -0,0 +1,183 @@
+"""Contract tests for capability interfaces.
+
+These tests verify that capability interfaces satisfy the architectural requirements:
+- VoiceLister: lists voices for a source
+- PreviewGenerator: generates preview audio
+- StreamingSynthesizer: yields audio chunks
+- CancelableSession: cancels in-progress synthesis
+"""
+
+import pytest
+
+from abogen.tts_plugin.capabilities import (
+    CancelableSession,
+    PreviewGenerator,
+    StreamingSynthesizer,
+    VoiceLister,
+)
+from abogen.tts_plugin.errors import CancelledError, EngineError
+from abogen.tts_plugin.manifest import VoiceManifest
+from abogen.tts_plugin.types import (
+    AudioFormat,
+    Duration,
+    ParameterValues,
+    SynthesisRequest,
+    SynthesizedAudio,
+    VoiceSelection,
+)
+
+from .conftest import FakeCancelableSession, FakeStreamingSession, FakeVoiceListerEngine
+
+
+class TestVoiceListerProtocolContract:
+    """Contract tests for VoiceLister protocol."""
+
+    def test_voice_lister_is_protocol(self) -> None:
+        assert hasattr(VoiceLister, "__protocol_attrs__")
+
+    def test_voice_lister_satisfied_by_engine(self) -> None:
+        engine = FakeVoiceListerEngine()
+        assert isinstance(engine, VoiceLister)
+
+    def test_list_voices_returns_list(self) -> None:
+        engine = FakeVoiceListerEngine()
+        voices = engine.listVoices("builtin")
+        assert isinstance(voices, list)
+
+    def test_list_voices_returns_voice_manifests(self) -> None:
+        engine = FakeVoiceListerEngine()
+        voices = engine.listVoices("builtin")
+        for voice in voices:
+            assert isinstance(voice, VoiceManifest)
+
+    def test_list_voices_has_required_fields(self) -> None:
+        engine = FakeVoiceListerEngine()
+        voices = engine.listVoices("builtin")
+        for voice in voices:
+            assert hasattr(voice, "id")
+            assert hasattr(voice, "name")
+            assert hasattr(voice, "tags")
+
+
+class TestPreviewGeneratorProtocolContract:
+    """Contract tests for PreviewGenerator protocol."""
+
+    def test_preview_generator_is_protocol(self) -> None:
+        assert hasattr(PreviewGenerator, "__protocol_attrs__")
+
+    def test_preview_generator_satisfied_by_engine(self) -> None:
+        from .conftest import FakePreviewEngine
+
+        engine = FakePreviewEngine()
+        assert isinstance(engine, PreviewGenerator)
+
+    def test_generate_preview_returns_synthesized_audio(self) -> None:
+        from .conftest import FakePreviewEngine
+
+        engine = FakePreviewEngine()
+        voice = VoiceSelection(source="builtin", key="af_nova")
+        result = engine.generatePreview(voice, "Hello")
+        assert isinstance(result, SynthesizedAudio)
+
+    def test_generate_preview_has_valid_data(self) -> None:
+        from .conftest import FakePreviewEngine
+
+        engine = FakePreviewEngine()
+        voice = VoiceSelection(source="builtin", key="af_nova")
+        result = engine.generatePreview(voice, "Hello")
+        assert isinstance(result.data, bytes)
+        assert len(result.data) > 0
+
+
+class TestStreamingSynthesizerProtocolContract:
+    """Contract tests for StreamingSynthesizer protocol."""
+
+    def test_streaming_synthesizer_is_protocol(self) -> None:
+        assert hasattr(StreamingSynthesizer, "__protocol_attrs__")
+
+    def test_streaming_session_satisfies_protocol(self) -> None:
+        session = FakeStreamingSession()
+        assert isinstance(session, StreamingSynthesizer)
+
+    def test_synthesize_stream_yields_bytes(self) -> None:
+        session = FakeStreamingSession()
+        request = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="af_nova"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        chunks = list(session.synthesizeStream(request))
+        assert len(chunks) > 0
+        for chunk in chunks:
+            assert isinstance(chunk, bytes)
+
+    def test_streaming_iterator_exhaustion(self) -> None:
+        """Architecture spec: Iterator exhaustion = synthesis complete."""
+        session = FakeStreamingSession()
+        request = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="af_nova"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        chunks = list(session.synthesizeStream(request))
+        assert len(chunks) == 3
+
+    def test_streaming_after_dispose_raises(self) -> None:
+        """Architecture spec: After dispose(), methods raise EngineError."""
+        session = FakeStreamingSession()
+        session.dispose()
+        request = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="af_nova"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        with pytest.raises(EngineError):
+            list(session.synthesizeStream(request))
+
+
+class TestCancelableSessionProtocolContract:
+    """Contract tests for CancelableSession protocol."""
+
+    def test_cancelable_session_is_protocol(self) -> None:
+        assert hasattr(CancelableSession, "__protocol_attrs__")
+
+    def test_cancelable_session_satisfies_protocol(self) -> None:
+        session = FakeCancelableSession()
+        assert isinstance(session, CancelableSession)
+
+    def test_cancel_causes_synthesize_to_raise_cancelled(self) -> None:
+        """Architecture spec: cancel() causes synthesize() to raise CancelledError."""
+        session = FakeCancelableSession()
+        request = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="af_nova"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+
+        # Cancel
+        session.cancel()
+
+        # synthesize should raise CancelledError
+        with pytest.raises(CancelledError):
+            session.synthesize(request)
+
+    def test_cancel_after_dispose_raises(self) -> None:
+        """Architecture spec: cancel() raises EngineError if called after dispose()."""
+        session = FakeCancelableSession()
+        session.dispose()
+        with pytest.raises(EngineError):
+            session.cancel()
+
+    def test_session_usable_after_cancel(self) -> None:
+        """Architecture spec: EngineSession remains usable after cancellation."""
+        session = FakeCancelableSession()
+
+        # Cancel
+        session.cancel()
+
+        # Dispose and create new session for synthesis
+        session.dispose()

tests/contracts/test_engine_contract.py

@@ -0,0 +1,106 @@
+"""Contract tests for Engine protocol.
+
+These tests verify that Engine implementations satisfy the architectural requirements:
+- createSession() returns EngineSession
+- dispose() is idempotent
+- After dispose(), createSession() raises EngineError
+- Engine is thread-safe for createSession()
+"""
+
+import pytest
+
+from abogen.tts_plugin.engine import Engine, EngineSession
+from abogen.tts_plugin.errors import EngineError
+
+from .conftest import FakeEngine, FakeEngineSession
+
+
+class TestEngineProtocolContract:
+    """Contract tests for the Engine protocol itself."""
+
+    def test_engine_is_protocol(self) -> None:
+        assert hasattr(Engine, "__protocol_attrs__")
+
+    def test_engine_session_is_protocol(self) -> None:
+        assert hasattr(EngineSession, "__protocol_attrs__")
+
+    def test_fake_engine_satisfies_protocol(self) -> None:
+        engine = FakeEngine()
+        assert isinstance(engine, Engine)
+
+    def test_fake_session_satisfies_protocol(self) -> None:
+        session = FakeEngineSession()
+        assert isinstance(session, EngineSession)
+
+
+class TestEngineCreateSessionContract:
+    """Contract tests for Engine.createSession()."""
+
+    def test_create_session_returns_engine_session(self) -> None:
+        engine = FakeEngine()
+        session = engine.createSession()
+        assert isinstance(session, EngineSession)
+
+    def test_create_session_returns_new_instance(self) -> None:
+        engine = FakeEngine()
+        session1 = engine.createSession()
+        session2 = engine.createSession()
+        assert session1 is not session2
+
+    def test_create_session_ownership_transfers(self) -> None:
+        """Architecture spec: Ownership transfers to caller."""
+        engine = FakeEngine()
+        session = engine.createSession()
+        assert isinstance(session, EngineSession)
+
+
+class TestEngineDisposeContract:
+    """Contract tests for Engine.dispose()."""
+
+    def test_dispose_is_idempotent(self) -> None:
+        """Architecture spec: dispose() is idempotent."""
+        engine = FakeEngine()
+        engine.dispose()
+        engine.dispose()  # Should not raise
+
+    def test_dispose_never_raises(self) -> None:
+        """Architecture spec: dispose() never raises exceptions."""
+        engine = FakeEngine()
+        engine.dispose()  # Should not raise
+
+    def test_create_session_after_dispose_raises(self) -> None:
+        """Architecture spec: After dispose(), all methods except dispose() raise EngineError."""
+        engine = FakeEngine()
+        engine.dispose()
+        with pytest.raises(EngineError):
+            engine.createSession()
+
+
+class TestEngineLifecycleContract:
+    """Contract tests for Engine lifecycle."""
+
+    def test_full_lifecycle(self) -> None:
+        """Test complete engine lifecycle: create -> sessions -> dispose."""
+        engine = FakeEngine()
+
+        # Create sessions
+        session1 = engine.createSession()
+        session2 = engine.createSession()
+
+        # Use sessions
+        assert isinstance(session1, EngineSession)
+        assert isinstance(session2, EngineSession)
+
+        # Dispose sessions
+        session1.dispose()
+        session2.dispose()
+
+        # Dispose engine
+        engine.dispose()
+
+    def test_engine_disposed_session_raises(self) -> None:
+        """Architecture spec: After dispose(), all methods except dispose() raise EngineError."""
+        engine = FakeEngine()
+        engine.dispose()
+        with pytest.raises(EngineError):
+            engine.createSession()

tests/contracts/test_errors_contract.py

@@ -0,0 +1,85 @@
+"""Contract tests for error hierarchy.
+
+These tests verify that the error hierarchy satisfies the architectural requirements:
+- All errors inherit from EngineError
+- EngineError inherits from Exception
+- Each error type is properly classified
+"""
+
+import pytest
+
+from abogen.tts_plugin.errors import (
+    CancelledError,
+    ConfigurationError,
+    EngineError,
+    InternalError,
+    InvalidInputError,
+    ModelLoadError,
+    ModelNotFoundError,
+    NetworkError,
+)
+
+
+class TestErrorHierarchyContract:
+    """Contract tests for the error hierarchy."""
+
+    def test_engine_error_is_exception(self) -> None:
+        assert issubclass(EngineError, Exception)
+
+    def test_all_errors_inherit_from_engine_error(self) -> None:
+        error_classes = [
+            ModelNotFoundError,
+            ModelLoadError,
+            NetworkError,
+            InvalidInputError,
+            ConfigurationError,
+            CancelledError,
+            InternalError,
+        ]
+        for error_class in error_classes:
+            assert issubclass(error_class, EngineError), (
+                f"{error_class.__name__} must inherit from EngineError"
+            )
+
+    def test_all_errors_are_catchable(self) -> None:
+        error_classes = [
+            EngineError,
+            ModelNotFoundError,
+            ModelLoadError,
+            NetworkError,
+            InvalidInputError,
+            ConfigurationError,
+            CancelledError,
+            InternalError,
+        ]
+        for error_class in error_classes:
+            with pytest.raises(EngineError):
+                raise error_class("test message")
+
+    def test_error_message_preserved(self) -> None:
+        msg = "Model not found: bert-base"
+        with pytest.raises(ModelNotFoundError, match=msg):
+            raise ModelNotFoundError(msg)
+
+    def test_error_can_be_caught_as_engine_error(self) -> None:
+        with pytest.raises(EngineError):
+            raise ModelNotFoundError("test")
+
+    def test_cancelled_error_is_engine_error(self) -> None:
+        """CancelledError is a subtype of EngineError per architecture spec."""
+        assert issubclass(CancelledError, EngineError)
+
+    def test_error_hierarchy_no_cycles(self) -> None:
+        """Verify no circular inheritance."""
+        error_classes = [
+            EngineError,
+            ModelNotFoundError,
+            ModelLoadError,
+            NetworkError,
+            InvalidInputError,
+            ConfigurationError,
+            CancelledError,
+            InternalError,
+        ]
+        for cls in error_classes:
+            assert cls not in cls.__bases__

tests/contracts/test_host_context_contract.py

@@ -0,0 +1,89 @@
+"""Contract tests for HostContext.
+
+These tests verify that HostContext satisfies the architectural requirements:
+- Minimal (3 fields maximum)
+- Frozen dataclass
+- config_dir: Path
+- logger: Logger
+- http_client: HttpClient protocol
+"""
+
+import logging
+from pathlib import Path
+
+import pytest
+
+from abogen.tts_plugin.host_context import HttpClient, HostContext
+
+
+class TestHostContextContract:
+    """Contract tests for HostContext dataclass."""
+
+    def test_is_frozen_dataclass(self) -> None:
+        assert hasattr(HostContext, "__dataclass_params__")
+        assert HostContext.__dataclass_params__.frozen is True
+
+    def test_required_fields(self, tmp_path: Path) -> None:
+        logger = logging.getLogger("test")
+
+        class FakeClient:
+            def get(self, url: str, **kwargs: object) -> object:
+                return None
+
+            def post(self, url: str, **kwargs: object) -> object:
+                return None
+
+        ctx = HostContext(
+            config_dir=tmp_path,
+            logger=logger,
+            http_client=FakeClient(),
+        )
+        assert ctx.config_dir == tmp_path
+        assert ctx.logger is logger
+
+    def test_immutability(self, tmp_path: Path) -> None:
+        class FakeClient:
+            def get(self, url: str, **kwargs: object) -> object:
+                return None
+
+            def post(self, url: str, **kwargs: object) -> object:
+                return None
+
+        ctx = HostContext(
+            config_dir=tmp_path,
+            logger=logging.getLogger("test"),
+            http_client=FakeClient(),
+        )
+        with pytest.raises(AttributeError):
+            ctx.config_dir = Path("/other")  # type: ignore[misc]
+
+    def test_max_three_fields(self) -> None:
+        """Architecture spec: HostContext is minimal (3 fields max)."""
+        import dataclasses
+
+        fields = dataclasses.fields(HostContext)
+        assert len(fields) <= 3
+
+
+class TestHttpClientProtocolContract:
+    """Contract tests for HttpClient protocol."""
+
+    def test_http_client_is_protocol(self) -> None:
+        assert hasattr(HttpClient, "__protocol_attrs__")
+
+    def test_http_client_has_get(self) -> None:
+        assert hasattr(HttpClient, "get")
+
+    def test_http_client_has_post(self) -> None:
+        assert hasattr(HttpClient, "post")
+
+    def test_http_client_satisfied(self) -> None:
+        class FakeClient:
+            def get(self, url: str, **kwargs: object) -> object:
+                return None
+
+            def post(self, url: str, **kwargs: object) -> object:
+                return None
+
+        client = FakeClient()
+        assert isinstance(client, HttpClient)

tests/contracts/test_integration.py

@@ -0,0 +1,420 @@
+"""Integration tests for the TTS Plugin Architecture.
+
+These tests verify:
+1. Consumer Flow: consumer → plugin → engine → session → synthesis → result
+2. Lifecycle: dispose, no leaks, error handling
+3. Regression: old path vs new path equivalence
+
+Tests use mock plugins to avoid requiring real TTS dependencies.
+"""
+
+import pytest
+from typing import Any, Iterator
+from unittest.mock import MagicMock, patch
+
+import numpy as np
+
+from abogen.tts_plugin.engine import Engine, EngineSession
+from abogen.tts_plugin.errors import EngineError
+from abogen.tts_plugin.plugin_manager import PluginManager, get_plugin_manager, reset_plugin_manager
+from abogen.tts_plugin.utils import Pipeline, create_pipeline
+from abogen.tts_plugin.types import (
+    AudioFormat,
+    Duration,
+    ParameterValues,
+    SynthesisRequest,
+    SynthesizedAudio,
+    VoiceSelection,
+)
+
+
+class MockEngineSession:
+    """Mock EngineSession that records calls for verification."""
+    
+    def __init__(self):
+        self._disposed = False
+        self.synthesize_calls = []
+    
+    def synthesize(self, request: SynthesisRequest) -> SynthesizedAudio:
+        if self._disposed:
+            raise EngineError("Session disposed")
+        
+        self.synthesize_calls.append(request)
+        
+        # Return fake audio
+        audio = np.ones(1000, dtype=np.float32) * 0.5
+        return SynthesizedAudio(
+            data=audio.tobytes(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+            duration=Duration(seconds=1000 / 24000),
+        )
+    
+    def dispose(self) -> None:
+        self._disposed = True
+
+
+class MockEngine:
+    """Mock Engine that creates MockEngineSessions."""
+    
+    def __init__(self, **kwargs):
+        self.kwargs = kwargs
+        self._disposed = False
+        self.sessions_created = []
+    
+    def createSession(self) -> MockEngineSession:
+        if self._disposed:
+            raise EngineError("Engine disposed")
+        session = MockEngineSession()
+        self.sessions_created.append(session)
+        return session
+    
+    def dispose(self) -> None:
+        self._disposed = True
+
+
+def create_mock_plugin(create_engine_func=None):
+    """Helper to create a mock plugin module."""
+    if create_engine_func is None:
+        create_engine_func = lambda **kwargs: MockEngine(**kwargs)
+    
+    from abogen.tts_plugin.manifest import PluginManifest, EngineManifest
+    
+    manifest = PluginManifest(
+        id="mock_tts",
+        name="Mock TTS",
+        version="1.0.0",
+        api_version="1.0",
+        description="Mock TTS for testing",
+        author="Test",
+        capabilities=(),
+        requires=None,
+        engine=EngineManifest(
+            voiceSources=(),
+            parameters=(),
+            audioFormats=(),
+        ),
+    )
+    
+    return {
+        "PLUGIN_MANIFEST": manifest,
+        "MODEL_REQUIREMENTS": [],
+        "create_engine": create_mock_plugin_engine if create_engine_func is None else create_engine_func,
+    }
+
+
+def create_mock_plugin_engine(**kwargs):
+    """Default mock plugin engine factory."""
+    return MockEngine(**kwargs)
+
+
+class TestConsumerFlow:
+    """Consumer Flow Test: consumer → plugin → engine → session → synthesis → result"""
+    
+    def test_full_consumer_flow(self):
+        """Verify complete flow from consumer to audio output."""
+        manager = PluginManager()
+        
+        # Register mock plugin
+        mock_plugin = create_mock_plugin()
+        manager._plugins["mock_tts"] = mock_plugin
+        manager._loaded = True
+        
+        # Step 1: Consumer gets plugin
+        assert manager.has_plugin("mock_tts") is True
+        
+        # Step 2: Plugin creates engine
+        engine = manager.create_engine("mock_tts")
+        assert engine is not None
+        assert isinstance(engine, MockEngine)
+        
+        # Step 3: Engine creates session
+        session = engine.createSession()
+        assert session is not None
+        assert isinstance(session, MockEngineSession)
+        
+        # Step 4: Session synthesizes
+        request = SynthesisRequest(
+            text="Hello world",
+            voice=VoiceSelection(source="builtin", key="default"),
+            parameters=ParameterValues(values={"speed": 1.0}),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        result = session.synthesize(request)
+        
+        # Step 5: Result returned
+        assert result is not None
+        assert isinstance(result, SynthesizedAudio)
+        assert len(result.data) > 0
+        assert result.format.mime == "audio/wav"
+        assert result.duration.seconds > 0
+    
+    def test_consumer_flow_via_pipeline(self):
+        """Verify flow through Pipeline utility matches direct flow."""
+        manager = PluginManager()
+        
+        # Register mock plugin
+        mock_plugin = create_mock_plugin()
+        manager._plugins["mock_tts"] = mock_plugin
+        manager._loaded = True
+        
+        # Use Pipeline utility
+        with patch("abogen.tts_plugin.utils.get_plugin_manager", return_value=manager):
+            backend = create_pipeline("mock_tts")
+            
+            # Call like old TTSBackend
+            segments = list(backend("Hello world", voice="default", speed=1.0))
+            
+            # Verify result
+            assert len(segments) >= 1
+            segment = segments[0]
+            assert hasattr(segment, "graphemes")
+            assert hasattr(segment, "audio")
+            assert segment.graphemes == "Hello world"
+
+
+class TestLifecycle:
+    """Lifecycle Test: dispose, no leaks, error handling"""
+    
+    def test_session_dispose_is_idempotent(self):
+        """dispose() can be called multiple times safely."""
+        session = MockEngineSession()
+        
+        session.dispose()
+        session.dispose()  # Should not raise
+        assert session._disposed is True
+    
+    def test_session_synthesize_after_dispose_raises(self):
+        """synthesize() after dispose() raises EngineError."""
+        session = MockEngineSession()
+        session.dispose()
+        
+        request = SynthesisRequest(
+            text="test",
+            voice=VoiceSelection(source="builtin", key="default"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        
+        with pytest.raises(EngineError):
+            session.synthesize(request)
+    
+    def test_engine_dispose_is_idempotent(self):
+        """Engine dispose() can be called multiple times safely."""
+        engine = MockEngine()
+        
+        engine.dispose()
+        engine.dispose()  # Should not raise
+        assert engine._disposed is True
+    
+    def test_engine_create_session_after_dispose_raises(self):
+        """createSession() after dispose() raises EngineError."""
+        engine = MockEngine()
+        engine.dispose()
+        
+        with pytest.raises(EngineError):
+            engine.createSession()
+    
+    def test_full_lifecycle(self):
+        """Test complete lifecycle: create → use → dispose."""
+        engine = MockEngine()
+        
+        # Create and use session
+        session = engine.createSession()
+        request = SynthesisRequest(
+            text="test",
+            voice=VoiceSelection(source="builtin", key="default"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        result = session.synthesize(request)
+        assert len(result.data) > 0
+        
+        # Dispose session
+        session.dispose()
+        assert session._disposed is True
+        
+        # Dispose engine
+        engine.dispose()
+        assert engine._disposed is True
+    
+    def test_no_session_leak_on_engine_dispose(self):
+        """Engine can be disposed even if sessions were created."""
+        engine = MockEngine()
+        
+        # Create multiple sessions
+        session1 = engine.createSession()
+        session2 = engine.createSession()
+        
+        # Use sessions
+        request = SynthesisRequest(
+            text="test",
+            voice=VoiceSelection(source="builtin", key="default"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        session1.synthesize(request)
+        session2.synthesize(request)
+        
+        # Dispose engine (sessions still exist but engine is disposed)
+        engine.dispose()
+        assert engine._disposed is True
+        
+        # Sessions can still be used (they hold reference to pipeline)
+        result = session1.synthesize(request)
+        assert len(result.data) > 0
+    
+    def test_error_handling_in_synthesis(self):
+        """Error during synthesis is handled correctly."""
+        class FailingSession:
+            def synthesize(self, request):
+                raise EngineError("Synthesis failed")
+            
+            def dispose(self):
+                pass
+        
+        session = FailingSession()
+        request = SynthesisRequest(
+            text="test",
+            voice=VoiceSelection(source="builtin", key="default"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        
+        with pytest.raises(EngineError, match="Synthesis failed"):
+            session.synthesize(request)
+
+
+class TestRegression:
+    """Regression Test: old path vs new path equivalence"""
+    
+    def test_old_path_vs_new_path_same_result(self):
+        """Both paths should produce equivalent results."""
+        # Setup mock plugin
+        manager = PluginManager()
+        mock_plugin = create_mock_plugin()
+        manager._plugins["mock_tts"] = mock_plugin
+        manager._loaded = True
+        
+        # New path: Plugin Manager → Engine → Session → Synthesis
+        with patch("abogen.tts_plugin.utils.get_plugin_manager", return_value=manager):
+            new_backend = create_pipeline("mock_tts")
+            new_segments = list(new_backend("Hello world", voice="default", speed=1.0))
+        
+        # Old path: Direct MockEngine (simulating old registry)
+        old_engine = MockEngine()
+        old_session = old_engine.createSession()
+        request = SynthesisRequest(
+            text="Hello world",
+            voice=VoiceSelection(source="builtin", key="default"),
+            parameters=ParameterValues(values={"speed": 1.0}),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        old_result = old_session.synthesize(request)
+        
+        # Compare results
+        # New path returns segments, old path returns SynthesizedAudio
+        # But both should have valid audio data
+        assert len(new_segments) >= 1
+        assert len(old_result.data) > 0
+        
+        # Both should have same format
+        assert new_segments[0].audio.dtype == np.float32
+    
+    def test_pipeline_matches_old_interface(self):
+        """Pipeline utility should match old TTSBackend interface."""
+        manager = PluginManager()
+        mock_plugin = create_mock_plugin()
+        manager._plugins["mock_tts"] = mock_plugin
+        manager._loaded = True
+        
+        with patch("abogen.tts_plugin.utils.get_plugin_manager", return_value=manager):
+            backend = create_pipeline("mock_tts", lang_code="a", device="cpu")
+            
+            # Old interface: pipeline(text, voice=..., speed=..., split_pattern=...)
+            segments = list(backend(
+                "Hello world",
+                voice="af_heart",
+                speed=1.0,
+                split_pattern=r"\n+"
+            ))
+            
+            # Should return segments with graphemes and audio
+            assert len(segments) >= 1
+            segment = segments[0]
+            assert segment.graphemes == "Hello world"
+            assert isinstance(segment.audio, np.ndarray)
+            assert segment.audio.dtype == np.float32
+            assert len(segment.audio) > 0
+
+
+class TestPluginManagerIntegration:
+    """Integration tests for PluginManager."""
+    
+    def test_plugin_manager_singleton_pattern(self):
+        """Global plugin manager follows singleton pattern."""
+        reset_plugin_manager()
+        
+        manager1 = get_plugin_manager()
+        manager2 = get_plugin_manager()
+        
+        assert manager1 is manager2
+        
+        reset_plugin_manager()
+        
+        manager3 = get_plugin_manager()
+        assert manager1 is not manager3
+    
+    def test_plugin_manager_discover_plugins(self):
+        """Plugin manager can discover plugins from directory."""
+        manager = PluginManager()
+        
+        # Discover from test plugins directory
+        manager.discover("tests/plugins")
+        
+        # Should find valid_plugin
+        # (This depends on test plugins existing)
+        plugins = manager.list_plugins()
+        assert isinstance(plugins, list)
+    
+    def test_plugin_manager_dispose_all(self):
+        """Plugin manager can dispose all cached engines."""
+        manager = PluginManager()
+        
+        # Register mock plugin
+        mock_plugin = create_mock_plugin()
+        manager._plugins["mock_tts"] = mock_plugin
+        manager._loaded = True
+        
+        # Create engines
+        engine1 = manager.get_or_create_engine("mock_tts")
+        engine2 = manager.get_or_create_engine("mock_tts")
+        
+        # Dispose all
+        manager.dispose_all()
+        
+        # Engines should be disposed
+        assert engine1._disposed is True
+        assert engine2._disposed is True
+        
+        # Cache should be empty
+        assert len(manager._engines) == 0
+
+
+class TestNoCompatLayer:
+    """Regression: confirm the compatibility layer has been removed."""
+
+    def test_compat_module_does_not_exist(self):
+        """abogen.tts_plugin.compat must not be importable."""
+        import importlib
+        with pytest.raises((ImportError, ModuleNotFoundError)):
+            importlib.import_module("abogen.tts_plugin.compat")
+
+    def test_consumers_use_plugin_architecture_directly(self):
+        """Key consumers import from abogen.tts_plugin.utils, not compat."""
+        import inspect, abogen.voice_profiles, abogen.voice_formulas, abogen.voice_cache
+
+        for mod in (abogen.voice_profiles, abogen.voice_formulas, abogen.voice_cache):
+            source = inspect.getsource(mod)
+            assert "tts_plugin.compat" not in source, (
+                f"{mod.__name__} still references tts_plugin.compat"
+            )

tests/contracts/test_loader_contract.py

@@ -0,0 +1,436 @@
+"""Comprehensive tests for the plugin loader infrastructure.
+
+These tests verify that the loader correctly:
+- Discovers plugins in directories
+- Imports plugin modules
+- Validates PLUGIN_MANIFEST, MODEL_REQUIREMENTS, create_engine
+- Validates api_version compatibility
+- Validates capabilities
+- Provides diagnostic messages for errors
+- Rejects invalid plugins
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pytest
+
+from abogen.tts_plugin.loader import (
+    HOST_API_VERSION,
+    PluginLoadError,
+    PluginLoadResult,
+    _check_api_version_compatibility,
+    _parse_api_version,
+    _validate_api_version,
+    _validate_capabilities,
+    _validate_manifest,
+    discover_plugins,
+    load_plugin,
+    load_plugin_from_dir,
+)
+from abogen.tts_plugin.manifest import (
+    EngineManifest,
+    ModelManifest,
+    PluginManifest,
+)
+
+
+# ──────────────────────────────────────────────────────────────
+# Path fixtures
+# ──────────────────────────────────────────────────────────────
+
+@pytest.fixture
+def plugins_dir() -> Path:
+    return Path(__file__).parent.parent / "plugins"
+
+
+@pytest.fixture
+def fake_plugin_dir(plugins_dir: Path) -> Path:
+    return plugins_dir / "fake_plugin"
+
+
+@pytest.fixture
+def missing_manifest_dir(plugins_dir: Path) -> Path:
+    return plugins_dir / "missing_manifest"
+
+
+@pytest.fixture
+def invalid_api_version_dir(plugins_dir: Path) -> Path:
+    return plugins_dir / "invalid_api_version"
+
+
+@pytest.fixture
+def invalid_capabilities_dir(plugins_dir: Path) -> Path:
+    return plugins_dir / "invalid_capabilities"
+
+
+@pytest.fixture
+def missing_create_engine_dir(plugins_dir: Path) -> Path:
+    return plugins_dir / "missing_create_engine"
+
+
+@pytest.fixture
+def import_error_dir(plugins_dir: Path) -> Path:
+    return plugins_dir / "import_error"
+
+
+@pytest.fixture
+def missing_model_requirements_dir(plugins_dir: Path) -> Path:
+    return plugins_dir / "missing_model_requirements"
+
+
+# ──────────────────────────────────────────────────────────────
+# Unit tests: _parse_api_version
+# ──────────────────────────────────────────────────────────────
+
+class TestParseApiVersion:
+    def test_valid_version(self) -> None:
+        assert _parse_api_version("1.0") == (1, 0)
+        assert _parse_api_version("2.5") == (2, 5)
+        assert _parse_api_version("10.20") == (10, 20)
+
+    def test_invalid_format(self) -> None:
+        assert _parse_api_version("1") is None
+        assert _parse_api_version("1.0.0") is None
+        assert _parse_api_version("abc") is None
+        assert _parse_api_version("") is None
+        assert _parse_api_version("1.x") is None
+
+
+# ──────────────────────────────────────────────────────────────
+# Unit tests: _check_api_version_compatibility
+# ──────────────────────────────────────────────────────────────
+
+class TestCheckApiVersionCompatibility:
+    def test_compatible_version(self) -> None:
+        assert _check_api_version_compatibility("1.0") is None
+        assert _check_api_version_compatibility("1.5") is None
+
+    def test_major_mismatch(self) -> None:
+        error = _check_api_version_compatibility("2.0")
+        assert error is not None
+        assert "major mismatch" in error
+
+    def test_invalid_format(self) -> None:
+        error = _check_api_version_compatibility("invalid")
+        assert error is not None
+        assert "Invalid api_version format" in error
+
+
+# ──────────────────────────────────────────────────────────────
+# Unit tests: _validate_manifest
+# ──────────────────────────────────────────────────────────────
+
+class TestValidateManifest:
+    def test_valid_manifest(self) -> None:
+        class FakeModule:
+            PLUGIN_MANIFEST = PluginManifest(
+                id="test", name="Test", version="1.0.0",
+                api_version="1.0", description="Test", author="Test",
+            )
+            MODEL_REQUIREMENTS: list = []
+            create_engine = lambda *a, **kw: None
+
+        errors = _validate_manifest(FakeModule(), Path("/tmp"))
+        assert errors == []
+
+    def test_missing_manifest(self) -> None:
+        class FakeModule:
+            MODEL_REQUIREMENTS: list = []
+            create_engine = lambda *a, **kw: None
+
+        errors = _validate_manifest(FakeModule(), Path("/tmp"))
+        assert any("Missing PLUGIN_MANIFEST" in e for e in errors)
+
+    def test_wrong_manifest_type(self) -> None:
+        class FakeModule:
+            PLUGIN_MANIFEST = "not a manifest"
+            MODEL_REQUIREMENTS: list = []
+            create_engine = lambda *a, **kw: None
+
+        errors = _validate_manifest(FakeModule(), Path("/tmp"))
+        assert any("PluginManifest instance" in e for e in errors)
+
+    def test_missing_model_requirements(self) -> None:
+        class FakeModule:
+            PLUGIN_MANIFEST = PluginManifest(
+                id="test", name="Test", version="1.0.0",
+                api_version="1.0", description="Test", author="Test",
+            )
+            create_engine = lambda *a, **kw: None
+
+        errors = _validate_manifest(FakeModule(), Path("/tmp"))
+        assert any("Missing MODEL_REQUIREMENTS" in e for e in errors)
+
+    def test_wrong_model_requirements_type(self) -> None:
+        class FakeModule:
+            PLUGIN_MANIFEST = PluginManifest(
+                id="test", name="Test", version="1.0.0",
+                api_version="1.0", description="Test", author="Test",
+            )
+            MODEL_REQUIREMENTS = "not a list"
+            create_engine = lambda *a, **kw: None
+
+        errors = _validate_manifest(FakeModule(), Path("/tmp"))
+        assert any("must be a list" in e for e in errors)
+
+    def test_invalid_model_requirements_item(self) -> None:
+        class FakeModule:
+            PLUGIN_MANIFEST = PluginManifest(
+                id="test", name="Test", version="1.0.0",
+                api_version="1.0", description="Test", author="Test",
+            )
+            MODEL_REQUIREMENTS = ["not a model manifest"]
+            create_engine = lambda *a, **kw: None
+
+        errors = _validate_manifest(FakeModule(), Path("/tmp"))
+        assert any("ModelManifest instance" in e for e in errors)
+
+    def test_missing_create_engine(self) -> None:
+        class FakeModule:
+            PLUGIN_MANIFEST = PluginManifest(
+                id="test", name="Test", version="1.0.0",
+                api_version="1.0", description="Test", author="Test",
+            )
+            MODEL_REQUIREMENTS: list = []
+
+        errors = _validate_manifest(FakeModule(), Path("/tmp"))
+        assert any("Missing create_engine" in e for e in errors)
+
+    def test_create_engine_not_callable(self) -> None:
+        class FakeModule:
+            PLUGIN_MANIFEST = PluginManifest(
+                id="test", name="Test", version="1.0.0",
+                api_version="1.0", description="Test", author="Test",
+            )
+            MODEL_REQUIREMENTS: list = []
+            create_engine = "not callable"
+
+        errors = _validate_manifest(FakeModule(), Path("/tmp"))
+        assert any("must be callable" in e for e in errors)
+
+
+# ──────────────────────────────────────────────────────────────
+# Unit tests: _validate_capabilities
+# ──────────────────────────────────────────────────────────────
+
+class TestValidateCapabilities:
+    def test_valid_capabilities(self) -> None:
+        manifest = PluginManifest(
+            id="test", name="Test", version="1.0.0",
+            api_version="1.0", description="Test", author="Test",
+            capabilities=("voice_list", "preview"),
+        )
+        errors = _validate_capabilities(manifest)
+        assert errors == []
+
+    def test_unknown_capability(self) -> None:
+        manifest = PluginManifest(
+            id="test", name="Test", version="1.0.0",
+            api_version="1.0", description="Test", author="Test",
+            capabilities=("voice_list", "unknown_cap"),
+        )
+        errors = _validate_capabilities(manifest)
+        assert any("unknown_cap" in e for e in errors)
+
+    def test_empty_capabilities(self) -> None:
+        manifest = PluginManifest(
+            id="test", name="Test", version="1.0.0",
+            api_version="1.0", description="Test", author="Test",
+            capabilities=(),
+        )
+        errors = _validate_capabilities(manifest)
+        assert errors == []
+
+
+# ──────────────────────────────────────────────────────────────
+# Unit tests: _validate_api_version
+# ──────────────────────────────────────────────────────────────
+
+class TestValidateApiVersion:
+    def test_compatible(self) -> None:
+        manifest = PluginManifest(
+            id="test", name="Test", version="1.0.0",
+            api_version="1.0", description="Test", author="Test",
+        )
+        errors = _validate_api_version(manifest)
+        assert errors == []
+
+    def test_incompatible(self) -> None:
+        manifest = PluginManifest(
+            id="test", name="Test", version="1.0.0",
+            api_version="2.0", description="Test", author="Test",
+        )
+        errors = _validate_api_version(manifest)
+        assert len(errors) > 0
+
+
+# ──────────────────────────────────────────────────────────────
+# Integration tests: load_plugin_from_dir
+# ──────────────────────────────────────────────────────────────
+
+class TestLoadPluginFromDir:
+    def test_load_valid_plugin(self, fake_plugin_dir: Path) -> None:
+        result = load_plugin_from_dir(fake_plugin_dir)
+        assert result.success is True
+        assert result.manifest is not None
+        assert result.manifest.id == "fake_plugin"
+        assert result.model_requirements is not None
+        assert result.create_engine is not None
+        assert result.module is not None
+        assert result.error is None
+
+    def test_plugin_satisfies_protocol(self, fake_plugin_dir: Path) -> None:
+        from abogen.tts_plugin.engine import Engine
+        from abogen.tts_plugin.host_context import HostContext
+        import logging
+
+        result = load_plugin_from_dir(fake_plugin_dir)
+        assert result.success is True
+
+        # Create engine using the loaded create_engine function
+        ctx = HostContext(
+            config_dir=Path("/tmp/test"),
+            logger=logging.getLogger("test"),
+            http_client=type("FakeClient", (), {"get": lambda s, **kw: None, "post": lambda s, **kw: None})(),
+        )
+        engine = result.create_engine(ctx, None, __import__("abogen.tts_plugin.types", fromlist=["EngineConfig"]).EngineConfig())
+        assert isinstance(engine, Engine)
+        engine.dispose()
+
+    def test_nonexistent_directory(self, tmp_path: Path) -> None:
+        result = load_plugin_from_dir(tmp_path / "nonexistent")
+        assert result.success is False
+        assert result.error is not None
+        assert "does not exist" in result.error.errors[0]
+
+    def test_missing_init_file(self, tmp_path: Path) -> None:
+        plugin_dir = tmp_path / "no_init"
+        plugin_dir.mkdir()
+        result = load_plugin_from_dir(plugin_dir)
+        assert result.success is False
+        assert result.error is not None
+        assert "__init__.py" in result.error.errors[0]
+
+    def test_import_error(self, import_error_dir: Path) -> None:
+        result = load_plugin_from_dir(import_error_dir)
+        assert result.success is False
+        assert result.error is not None
+        assert "Failed to import" in result.error.errors[0]
+
+
+# ──────────────────────────────────────────────────────────────
+# Integration tests: invalid plugins
+# ──────────────────────────────────────────────────────────────
+
+class TestInvalidPlugins:
+    def test_missing_manifest(self, missing_manifest_dir: Path) -> None:
+        result = load_plugin_from_dir(missing_manifest_dir)
+        assert result.success is False
+        assert result.error is not None
+        assert any("Missing PLUGIN_MANIFEST" in e for e in result.error.errors)
+
+    def test_invalid_api_version(self, invalid_api_version_dir: Path) -> None:
+        result = load_plugin_from_dir(invalid_api_version_dir)
+        assert result.success is False
+        assert result.error is not None
+        assert any("major mismatch" in e for e in result.error.errors)
+
+    def test_invalid_capabilities(self, invalid_capabilities_dir: Path) -> None:
+        result = load_plugin_from_dir(invalid_capabilities_dir)
+        assert result.success is False
+        assert result.error is not None
+        assert any("Unknown capability" in e for e in result.error.errors)
+
+    def test_missing_create_engine(self, missing_create_engine_dir: Path) -> None:
+        result = load_plugin_from_dir(missing_create_engine_dir)
+        assert result.success is False
+        assert result.error is not None
+        assert any("Missing create_engine" in e for e in result.error.errors)
+
+    def test_missing_model_requirements(self, missing_model_requirements_dir: Path) -> None:
+        result = load_plugin_from_dir(missing_model_requirements_dir)
+        assert result.success is False
+        assert result.error is not None
+        assert any("Missing MODEL_REQUIREMENTS" in e for e in result.error.errors)
+
+
+# ──────────────────────────────────────────────────────────────
+# Integration tests: discover_plugins
+# ──────────────────────────────────────────────────────────────
+
+class TestDiscoverPlugins:
+    def test_discover_from_valid_dir(self, plugins_dir: Path) -> None:
+        results = discover_plugins([plugins_dir])
+        # Should find multiple plugins (valid and invalid)
+        assert len(results) > 0
+
+    def test_discover_includes_valid_plugin(self, plugins_dir: Path) -> None:
+        results = discover_plugins([plugins_dir])
+        valid = [r for r in results if r.success]
+        assert len(valid) >= 1
+        assert any(r.manifest and r.manifest.id == "fake_plugin" for r in valid)
+
+    def test_discover_includes_invalid_plugins(self, plugins_dir: Path) -> None:
+        results = discover_plugins([plugins_dir])
+        invalid = [r for r in results if not r.success]
+        assert len(invalid) >= 1
+
+    def test_discover_nonexistent_dir(self, tmp_path: Path) -> None:
+        results = discover_plugins([tmp_path / "nonexistent"])
+        assert results == []
+
+    def test_discover_multiple_dirs(self, plugins_dir: Path, tmp_path: Path) -> None:
+        results = discover_plugins([plugins_dir, tmp_path / "nonexistent"])
+        assert len(results) > 0
+
+
+# ──────────────────────────────────────────────────────────────
+# Diagnostic messages tests
+# ──────────────────────────────────────────────────────────────
+
+class TestDiagnosticMessages:
+    def test_error_contains_plugin_id(self, missing_manifest_dir: Path) -> None:
+        result = load_plugin_from_dir(missing_manifest_dir)
+        assert result.error is not None
+        assert result.error.plugin_id == "missing_manifest"
+
+    def test_error_contains_path(self, missing_manifest_dir: Path) -> None:
+        result = load_plugin_from_dir(missing_manifest_dir)
+        assert result.error is not None
+        assert result.error.path == missing_manifest_dir
+
+    def test_error_contains_messages(self, missing_manifest_dir: Path) -> None:
+        result = load_plugin_from_dir(missing_manifest_dir)
+        assert result.error is not None
+        assert len(result.error.errors) > 0
+
+    def test_multiple_errors(self, invalid_api_version_dir: Path) -> None:
+        # This plugin has multiple issues
+        result = load_plugin_from_dir(invalid_api_version_dir)
+        assert result.error is not None
+        # Should have at least the api_version error
+        assert len(result.error.errors) >= 1
+
+
+# ──────────────────────────────────────────────────────────────
+# No partial registration tests
+# ──────────────────────────────────────────────────────────────
+
+class TestNoPartialRegistration:
+    def test_invalid_plugin_no_manifest_attr(self, missing_manifest_dir: Path) -> None:
+        """After failed load, module should not remain in sys.modules."""
+        result = load_plugin_from_dir(missing_manifest_dir)
+        assert result.success is False
+        # Module should not be registered
+        module_name = f"abogen.tts_plugin._loaded.missing_manifest"
+        assert module_name not in sys.modules
+
+    def test_import_error_no_registration(self, import_error_dir: Path) -> None:
+        """After import error, module should not remain in sys.modules."""
+        result = load_plugin_from_dir(import_error_dir)
+        assert result.success is False
+        module_name = f"abogen.tts_plugin._loaded.import_error"
+        assert module_name not in sys.modules

tests/contracts/test_manifest_contract.py

@@ -0,0 +1,290 @@
+"""Contract tests for plugin manifest types.
+
+These tests verify that manifest types satisfy the architectural requirements:
+- All required fields are present
+- api_version follows semver format
+- capabilities are properly defined
+- engine manifest describes the engine correctly
+"""
+
+import re
+
+import pytest
+
+from abogen.tts_plugin.manifest import (
+    AudioFormatManifest,
+    EngineManifest,
+    EnumOption,
+    GpuRequirement,
+    ModelManifest,
+    ParameterManifest,
+    PluginManifest,
+    RequirementManifest,
+    VoiceManifest,
+    VoiceSourceManifest,
+)
+
+
+class TestPluginManifestContract:
+    """Contract tests for PluginManifest."""
+
+    def test_required_fields(self) -> None:
+        manifest = PluginManifest(
+            id="test-plugin",
+            name="Test Plugin",
+            version="1.0.0",
+            api_version="1.0",
+            description="A test plugin",
+            author="Test Author",
+        )
+        assert manifest.id == "test-plugin"
+        assert manifest.name == "Test Plugin"
+        assert manifest.version == "1.0.0"
+        assert manifest.api_version == "1.0"
+        assert manifest.description == "A test plugin"
+        assert manifest.author == "Test Author"
+
+    def test_api_version_semver_format(self) -> None:
+        """Architecture spec: api_version format is semver (MAJOR.MINOR)."""
+        valid_versions = ["1.0", "2.1", "10.5"]
+        for version in valid_versions:
+            manifest = PluginManifest(
+                id="test",
+                name="Test",
+                version="1.0.0",
+                api_version=version,
+                description="Test",
+                author="Test",
+            )
+            assert re.match(r"^\d+\.\d+$", manifest.api_version)
+
+    def test_capabilities_default_empty(self) -> None:
+        manifest = PluginManifest(
+            id="test",
+            name="Test",
+            version="1.0.0",
+            api_version="1.0",
+            description="Test",
+            author="Test",
+        )
+        assert manifest.capabilities == ()
+
+    def test_capabilities_tuple(self) -> None:
+        manifest = PluginManifest(
+            id="test",
+            name="Test",
+            version="1.0.0",
+            api_version="1.0",
+            description="Test",
+            author="Test",
+            capabilities=("voice_list", "preview"),
+        )
+        assert "voice_list" in manifest.capabilities
+        assert "preview" in manifest.capabilities
+
+    def test_requires_default(self) -> None:
+        manifest = PluginManifest(
+            id="test",
+            name="Test",
+            version="1.0.0",
+            api_version="1.0",
+            description="Test",
+            author="Test",
+        )
+        assert isinstance(manifest.requires, RequirementManifest)
+
+    def test_engine_default(self) -> None:
+        manifest = PluginManifest(
+            id="test",
+            name="Test",
+            version="1.0.0",
+            api_version="1.0",
+            description="Test",
+            author="Test",
+        )
+        assert isinstance(manifest.engine, EngineManifest)
+
+
+class TestEngineManifestContract:
+    """Contract tests for EngineManifest."""
+
+    def test_required_fields(self) -> None:
+        manifest = EngineManifest(
+            voiceSources=(
+                VoiceSourceManifest(id="builtin", name="Builtin", type="list"),
+            ),
+            parameters=(
+                ParameterManifest(
+                    id="speed", name="Speed", description="Speed", type="float", default=1.0
+                ),
+            ),
+            audioFormats=(AudioFormatManifest(mime="audio/wav", extension="wav"),),
+        )
+        assert len(manifest.voiceSources) == 1
+        assert len(manifest.parameters) == 1
+        assert len(manifest.audioFormats) == 1
+
+    def test_defaults_empty(self) -> None:
+        manifest = EngineManifest()
+        assert manifest.voiceSources == ()
+        assert manifest.parameters == ()
+        assert manifest.audioFormats == ()
+
+
+class TestVoiceSourceManifestContract:
+    """Contract tests for VoiceSourceManifest."""
+
+    def test_required_fields(self) -> None:
+        vs = VoiceSourceManifest(id="builtin", name="Builtin", type="list")
+        assert vs.id == "builtin"
+        assert vs.name == "Builtin"
+        assert vs.type == "list"
+
+    def test_valid_types(self) -> None:
+        valid_types = ["list", "speaker_id", "clone", "blend", "generate", "none"]
+        for vtype in valid_types:
+            vs = VoiceSourceManifest(id="test", name="Test", type=vtype)
+            assert vs.type == vtype
+
+    def test_config_optional(self) -> None:
+        vs = VoiceSourceManifest(id="test", name="Test", type="list")
+        assert vs.config is None
+
+    def test_config_any(self) -> None:
+        config = {"voices": ["af_nova", "af_sky"]}
+        vs = VoiceSourceManifest(id="test", name="Test", type="list", config=config)
+        assert vs.config == config
+
+
+class TestVoiceManifestContract:
+    """Contract tests for VoiceManifest."""
+
+    def test_required_fields(self) -> None:
+        v = VoiceManifest(id="af_nova", name="Nova")
+        assert v.id == "af_nova"
+        assert v.name == "Nova"
+
+    def test_tags_default_empty(self) -> None:
+        v = VoiceManifest(id="af_nova", name="Nova")
+        assert v.tags == ()
+
+    def test_tags_tuple(self) -> None:
+        v = VoiceManifest(id="af_nova", name="Nova", tags=("en", "female"))
+        assert "en" in v.tags
+        assert "female" in v.tags
+
+
+class TestParameterManifestContract:
+    """Contract tests for ParameterManifest."""
+
+    def test_required_fields(self) -> None:
+        p = ParameterManifest(
+            id="speed", name="Speed", description="Speech speed", type="float", default=1.0
+        )
+        assert p.id == "speed"
+        assert p.name == "Speed"
+        assert p.description == "Speech speed"
+        assert p.type == "float"
+        assert p.default == 1.0
+
+    def test_valid_types(self) -> None:
+        valid_types = ["float", "int", "string", "boolean", "enum"]
+        for ptype in valid_types:
+            p = ParameterManifest(
+                id="test", name="Test", description="Test", type=ptype, default=None
+            )
+            assert p.type == ptype
+
+    def test_optional_numeric_bounds(self) -> None:
+        p = ParameterManifest(
+            id="speed",
+            name="Speed",
+            description="Speed",
+            type="float",
+            default=1.0,
+            min=0.5,
+            max=2.0,
+            step=0.1,
+        )
+        assert p.min == 0.5
+        assert p.max == 2.0
+        assert p.step == 0.1
+
+    def test_enum_options(self) -> None:
+        options = (
+            EnumOption(value="low", label="Low"),
+            EnumOption(value="high", label="High"),
+        )
+        p = ParameterManifest(
+            id="quality",
+            name="Quality",
+            description="Quality",
+            type="enum",
+            default="low",
+            options=options,
+        )
+        assert len(p.options) == 2
+        assert p.options[0].value == "low"
+
+
+class TestAudioFormatManifestContract:
+    """Contract tests for AudioFormatManifest."""
+
+    def test_required_fields(self) -> None:
+        af = AudioFormatManifest(mime="audio/wav", extension="wav")
+        assert af.mime == "audio/wav"
+        assert af.extension == "wav"
+
+
+class TestEnumOptionContract:
+    """Contract tests for EnumOption."""
+
+    def test_required_fields(self) -> None:
+        opt = EnumOption(value="low", label="Low Quality")
+        assert opt.value == "low"
+        assert opt.label == "Low Quality"
+
+
+class TestRequirementManifestContract:
+    """Contract tests for RequirementManifest."""
+
+    def test_defaults(self) -> None:
+        req = RequirementManifest()
+        assert req.gpu is None
+        assert req.memory is None
+        assert req.internet is None
+
+    def test_with_gpu(self) -> None:
+        gpu = GpuRequirement(required=True, type="cuda", memory=8.0)
+        req = RequirementManifest(gpu=gpu)
+        assert req.gpu.required is True
+        assert req.gpu.type == "cuda"
+        assert req.gpu.memory == 8.0
+
+    def test_with_internet(self) -> None:
+        req = RequirementManifest(internet=True)
+        assert req.internet is True
+
+
+class TestGpuRequirementContract:
+    """Contract tests for GpuRequirement."""
+
+    def test_defaults(self) -> None:
+        gpu = GpuRequirement()
+        assert gpu.required is False
+        assert gpu.type is None
+        assert gpu.memory is None
+
+    def test_required_gpu(self) -> None:
+        gpu = GpuRequirement(required=True, type="cuda", memory=8.0)
+        assert gpu.required is True
+
+
+class TestModelManifestContract:
+    """Contract tests for ModelManifest."""
+
+    def test_required_fields(self) -> None:
+        m = ModelManifest(id="xtts_v2", name="XTTS v2", size="2GB")
+        assert m.id == "xtts_v2"
+        assert m.name == "XTTS v2"
+        assert m.size == "2GB"

tests/contracts/test_plugin_contract.py

@@ -0,0 +1,146 @@
+"""Contract tests for plugin contract.
+
+These tests verify that plugin modules satisfy the architectural requirements:
+- Must export PLUGIN_MANIFEST: PluginManifest
+- Must export MODEL_REQUIREMENTS: list[ModelManifest]
+- Must export create_engine: Callable[[HostContext, Path | None, EngineConfig], Engine]
+- create_engine() must be atomic
+"""
+
+import logging
+from pathlib import Path
+from typing import Any
+
+import pytest
+
+from abogen.tts_plugin.engine import Engine
+from abogen.tts_plugin.host_context import HostContext
+from abogen.tts_plugin.manifest import EngineManifest, ModelManifest, PluginManifest
+from abogen.tts_plugin.plugin import Plugin
+from abogen.tts_plugin.types import EngineConfig
+
+from .conftest import FakeEngine
+
+
+class FakePluginModule:
+    """Stub plugin module that satisfies the plugin contract."""
+
+    PLUGIN_MANIFEST = PluginManifest(
+        id="fake-plugin",
+        name="Fake Plugin",
+        version="1.0.0",
+        api_version="1.0",
+        description="A fake plugin for testing",
+        author="Test Author",
+        capabilities=(),
+        engine=EngineManifest(),
+    )
+
+    MODEL_REQUIREMENTS: list[ModelManifest] = []
+
+    @staticmethod
+    def create_engine(
+        context: HostContext,
+        model_path: Path | None,
+        config: EngineConfig,
+    ) -> Engine:
+        return FakeEngine()
+
+
+class TestPluginProtocolContract:
+    """Contract tests for the Plugin protocol."""
+
+    def test_plugin_is_protocol(self) -> None:
+        assert hasattr(Plugin, "__protocol_attrs__")
+
+
+class TestPluginExportsContract:
+    """Contract tests for required plugin exports."""
+
+    def test_plugin_has_plugin_manifest(self) -> None:
+        """Architecture spec: Plugin must export PLUGIN_MANIFEST."""
+        assert hasattr(FakePluginModule, "PLUGIN_MANIFEST")
+        assert isinstance(FakePluginModule.PLUGIN_MANIFEST, PluginManifest)
+
+    def test_plugin_has_model_requirements(self) -> None:
+        """Architecture spec: Plugin must export MODEL_REQUIREMENTS."""
+        assert hasattr(FakePluginModule, "MODEL_REQUIREMENTS")
+        assert isinstance(FakePluginModule.MODEL_REQUIREMENTS, list)
+
+    def test_plugin_has_create_engine(self) -> None:
+        """Architecture spec: Plugin must export create_engine."""
+        assert hasattr(FakePluginModule, "create_engine")
+        assert callable(FakePluginModule.create_engine)
+
+    def test_plugin_manifest_required_fields(self) -> None:
+        """Architecture spec: PluginManifest has required fields."""
+        manifest = FakePluginModule.PLUGIN_MANIFEST
+        assert manifest.id
+        assert manifest.name
+        assert manifest.version
+        assert manifest.api_version
+        assert manifest.description
+        assert manifest.author
+
+    def test_plugin_manifest_capabilities_is_tuple(self) -> None:
+        manifest = FakePluginModule.PLUGIN_MANIFEST
+        assert isinstance(manifest.capabilities, tuple)
+
+
+class TestCreateEngineContract:
+    """Contract tests for create_engine() function."""
+
+    def test_create_engine_returns_engine(self) -> None:
+        """Architecture spec: create_engine() returns Engine."""
+        ctx = HostContext(
+            config_dir=Path("/tmp/test"),
+            logger=logging.getLogger("test"),
+            http_client=type("FakeClient", (), {"get": lambda self, **kw: None, "post": lambda self, **kw: None})(),
+        )
+        engine = FakePluginModule.create_engine(ctx, None, EngineConfig())
+        assert isinstance(engine, Engine)
+
+    def test_create_engine_atomic(self) -> None:
+        """Architecture spec: create_engine() is atomic (all-or-nothing)."""
+        ctx = HostContext(
+            config_dir=Path("/tmp/test"),
+            logger=logging.getLogger("test"),
+            http_client=type("FakeClient", (), {"get": lambda self, **kw: None, "post": lambda self, **kw: None})(),
+        )
+        engine = FakePluginModule.create_engine(ctx, None, EngineConfig())
+        assert isinstance(engine, Engine)
+        engine.dispose()
+
+    def test_create_engine_with_none_model_path(self) -> None:
+        """Architecture spec: model_path can be None for cloud/no-model engines."""
+        ctx = HostContext(
+            config_dir=Path("/tmp/test"),
+            logger=logging.getLogger("test"),
+            http_client=type("FakeClient", (), {"get": lambda self, **kw: None, "post": lambda self, **kw: None})(),
+        )
+        engine = FakePluginModule.create_engine(ctx, None, EngineConfig())
+        assert isinstance(engine, Engine)
+        engine.dispose()
+
+    def test_create_engine_with_model_path(self) -> None:
+        """Architecture spec: model_path is Path | None."""
+        ctx = HostContext(
+            config_dir=Path("/tmp/test"),
+            logger=logging.getLogger("test"),
+            http_client=type("FakeClient", (), {"get": lambda self, **kw: None, "post": lambda self, **kw: None})(),
+        )
+        engine = FakePluginModule.create_engine(ctx, Path("/models/test"), EngineConfig())
+        assert isinstance(engine, Engine)
+        engine.dispose()
+
+
+class TestModelRequirementsContract:
+    """Contract tests for MODEL_REQUIREMENTS."""
+
+    def test_model_requirements_is_list(self) -> None:
+        assert isinstance(FakePluginModule.MODEL_REQUIREMENTS, list)
+
+    def test_model_requirements_contains_model_manifests(self) -> None:
+        """If non-empty, each item must be a ModelManifest."""
+        for req in FakePluginModule.MODEL_REQUIREMENTS:
+            assert isinstance(req, ModelManifest)

tests/contracts/test_plugin_manager_contract.py

@@ -0,0 +1,274 @@
+"""Integration tests for Plugin Manager and direct utility functions."""
+
+import pytest
+from unittest.mock import MagicMock, patch
+
+from abogen.tts_plugin.plugin_manager import PluginManager, get_plugin_manager, reset_plugin_manager
+from abogen.tts_plugin.utils import Pipeline, create_pipeline
+from abogen.tts_plugin.engine import Engine, EngineSession
+from abogen.tts_plugin.types import SynthesisRequest, SynthesizedAudio, AudioFormat
+
+
+class FakeEngine:
+    """Fake Engine for testing."""
+    
+    def __init__(self, **kwargs):
+        self.kwargs = kwargs
+        self._disposed = False
+    
+    def createSession(self):
+        return FakeEngineSession()
+    
+    def dispose(self):
+        self._disposed = True
+    
+    @property
+    def manifest(self):
+        return MagicMock()
+
+
+class FakeEngineSession:
+    """Fake EngineSession for testing."""
+    
+    def __init__(self):
+        self._disposed = False
+    
+    def synthesize(self, request):
+        # Return fake audio
+        import numpy as np
+        from abogen.tts_plugin.types import AudioFormat, Duration, SynthesizedAudio
+        
+        audio = np.zeros(1000, dtype=np.float32)
+        return SynthesizedAudio(
+            data=audio.tobytes(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+            duration=Duration(seconds=0.04167),  # 1000 samples / 24000 Hz
+        )
+    
+    def dispose(self):
+        self._disposed = True
+
+
+class TestPluginManager:
+    """Test PluginManager functionality."""
+    
+    def test_plugin_manager_creation(self):
+        """PluginManager can be created."""
+        manager = PluginManager()
+        assert manager is not None
+    
+    def test_plugin_manager_list_discovers_plugins(self):
+        """PluginManager discovers plugins from plugins directory."""
+        manager = PluginManager()
+        manager.discover("plugins")
+        plugins = manager.list_plugins()
+        # Should discover kokoro plugin if it exists
+        assert isinstance(plugins, list)
+    
+    def test_plugin_manager_has_plugin_after_discover(self):
+        """PluginManager reports plugins after discovery."""
+        manager = PluginManager()
+        manager.discover("plugins")
+        # kokoro plugin should be discovered if plugins/kokoro exists
+        # This is expected behavior
+        assert isinstance(manager._plugins, dict)
+    
+    def test_plugin_manager_get_plugin_after_discover(self):
+        """PluginManager returns plugin info after discovery."""
+        manager = PluginManager()
+        manager.discover("plugins")
+        # kokoro plugin should be discovered if plugins/kokoro exists
+        assert isinstance(manager._plugins, dict)
+    
+    def test_plugin_manager_create_engine_not_found(self):
+        """PluginManager raises KeyError for unknown plugins."""
+        manager = PluginManager()
+        with pytest.raises(KeyError, match="Plugin not found"):
+            manager.create_engine("nonexistent")
+    
+    def test_plugin_manager_discover_with_empty_dir(self):
+        """PluginManager handles missing plugins directory."""
+        manager = PluginManager()
+        manager.discover("/nonexistent/path")
+        plugins = manager.list_plugins()
+        assert plugins == []
+    
+    def test_global_plugin_manager_singleton(self):
+        """Global PluginManager is a singleton."""
+        reset_plugin_manager()
+        manager1 = get_plugin_manager()
+        manager2 = get_plugin_manager()
+        assert manager1 is manager2
+        reset_plugin_manager()
+    
+    def test_reset_plugin_manager(self):
+        """reset_plugin_manager clears the singleton."""
+        manager1 = get_plugin_manager()
+        reset_plugin_manager()
+        manager2 = get_plugin_manager()
+        assert manager1 is not manager2
+        reset_plugin_manager()
+
+
+class TestPipeline:
+    """Test Pipeline functionality."""
+    
+    def test_pipeline_creation(self):
+        """Pipeline can be created."""
+        engine = FakeEngine()
+        backend = Pipeline(engine)
+        assert backend is not None
+    
+    def test_pipeline_callable(self):
+        """Pipeline is callable like old TTSBackend."""
+        engine = FakeEngine()
+        backend = Pipeline(engine)
+        
+        # Should be callable
+        assert callable(backend)
+    
+    def test_pipeline_synthesize(self):
+        """Pipeline can synthesize text."""
+        engine = FakeEngine()
+        backend = Pipeline(engine)
+        
+        # Call the backend
+        segments = list(backend("Hello world", voice="default", speed=1.0))
+        
+        # Should return at least one segment
+        assert len(segments) >= 1
+        
+        # Segment should have graphemes and audio
+        segment = segments[0]
+        assert hasattr(segment, "graphemes")
+        assert hasattr(segment, "audio")
+        assert segment.graphemes == "Hello world"
+    
+    def test_pipeline_dispose(self):
+        """Pipeline can be disposed."""
+        engine = FakeEngine()
+        backend = Pipeline(engine)
+        
+        # Create a session by calling
+        list(backend("test"))
+        
+        # Dispose should not raise
+        backend.dispose()
+        
+        # Double dispose should be safe
+        backend.dispose()
+
+
+class TestCreatePipelineCompat:
+    """Test create_pipeline utility function."""
+    
+    def test_create_pipeline_returns_callable(self):
+        """create_pipeline returns a callable backend."""
+        from abogen.tts_plugin.host_context import HostContext
+        from abogen.tts_plugin.types import EngineConfig
+
+        # Mock the plugin manager
+        with patch("abogen.tts_plugin.utils.get_plugin_manager") as mock_get_manager:
+            mock_manager = MagicMock()
+            mock_get_manager.return_value = mock_manager
+            
+            mock_engine = FakeEngine()
+            mock_manager.create_engine.return_value = mock_engine
+            
+            backend = create_pipeline("kokoro", lang_code="a", device="cpu")
+            
+            assert callable(backend)
+            mock_manager.create_engine.assert_called_once()
+            call_args = mock_manager.create_engine.call_args
+            assert call_args.args[0] == "kokoro"
+            assert isinstance(call_args.kwargs["context"], HostContext)
+            assert call_args.kwargs["model_path"] is None
+            assert isinstance(call_args.kwargs["config"], EngineConfig)
+            assert call_args.kwargs["config"].device == "cpu"
+            assert call_args.kwargs["config"].lang_code == "a"
+    
+    def test_create_pipeline_raises_for_unknown_plugin(self):
+        """create_pipeline raises KeyError for unknown plugins."""
+        with patch("abogen.tts_plugin.utils.get_plugin_manager") as mock_get_manager:
+            mock_manager = MagicMock()
+            mock_get_manager.return_value = mock_manager
+            mock_manager.create_engine.side_effect = KeyError("Plugin not found")
+            
+            with pytest.raises(KeyError):
+                create_pipeline("nonexistent")
+
+
+class TestPluginManagerWithFakePlugins:
+    """Test PluginManager with fake plugin loading."""
+    
+    def test_plugin_manager_create_engine_from_plugin(self):
+        """PluginManager creates engine from loaded plugin."""
+        manager = PluginManager()
+        
+        # Manually add a fake plugin
+        def fake_create_engine(**kwargs):
+            return FakeEngine(**kwargs)
+        
+        manager._plugins["fake"] = {
+            "manifest": MagicMock(),
+            "create_engine": fake_create_engine,
+        }
+        manager._loaded = True
+        
+        # Create engine
+        engine = manager.create_engine("fake", param="value")
+        
+        assert isinstance(engine, FakeEngine)
+        assert engine.kwargs == {"param": "value"}
+    
+    def test_plugin_manager_get_or_create_engine(self):
+        """PluginManager caches engines."""
+        manager = PluginManager()
+        
+        call_count = 0
+        
+        def fake_create_engine(**kwargs):
+            nonlocal call_count
+            call_count += 1
+            return FakeEngine(**kwargs)
+        
+        manager._plugins["fake"] = {
+            "manifest": MagicMock(),
+            "create_engine": fake_create_engine,
+        }
+        manager._loaded = True
+        
+        # Get engine twice
+        engine1 = manager.get_or_create_engine("fake")
+        engine2 = manager.get_or_create_engine("fake")
+        
+        # Should be same instance
+        assert engine1 is engine2
+        assert call_count == 1
+    
+    def test_plugin_manager_dispose_all(self):
+        """PluginManager disposes all cached engines."""
+        manager = PluginManager()
+        
+        def fake_create_engine(**kwargs):
+            return FakeEngine(**kwargs)
+        
+        manager._plugins["fake"] = {
+            "manifest": MagicMock(),
+            "create_engine": fake_create_engine,
+        }
+        manager._loaded = True
+        
+        # Create and cache engines
+        engine1 = manager.get_or_create_engine("fake")
+        engine2 = manager.get_or_create_engine("fake")
+        
+        # Dispose all
+        manager.dispose_all()
+        
+        # Engines should be disposed
+        assert engine1._disposed is True
+        assert engine2._disposed is True
+        
+        # Cache should be empty
+        assert len(manager._engines) == 0

tests/contracts/test_session_contract.py

@@ -0,0 +1,135 @@
+"""Contract tests for EngineSession protocol.
+
+These tests verify that EngineSession implementations satisfy the architectural requirements:
+- synthesize() returns SynthesizedAudio
+- dispose() is idempotent
+- After dispose(), synthesize() raises EngineError
+- Session remains usable after synthesize() failure
+"""
+
+import pytest
+
+from abogen.tts_plugin.engine import EngineSession
+from abogen.tts_plugin.errors import EngineError
+from abogen.tts_plugin.types import (
+    AudioFormat,
+    Duration,
+    ParameterValues,
+    SynthesisRequest,
+    SynthesizedAudio,
+    VoiceSelection,
+)
+
+from .conftest import FakeEngineSession
+
+
+class TestEngineSessionProtocolContract:
+    """Contract tests for the EngineSession protocol itself."""
+
+    def test_engine_session_is_protocol(self) -> None:
+        assert hasattr(EngineSession, "__protocol_attrs__")
+
+    def test_fake_session_satisfies_protocol(self) -> None:
+        session = FakeEngineSession()
+        assert isinstance(session, EngineSession)
+
+
+class TestSessionSynthesizeContract:
+    """Contract tests for EngineSession.synthesize()."""
+
+    def test_synthesize_returns_synthesized_audio(self) -> None:
+        session = FakeEngineSession()
+        request = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="af_nova"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        result = session.synthesize(request)
+        assert isinstance(result, SynthesizedAudio)
+
+    def test_synthesize_returns_valid_audio_data(self) -> None:
+        session = FakeEngineSession()
+        request = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="af_nova"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        result = session.synthesize(request)
+        assert isinstance(result.data, bytes)
+        assert len(result.data) > 0
+        assert isinstance(result.format, AudioFormat)
+        assert isinstance(result.duration, Duration)
+
+
+class TestSessionDisposeContract:
+    """Contract tests for EngineSession.dispose()."""
+
+    def test_dispose_is_idempotent(self) -> None:
+        """Architecture spec: dispose() is idempotent."""
+        session = FakeEngineSession()
+        session.dispose()
+        session.dispose()  # Should not raise
+
+    def test_dispose_never_raises(self) -> None:
+        """Architecture spec: dispose() never raises exceptions."""
+        session = FakeEngineSession()
+        session.dispose()  # Should not raise
+
+
+class TestSessionAfterDisposeContract:
+    """Contract tests for behavior after dispose()."""
+
+    def test_synthesize_after_dispose_raises(self) -> None:
+        """Architecture spec: After dispose(), all methods except dispose() raise EngineError."""
+        session = FakeEngineSession()
+        session.dispose()
+        request = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="af_nova"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        with pytest.raises(EngineError):
+            session.synthesize(request)
+
+
+class TestSessionLifecycleContract:
+    """Contract tests for EngineSession lifecycle."""
+
+    def test_full_lifecycle(self) -> None:
+        """Test complete session lifecycle: create -> synthesize -> dispose."""
+        session = FakeEngineSession()
+        request = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="af_nova"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+
+        # Synthesize
+        result = session.synthesize(request)
+        assert isinstance(result, SynthesizedAudio)
+
+        # Dispose
+        session.dispose()
+
+    def test_multiple_synthesize_before_dispose(self) -> None:
+        """Architecture spec: Session remains usable after synthesize() failure."""
+        session = FakeEngineSession()
+        request = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="af_nova"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+
+        # Multiple synthesize calls
+        result1 = session.synthesize(request)
+        result2 = session.synthesize(request)
+        assert isinstance(result1, SynthesizedAudio)
+        assert isinstance(result2, SynthesizedAudio)
+
+        # Dispose
+        session.dispose()

tests/contracts/test_types_contract.py

@@ -0,0 +1,244 @@
+"""Contract tests for core domain value objects.
+
+These tests verify that value objects satisfy the architectural requirements:
+- Frozen (immutable) dataclasses
+- Correct field definitions
+- Proper equality behavior
+"""
+
+import pytest
+
+from abogen.tts_plugin.types import (
+    AudioFormat,
+    Duration,
+    EngineConfig,
+    ParameterValues,
+    SynthesisRequest,
+    SynthesizedAudio,
+    VoiceSelection,
+)
+
+
+class TestAudioFormatContract:
+    """Contract tests for AudioFormat value object."""
+
+    def test_is_frozen_dataclass(self) -> None:
+        assert hasattr(AudioFormat, "__dataclass_params__")
+        assert AudioFormat.__dataclass_params__.frozen is True
+
+    def test_required_fields(self) -> None:
+        af = AudioFormat(mime="audio/wav", extension="wav")
+        assert af.mime == "audio/wav"
+        assert af.extension == "wav"
+
+    def test_immutability(self) -> None:
+        af = AudioFormat(mime="audio/wav", extension="wav")
+        with pytest.raises(AttributeError):
+            af.mime = "audio/mpeg"  # type: ignore[misc]
+
+    def test_equality(self) -> None:
+        af1 = AudioFormat(mime="audio/wav", extension="wav")
+        af2 = AudioFormat(mime="audio/wav", extension="wav")
+        assert af1 == af2
+
+    def test_inequality(self) -> None:
+        af1 = AudioFormat(mime="audio/wav", extension="wav")
+        af2 = AudioFormat(mime="audio/mpeg", extension="mp3")
+        assert af1 != af2
+
+    def test_hashable(self) -> None:
+        af = AudioFormat(mime="audio/wav", extension="wav")
+        assert hash(af) == hash(AudioFormat(mime="audio/wav", extension="wav"))
+
+
+class TestDurationContract:
+    """Contract tests for Duration value object."""
+
+    def test_is_frozen_dataclass(self) -> None:
+        assert hasattr(Duration, "__dataclass_params__")
+        assert Duration.__dataclass_params__.frozen is True
+
+    def test_required_fields(self) -> None:
+        d = Duration(seconds=1.5)
+        assert d.seconds == 1.5
+
+    def test_immutability(self) -> None:
+        d = Duration(seconds=1.0)
+        with pytest.raises(AttributeError):
+            d.seconds = 2.0  # type: ignore[misc]
+
+    def test_equality(self) -> None:
+        d1 = Duration(seconds=1.0)
+        d2 = Duration(seconds=1.0)
+        assert d1 == d2
+
+
+class TestVoiceSelectionContract:
+    """Contract tests for VoiceSelection value object."""
+
+    def test_is_frozen_dataclass(self) -> None:
+        assert hasattr(VoiceSelection, "__dataclass_params__")
+        assert VoiceSelection.__dataclass_params__.frozen is True
+
+    def test_required_fields(self) -> None:
+        vs = VoiceSelection(source="builtin", key="af_nova")
+        assert vs.source == "builtin"
+        assert vs.key == "af_nova"
+
+    def test_payload_default_none(self) -> None:
+        vs = VoiceSelection(source="builtin", key="af_nova")
+        assert vs.payload is None
+
+    def test_payload_optional(self) -> None:
+        vs = VoiceSelection(source="clone", key="my_voice", payload=b"audio_data")
+        assert vs.payload == b"audio_data"
+
+    def test_immutability(self) -> None:
+        vs = VoiceSelection(source="builtin", key="af_nova")
+        with pytest.raises(AttributeError):
+            vs.source = "other"  # type: ignore[misc]
+
+
+class TestParameterValuesContract:
+    """Contract tests for ParameterValues value object."""
+
+    def test_is_frozen_dataclass(self) -> None:
+        assert hasattr(ParameterValues, "__dataclass_params__")
+        assert ParameterValues.__dataclass_params__.frozen is True
+
+    def test_default_empty(self) -> None:
+        pv = ParameterValues()
+        assert pv.values == {}
+
+    def test_with_values(self) -> None:
+        pv = ParameterValues(values={"speed": 1.0, "pitch": 0.5})
+        assert pv.values["speed"] == 1.0
+        assert pv.values["pitch"] == 0.5
+
+    def test_immutability(self) -> None:
+        pv = ParameterValues(values={"speed": 1.0})
+        with pytest.raises(AttributeError):
+            pv.values = {}  # type: ignore[misc]
+
+
+class TestSynthesisRequestContract:
+    """Contract tests for SynthesisRequest value object."""
+
+    def test_is_frozen_dataclass(self) -> None:
+        assert hasattr(SynthesisRequest, "__dataclass_params__")
+        assert SynthesisRequest.__dataclass_params__.frozen is True
+
+    def test_required_fields(self) -> None:
+        req = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="af_nova"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        assert req.text == "Hello"
+        assert req.voice.source == "builtin"
+        assert req.format.mime == "audio/wav"
+
+    def test_immutability(self) -> None:
+        req = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="af_nova"),
+            parameters=ParameterValues(),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        with pytest.raises(AttributeError):
+            req.text = "World"  # type: ignore[misc]
+
+
+class TestSynthesizedAudioContract:
+    """Contract tests for SynthesizedAudio value object."""
+
+    def test_is_frozen_dataclass(self) -> None:
+        assert hasattr(SynthesizedAudio, "__dataclass_params__")
+        assert SynthesizedAudio.__dataclass_params__.frozen is True
+
+    def test_required_fields(self) -> None:
+        audio = SynthesizedAudio(
+            data=b"\x00" * 100,
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+            duration=Duration(seconds=1.0),
+        )
+        assert audio.data == b"\x00" * 100
+        assert audio.format.mime == "audio/wav"
+        assert audio.duration.seconds == 1.0
+
+    def test_immutability(self) -> None:
+        audio = SynthesizedAudio(
+            data=b"\x00" * 100,
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+            duration=Duration(seconds=1.0),
+        )
+        with pytest.raises(AttributeError):
+            audio.data = b"\x00"  # type: ignore[misc]
+
+
+class TestEngineConfigContract:
+    """Contract tests for EngineConfig value object."""
+
+    def test_is_frozen_dataclass(self) -> None:
+        assert hasattr(EngineConfig, "__dataclass_params__")
+        assert EngineConfig.__dataclass_params__.frozen is True
+
+    def test_default_device(self) -> None:
+        config = EngineConfig()
+        assert config.device == "cpu"
+
+    def test_custom_device(self) -> None:
+        config = EngineConfig(device="cuda:0")
+        assert config.device == "cuda:0"
+
+    def test_default_lang_code(self) -> None:
+        config = EngineConfig()
+        assert config.lang_code == "a"
+
+    def test_custom_lang_code(self) -> None:
+        config = EngineConfig(lang_code="j")
+        assert config.lang_code == "j"
+
+    def test_immutability(self) -> None:
+        config = EngineConfig()
+        with pytest.raises(AttributeError):
+            config.device = "cuda:0"  # type: ignore[misc]
+
+    def test_immutability_lang_code(self) -> None:
+        config = EngineConfig()
+        with pytest.raises(AttributeError):
+            config.lang_code = "j"  # type: ignore[misc]
+
+    def test_unknown_keys_ignored_per_spec(self) -> None:
+        """Architecture spec: Unknown keys are ignored (no error).
+
+        EngineConfig is frozen, so unknown keys cannot be set after creation.
+        This test verifies the default behavior matches the spec.
+        """
+        config = EngineConfig()
+        assert config.device == "cpu"
+
+    def test_plugins_may_ignore_irrelevant_fields(self) -> None:
+        """Architecture Amendment #1: Plugins ignore unsupported fields.
+
+        EngineConfig may contain fields that are not relevant to every plugin.
+        Plugins MUST ignore fields they do not need, not raise on them.
+        """
+        config = EngineConfig(device="cuda:0", lang_code="j")
+        assert config.device == "cuda:0"
+        assert config.lang_code == "j"
+        # A plugin that only needs device simply reads config.device
+        # and ignores config.lang_code — this must not raise.
+
+    def test_engine_config_contains_engine_instance_configuration(self) -> None:
+        """Architecture Amendment #1: EngineConfig definition.
+
+        EngineConfig contains parameters that define how a particular
+        Engine instance is created and that remain constant throughout
+        the lifetime of that Engine.
+        """
+        config = EngineConfig(device="cpu", lang_code="a")
+        # Both fields are init-time, immutable, engine-scoped.
+        assert config.device == "cpu"
+        assert config.lang_code == "a"

tests/plugins/__init__.py

@@ -0,0 +1,11 @@
+"""Auto-discovery tests for TTS plugins.
+
+This package contains generic tests that automatically run for every plugin
+in the plugins/ directory. Tests verify:
+- Manifest structure
+- Engine creation and dispose contract
+- Capability implementation
+
+Plugin-specific tests remain in tests/test_<plugin>_plugin.py for
+integration with real dependencies (e.g., KPipeline for Kokoro).
+"""

tests/plugins/conftest.py

@@ -0,0 +1,169 @@
+"""Fixtures for auto-discovery plugin tests.
+
+This module provides shared fixtures for testing all plugins:
+- plugin_ids: List of all plugin IDs from plugins/ directory
+- plugin_dir: Path to a specific plugin directory (parametrized)
+- loaded_plugin: Loaded plugin data (manifest, model_requirements, create_engine)
+- host_context: Test HostContext with fake HTTP client
+"""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Any
+
+import pytest
+
+from abogen.tts_plugin.host_context import HostContext
+from abogen.tts_plugin.loader import load_plugin_from_dir
+
+
+def _discover_plugin_ids() -> list[str]:
+    """Discover all plugin IDs from the plugins/ directory.
+    
+    Returns:
+        List of plugin IDs (directory names with __init__.py).
+    """
+    plugins_dir = Path(__file__).parent.parent.parent / "plugins"
+    plugin_ids = []
+    if plugins_dir.exists():
+        for item in sorted(plugins_dir.iterdir()):
+            if item.is_dir() and not item.name.startswith("."):
+                init_file = item / "__init__.py"
+                if init_file.exists():
+                    plugin_ids.append(item.name)
+    return plugin_ids
+
+
+@pytest.fixture(scope="module")
+def plugins_dir() -> Path:
+    """Return the path to the plugins directory."""
+    return Path(__file__).parent.parent.parent / "plugins"
+
+
+@pytest.fixture(scope="module")
+def plugin_ids(plugins_dir: Path) -> list[str]:
+    """Return a list of all plugin IDs from the plugins/ directory.
+    
+    This fixture discovers all subdirectories in plugins/ that contain
+    an __init__.py file (i.e., valid plugin directories).
+    
+    Returns:
+        List of plugin IDs (directory names).
+    """
+    return _discover_plugin_ids()
+
+
+@pytest.fixture(params=_discover_plugin_ids())
+def plugin_id(request: pytest.FixtureRequest) -> str:
+    """Parametrized fixture returning each plugin ID.
+    
+    This fixture is parametrized to run tests for each plugin.
+    """
+    return request.param
+
+
+@pytest.fixture
+def plugin_dir(plugins_dir: Path, plugin_id: str) -> Path:
+    """Return the path to a specific plugin directory.
+    
+    Args:
+        plugins_dir: Base plugins directory.
+        plugin_id: The plugin ID (directory name).
+        
+    Returns:
+        Path to the plugin directory.
+    """
+    return plugins_dir / plugin_id
+
+
+@pytest.fixture
+def loaded_plugin(plugin_dir: Path):
+    """Load a plugin and return its load result.
+    
+    This fixture loads the plugin using the loader, providing access to:
+    - manifest: PluginManifest
+    - model_requirements: tuple[ModelManifest, ...]
+    - create_engine: Callable
+    - module: The loaded module
+    
+    Args:
+        plugin_dir: Path to the plugin directory.
+        
+    Returns:
+        PluginLoadResult from load_plugin_from_dir.
+        
+    Raises:
+        pytest.skip: If plugin fails to load (should not happen for valid plugins).
+    """
+    from abogen.tts_plugin.loader import PluginLoadResult
+    
+    result = load_plugin_from_dir(plugin_dir)
+    if not result.success:
+        pytest.fail(
+            f"Plugin {plugin_dir.name} failed to load: "
+            f"{result.error.errors if result.error else 'Unknown error'}"
+        )
+    return result
+
+
+@pytest.fixture
+def host_context(tmp_path: Path) -> HostContext:
+    """Create a test HostContext with fake HTTP client.
+    
+    Args:
+        tmp_path: Pytest tmp_path fixture for config directory.
+        
+    Returns:
+        HostContext with fake HTTP client and test logger.
+    """
+    class FakeHttpClient:
+        """Fake HTTP client for testing."""
+        
+        def get(self, url: str, **kwargs: object) -> object:
+            """Fake GET request."""
+            return None
+        
+        def post(self, url: str, **kwargs: object) -> object:
+            """Fake POST request."""
+            return None
+    
+    return HostContext(
+        config_dir=tmp_path,
+        logger=logging.getLogger("test"),
+        http_client=FakeHttpClient(),
+    )
+
+
+@pytest.fixture
+def engine_config() -> Any:
+    """Return a default EngineConfig for testing.
+    
+    Returns:
+        EngineConfig with device="cpu" for testing.
+    """
+    from abogen.tts_plugin.types import EngineConfig
+    return EngineConfig(device="cpu")
+
+
+@pytest.fixture
+def create_engine(loaded_plugin, host_context, engine_config):
+    """Create an engine instance from a loaded plugin.
+    
+    This fixture creates an engine and ensures proper cleanup via dispose.
+    
+    Args:
+        loaded_plugin: Loaded plugin result.
+        host_context: Test HostContext.
+        engine_config: EngineConfig for engine creation.
+        
+    Yields:
+        Engine instance.
+    """
+    from abogen.tts_plugin.engine import Engine
+    
+    engine = loaded_plugin.create_engine(host_context, None, engine_config)
+    assert isinstance(engine, Engine)
+    yield engine
+    engine.dispose()

tests/plugins/fake_plugin/__init__.py

@@ -0,0 +1,92 @@
+"""Fake plugin for testing the plugin loader.
+
+This is a minimal valid plugin that satisfies the Plugin API contract.
+It does NOT perform any real TTS synthesis.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+from abogen.tts_plugin.engine import Engine, EngineSession
+from abogen.tts_plugin.errors import EngineError
+from abogen.tts_plugin.host_context import HostContext
+from abogen.tts_plugin.manifest import (
+    AudioFormatManifest,
+    EngineManifest,
+    PluginManifest,
+    VoiceSourceManifest,
+)
+from abogen.tts_plugin.types import (
+    AudioFormat,
+    Duration,
+    EngineConfig,
+    ParameterValues,
+    SynthesisRequest,
+    SynthesizedAudio,
+)
+
+
+class FakeSession:
+    """Minimal EngineSession implementation for testing."""
+
+    def __init__(self) -> None:
+        self._disposed = False
+
+    def synthesize(self, request: SynthesisRequest) -> SynthesizedAudio:
+        if self._disposed:
+            raise EngineError("Session disposed")
+        return SynthesizedAudio(
+            data=b"\x00" * 100,
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+            duration=Duration(seconds=1.0),
+        )
+
+    def dispose(self) -> None:
+        self._disposed = True
+
+
+class FakeEngine:
+    """Minimal Engine implementation for testing."""
+
+    def __init__(self) -> None:
+        self._disposed = False
+
+    def createSession(self) -> EngineSession:
+        if self._disposed:
+            raise EngineError("Engine disposed")
+        return FakeSession()
+
+    def dispose(self) -> None:
+        self._disposed = True
+
+
+PLUGIN_MANIFEST = PluginManifest(
+    id="fake_plugin",
+    name="Fake Plugin",
+    version="1.0.0",
+    api_version="1.0",
+    description="A fake plugin for testing",
+    author="Test Author",
+    capabilities=(),
+    engine=EngineManifest(
+        voiceSources=(
+            VoiceSourceManifest(id="builtin", name="Builtin", type="list"),
+        ),
+        audioFormats=(
+            AudioFormatManifest(mime="audio/wav", extension="wav"),
+        ),
+    ),
+)
+
+MODEL_REQUIREMENTS: list[Any] = []
+
+
+def create_engine(
+    context: HostContext,
+    model_path: Path | None,
+    config: EngineConfig,
+) -> Engine:
+    """Create a fake engine instance."""
+    return FakeEngine()

tests/plugins/import_error/__init__.py

@@ -0,0 +1,18 @@
+"""Invalid plugin: raises ImportError during import."""
+
+from __future__ import annotations
+
+# This plugin intentionally raises an ImportError
+raise ImportError("Simulated import error for testing")
+
+# The following code will never be reached, but is here for documentation
+from abogen.tts_plugin.manifest import PluginManifest
+
+PLUGIN_MANIFEST = PluginManifest(
+    id="import_error",
+    name="Import Error Plugin",
+    version="1.0.0",
+    api_version="1.0",
+    description="Plugin that fails to import",
+    author="Test Author",
+)

tests/plugins/invalid_api_version/__init__.py

@@ -0,0 +1,29 @@
+"""Invalid plugin: incompatible api_version (major version mismatch)."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+from abogen.tts_plugin.manifest import PluginManifest
+from abogen.tts_plugin.types import EngineConfig
+
+# api_version "2.0" has major version 2, but host expects major version 1
+PLUGIN_MANIFEST = PluginManifest(
+    id="invalid_api_version",
+    name="Invalid API Version Plugin",
+    version="1.0.0",
+    api_version="2.0",  # Major version mismatch!
+    description="Plugin with incompatible api_version",
+    author="Test Author",
+)
+
+MODEL_REQUIREMENTS: list[Any] = []
+
+
+def create_engine(
+    context: Any,
+    model_path: Path | None,
+    config: EngineConfig,
+) -> Any:
+    raise NotImplementedError("This plugin is invalid")

tests/plugins/invalid_capabilities/__init__.py

@@ -0,0 +1,29 @@
+"""Invalid plugin: unknown capabilities."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+from abogen.tts_plugin.manifest import PluginManifest
+from abogen.tts_plugin.types import EngineConfig
+
+PLUGIN_MANIFEST = PluginManifest(
+    id="invalid_capabilities",
+    name="Invalid Capabilities Plugin",
+    version="1.0.0",
+    api_version="1.0",
+    description="Plugin with unknown capabilities",
+    author="Test Author",
+    capabilities=("voice_list", "unknown_capability", "another_unknown"),
+)
+
+MODEL_REQUIREMENTS: list[Any] = []
+
+
+def create_engine(
+    context: Any,
+    model_path: Path | None,
+    config: EngineConfig,
+) -> Any:
+    raise NotImplementedError("This plugin is invalid")

tests/plugins/missing_create_engine/__init__.py

@@ -0,0 +1,18 @@
+"""Invalid plugin: missing create_engine function."""
+
+from __future__ import annotations
+
+from abogen.tts_plugin.manifest import PluginManifest
+
+PLUGIN_MANIFEST = PluginManifest(
+    id="missing_create_engine",
+    name="Missing Create Engine Plugin",
+    version="1.0.0",
+    api_version="1.0",
+    description="Plugin missing create_engine",
+    author="Test Author",
+)
+
+MODEL_REQUIREMENTS: list = []
+
+# This plugin intentionally does NOT export create_engine

tests/plugins/missing_manifest/__init__.py

@@ -0,0 +1,10 @@
+"""Invalid plugin: missing PLUGIN_MANIFEST."""
+
+from __future__ import annotations
+
+# This plugin intentionally does NOT export PLUGIN_MANIFEST
+MODEL_REQUIREMENTS: list = []
+
+
+def create_engine(context, model_path, config):
+    raise NotImplementedError("This plugin is invalid")

tests/plugins/missing_model_requirements/__init__.py

@@ -0,0 +1,28 @@
+"""Invalid plugin: missing MODEL_REQUIREMENTS."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+from abogen.tts_plugin.manifest import PluginManifest
+from abogen.tts_plugin.types import EngineConfig
+
+PLUGIN_MANIFEST = PluginManifest(
+    id="missing_model_requirements",
+    name="Missing Model Requirements Plugin",
+    version="1.0.0",
+    api_version="1.0",
+    description="Plugin missing MODEL_REQUIREMENTS",
+    author="Test Author",
+)
+
+# This plugin intentionally does NOT export MODEL_REQUIREMENTS
+
+
+def create_engine(
+    context: Any,
+    model_path: Path | None,
+    config: EngineConfig,
+) -> Any:
+    raise NotImplementedError("This plugin is invalid")

tests/plugins/test_all_plugins.py

@@ -0,0 +1,235 @@
+"""Auto-discovery tests for all TTS plugins.
+
+These tests automatically run for every plugin in the plugins/ directory.
+Tests are grouped into three categories:
+
+1. TestAllPluginsManifest - Validates manifest structure
+   - Required fields (id, name, version, api_version, etc.)
+   - API version format (semver MAJOR.MINOR)
+   - Voices field (optional)
+
+2. TestAllPluginsEngine - Validates engine lifecycle
+   - create_engine returns valid Engine
+   - dispose is idempotent
+   - dispose → createSession raises EngineError
+
+3. TestAllPluginsCapabilities - Validates capability implementation
+   - Declared capabilities are implemented
+   - voice_list → VoiceLister interface
+"""
+
+from __future__ import annotations
+
+import re
+
+import pytest
+
+from abogen.tts_plugin.engine import Engine, EngineSession
+from abogen.tts_plugin.errors import EngineError
+from abogen.tts_plugin.manifest import PluginManifest
+from abogen.tts_plugin.types import EngineConfig, VoiceSelection, AudioFormat, ParameterValues, SynthesisRequest
+
+
+class TestAllPluginsManifest:
+    """Test that all plugins have valid manifest structure."""
+
+    def test_manifest_has_required_fields(self, loaded_plugin):
+        """Verify manifest has all required fields."""
+        manifest = loaded_plugin.manifest
+        assert manifest is not None
+        
+        # Required fields
+        assert manifest.id, "Plugin id must not be empty"
+        assert manifest.name, "Plugin name must not be empty"
+        assert manifest.version, "Plugin version must not be empty"
+        assert manifest.api_version, "Plugin api_version must not be empty"
+        assert manifest.description, "Plugin description must not be empty"
+        assert manifest.author, "Plugin author must not be empty"
+
+    def test_manifest_id_matches_directory(self, loaded_plugin, plugin_dir):
+        """Verify manifest id matches the plugin directory name."""
+        manifest = loaded_plugin.manifest
+        assert manifest.id == plugin_dir.name, \
+            f"Manifest id '{manifest.id}' must match directory name '{plugin_dir.name}'"
+
+    def test_api_version_format(self, loaded_plugin):
+        """Verify api_version follows semver format (MAJOR.MINOR)."""
+        manifest = loaded_plugin.manifest
+        api_version = manifest.api_version
+        
+        # Must match MAJOR.MINOR format
+        pattern = r"^\d+\.\d+$"
+        assert re.match(pattern, api_version), \
+            f"api_version '{api_version}' must be in format MAJOR.MINOR (e.g., 1.0)"
+
+    def test_api_version_compatibility(self, loaded_plugin):
+        """Verify api_version major version matches host API version."""
+        from abogen.tts_plugin.loader import HOST_API_VERSION
+        
+        manifest = loaded_plugin.manifest
+        plugin_ver = manifest.api_version
+        host_ver = HOST_API_VERSION
+        
+        # Extract major versions
+        plugin_major = int(plugin_ver.split(".")[0])
+        host_major = int(host_ver.split(".")[0])
+        
+        assert plugin_major == host_major, \
+            f"API version major mismatch: plugin={plugin_major}, host={host_major}"
+
+    def test_voices_field_is_optional(self, loaded_plugin):
+        """Verify voices field is optional (can be None or tuple)."""
+        manifest = loaded_plugin.manifest
+        voices = manifest.voices
+        
+        # voices can be None (not declared) or tuple (empty or with voices)
+        assert voices is None or isinstance(voices, tuple), \
+            f"voices must be None or tuple, got {type(voices).__name__}"
+
+    def test_capabilities_is_tuple(self, loaded_plugin):
+        """Verify capabilities is a tuple."""
+        manifest = loaded_plugin.manifest
+        assert isinstance(manifest.capabilities, tuple), \
+            f"capabilities must be a tuple, got {type(manifest.capabilities).__name__}"
+
+    def test_engine_manifest_exists(self, loaded_plugin):
+        """Verify engine manifest exists and has required fields."""
+        manifest = loaded_plugin.manifest
+        engine_manifest = manifest.engine
+        
+        assert engine_manifest is not None
+        assert isinstance(engine_manifest.voiceSources, tuple)
+        assert isinstance(engine_manifest.parameters, tuple)
+        assert isinstance(engine_manifest.audioFormats, tuple)
+
+
+class TestAllPluginsEngine:
+    """Test that all plugins satisfy the Engine contract."""
+
+    def test_create_engine_returns_engine(self, loaded_plugin, host_context, engine_config):
+        """Verify create_engine returns an Engine instance."""
+        engine = loaded_plugin.create_engine(host_context, None, engine_config)
+        assert isinstance(engine, Engine), \
+            f"create_engine must return Engine instance, got {type(engine).__name__}"
+        engine.dispose()
+
+    def test_dispose_is_idempotent(self, loaded_plugin, host_context, engine_config):
+        """Verify dispose can be called multiple times without error."""
+        engine = loaded_plugin.create_engine(host_context, None, engine_config)
+        
+        # First dispose
+        engine.dispose()
+        
+        # Second dispose should not raise
+        engine.dispose()
+
+    def test_create_session_after_dispose_raises(self, loaded_plugin, host_context, engine_config):
+        """Verify createSession raises EngineError after dispose."""
+        engine = loaded_plugin.create_engine(host_context, None, engine_config)
+        engine.dispose()
+        
+        with pytest.raises(EngineError):
+            engine.createSession()
+
+    def test_dispose_after_dispose_raises(self, loaded_plugin, host_context, engine_config):
+        """Verify dispose after dispose does not raise (idempotent)."""
+        engine = loaded_plugin.create_engine(host_context, None, engine_config)
+        engine.dispose()
+        
+        # Should not raise
+        engine.dispose()
+
+    def test_create_session_returns_valid_session(self, loaded_plugin, host_context, engine_config):
+        """Verify createSession returns an EngineSession instance."""
+        engine = loaded_plugin.create_engine(host_context, None, engine_config)
+        session = engine.createSession()
+        
+        assert isinstance(session, EngineSession), \
+            f"createSession must return EngineSession instance, got {type(session).__name__}"
+        session.dispose()
+        engine.dispose()
+
+    def test_session_dispose_is_idempotent(self, loaded_plugin, host_context, engine_config):
+        """Verify session dispose can be called multiple times."""
+        engine = loaded_plugin.create_engine(host_context, None, engine_config)
+        session = engine.createSession()
+        
+        # First dispose
+        session.dispose()
+        
+        # Second dispose should not raise
+        session.dispose()
+        engine.dispose()
+
+    def test_session_synthesize_after_dispose_raises(self, loaded_plugin, host_context, engine_config):
+        """Verify session.synthesize raises EngineError after dispose."""
+        engine = loaded_plugin.create_engine(host_context, None, engine_config)
+        session = engine.createSession()
+        session.dispose()
+        
+        # Create a minimal request (won't actually synthesize, just test error)
+        request = SynthesisRequest(
+            text="test",
+            voice=VoiceSelection(source="test", key="test"),
+            parameters=ParameterValues(values={}),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        
+        with pytest.raises(EngineError):
+            session.synthesize(request)
+        engine.dispose()
+
+
+class TestAllPluginsCapabilities:
+    """Test that all plugins implement their declared capabilities."""
+
+    def test_voice_list_capability(self, loaded_plugin, host_context, engine_config):
+        """Verify plugins with 'voice_list' capability implement VoiceLister."""
+        manifest = loaded_plugin.manifest
+        
+        if "voice_list" not in manifest.capabilities:
+            pytest.skip("Plugin does not declare 'voice_list' capability")
+        
+        engine = loaded_plugin.create_engine(host_context, None, engine_config)
+        
+        # Check if listVoices method exists
+        assert hasattr(engine, "listVoices"), \
+            "Plugin with 'voice_list' capability must have listVoices method"
+        
+        # listVoices should be callable
+        assert callable(engine.listVoices), "listVoices must be callable"
+        
+        # listVoices should return a list/tuple of voices
+        # Get first voice source from manifest
+        if manifest.engine.voiceSources:
+            source_id = manifest.engine.voiceSources[0].id
+            voices = engine.listVoices(source_id)
+            
+            assert isinstance(voices, (list, tuple)), \
+                f"listVoices must return list or tuple, got {type(voices).__name__}"
+            
+            # Each voice should have id, name, tags
+            for voice in voices:
+                assert hasattr(voice, "id"), "Voice must have 'id' attribute"
+                assert hasattr(voice, "name"), "Voice must have 'name' attribute"
+                assert hasattr(voice, "tags"), "Voice must have 'tags' attribute"
+                assert isinstance(voice.tags, tuple), "Voice tags must be a tuple"
+        
+        engine.dispose()
+
+    def test_known_capabilities_are_valid(self, loaded_plugin):
+        """Verify all declared capabilities are known."""
+        manifest = loaded_plugin.manifest
+        
+        known_capabilities = frozenset({
+            "voice_list",
+            "preview",
+            "voice_clone",
+            "voice_blend",
+            "streaming",
+            "cancel",
+        })
+        
+        for cap in manifest.capabilities:
+            assert cap in known_capabilities, \
+                f"Unknown capability: '{cap}'. Known capabilities: {known_capabilities}"

tests/test_conversion_voice_resolution.py

@@ -1,7 +1,7 @@
 from types import SimpleNamespace
 from typing import cast
 
-from abogen.constants import VOICES_INTERNAL
+from abogen.tts_plugin.utils import get_voices
 from abogen.webui.conversion_runner import (
     _chapter_voice_spec,
     _chunk_voice_spec,
@@ -49,4 +49,4 @@ def test_voice_collection_includes_formula_components():
     voices = _collect_required_voice_ids(job)
 
     assert {"af_nova", "am_liam"}.issubset(voices)
-    assert voices.issuperset(VOICES_INTERNAL)
+    assert voices.issuperset(get_voices("kokoro"))

tests/test_epub_exporter.py

@@ -197,7 +197,7 @@ def test_epub3_preserves_original_whitespace(tmp_path) -> None:
     )
     assert match is not None
     original_text = html.unescape(match.group(1))
-    assert "Second line\n\nThird paragraph." in original_text
+    assert "Second line\n\nThird paragraph." in original_text.replace("\r\n", "\n")
 
 
 def test_epub3_sentence_chunks_render_as_paragraphs(tmp_path) -> None:

tests/test_kokoro_plugin.py

@@ -0,0 +1,196 @@
+"""Tests for the Kokoro TTS Plugin.
+
+These tests verify that the Kokoro plugin:
+- Loads correctly through the Plugin Loader
+- Has a valid manifest
+- Creates a valid Engine
+- Satisfies the Engine/EngineSession contract (via EngineContractMixin)
+- Implements VoiceLister capability
+"""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Any
+
+import pytest
+
+from abogen.tts_plugin.engine import Engine, EngineSession
+from abogen.tts_plugin.host_context import HostContext
+from abogen.tts_plugin.loader import load_plugin_from_dir
+from abogen.tts_plugin.manifest import PluginManifest
+from abogen.tts_plugin.types import (
+    AudioFormat,
+    EngineConfig,
+    ParameterValues,
+    SynthesisRequest,
+    VoiceSelection,
+)
+
+from tests.contracts.engine_contract import EngineContractMixin
+
+
+# ──────────────────────────────────────────────────────────────
+# Helpers
+# ──────────────────────────────────────────────────────────────
+
+def _kokoro_available() -> bool:
+    try:
+        from kokoro import KPipeline  # type: ignore[import-not-found]
+        return True
+    except ImportError:
+        return False
+
+
+def _make_mock_engine() -> Any:
+    from plugins.kokoro.engine import KokoroEngine
+
+    class MockPipeline:
+        def __call__(self, text, voice, speed, split_pattern=None):
+            class MockSegment:
+                def __init__(self):
+                    self.audio = MockAudio()
+            class MockAudio:
+                def numpy(self):
+                    import numpy as np
+                    return np.zeros(24000, dtype="float32")
+            return [MockSegment()]
+
+    engine = KokoroEngine(MockPipeline())
+
+    # Override listVoices for testing (real engine reads from manifest)
+    from abogen.tts_plugin.manifest import VoiceManifest
+    engine.listVoices = lambda source_id: [
+        VoiceManifest(id="test_voice_1", name="Test Voice 1", tags=("en",)),
+        VoiceManifest(id="test_voice_2", name="Test Voice 2", tags=("es",)),
+    ]
+    return engine
+
+
+# ──────────────────────────────────────────────────────────────
+# Fixtures
+# ──────────────────────────────────────────────────────────────
+
+@pytest.fixture
+def kokoro_plugin_dir() -> Path:
+    return Path(__file__).parent.parent / "plugins" / "kokoro"
+
+
+@pytest.fixture
+def host_context(tmp_path: Path) -> HostContext:
+    class FakeHttpClient:
+        def get(self, url: str, **kwargs: object) -> object:
+            return None
+        def post(self, url: str, **kwargs: object) -> object:
+            return None
+
+    return HostContext(
+        config_dir=tmp_path,
+        logger=logging.getLogger("test"),
+        http_client=FakeHttpClient(),
+    )
+
+
+@pytest.fixture
+def engine() -> Engine:
+    return _make_mock_engine()
+
+
+# ──────────────────────────────────────────────────────────────
+# Plugin Loading Tests
+# ──────────────────────────────────────────────────────────────
+
+class TestKokoroPluginLoading:
+
+    def test_plugin_loads_successfully(self, kokoro_plugin_dir: Path) -> None:
+        result = load_plugin_from_dir(kokoro_plugin_dir)
+        assert result.success is True
+        assert result.manifest is not None
+        assert result.create_engine is not None
+
+    def test_plugin_has_valid_manifest(self, kokoro_plugin_dir: Path) -> None:
+        result = load_plugin_from_dir(kokoro_plugin_dir)
+        assert result.success is True
+        manifest = result.manifest
+        assert isinstance(manifest, PluginManifest)
+        assert manifest.id == "kokoro"
+        assert manifest.name == "Kokoro"
+        assert manifest.api_version == "1.0"
+
+    def test_plugin_has_model_requirements(self, kokoro_plugin_dir: Path) -> None:
+        result = load_plugin_from_dir(kokoro_plugin_dir)
+        assert result.success is True
+        assert result.model_requirements is not None
+        assert isinstance(result.model_requirements, tuple)
+
+    def test_plugin_manifest_capabilities(self, kokoro_plugin_dir: Path) -> None:
+        result = load_plugin_from_dir(kokoro_plugin_dir)
+        assert result.success is True
+        assert "voice_list" in result.manifest.capabilities
+
+    def test_plugin_manifest_engine(self, kokoro_plugin_dir: Path) -> None:
+        result = load_plugin_from_dir(kokoro_plugin_dir)
+        assert result.success is True
+        engine_manifest = result.manifest.engine
+        assert len(engine_manifest.voiceSources) > 0
+        assert len(engine_manifest.audioFormats) > 0
+        assert len(engine_manifest.parameters) > 0
+
+
+# ──────────────────────────────────────────────────────────────
+# Engine Creation (real backend, skipped if not installed)
+# ──────────────────────────────────────────────────────────────
+
+class TestKokoroEngineCreation:
+
+    @pytest.mark.skipif(not _kokoro_available(), reason="Kokoro not installed")
+    def test_create_engine(self, kokoro_plugin_dir: Path, host_context: HostContext) -> None:
+        result = load_plugin_from_dir(kokoro_plugin_dir)
+        assert result.success is True
+        engine = result.create_engine(host_context, None, EngineConfig())
+        assert isinstance(engine, Engine)
+        engine.dispose()
+
+    @pytest.mark.skipif(not _kokoro_available(), reason="Kokoro not installed")
+    def test_engine_satisfies_protocol(self, kokoro_plugin_dir: Path, host_context: HostContext) -> None:
+        result = load_plugin_from_dir(kokoro_plugin_dir)
+        assert result.success is True
+        engine = result.create_engine(host_context, None, EngineConfig())
+        assert isinstance(engine, Engine)
+        engine.dispose()
+
+
+# ──────────────────────────────────────────────────────────────
+# Engine / Session Contract (inherited from base)
+# ──────────────────────────────────────────────────────────────
+
+class TestKokoroEngineContract(EngineContractMixin):
+    """Every test from EngineContractMixin runs against KokoroEngine."""
+
+    @pytest.fixture
+    def default_voice(self) -> str:
+        return "af_nova"
+
+
+# ──────────────────────────────────────────────────────────────
+# VoiceLister Tests
+# ──────────────────────────────────────────────────────────────
+
+class TestKokoroVoiceLister:
+
+    def test_list_voices(self) -> None:
+        engine = _make_mock_engine()
+        voices = engine.listVoices("builtin")
+        assert len(voices) > 0
+        assert all(hasattr(v, "id") for v in voices)
+        assert all(hasattr(v, "name") for v in voices)
+        engine.dispose()
+
+    def test_voices_have_tags(self) -> None:
+        engine = _make_mock_engine()
+        voices = engine.listVoices("builtin")
+        for voice in voices:
+            assert isinstance(voice.tags, tuple)
+            assert len(voice.tags) > 0
+        engine.dispose()

tests/test_preview_applies_manual_overrides.py

@@ -19,7 +19,7 @@ def test_preview_applies_manual_override_before_normalization(monkeypatch):
 
     # And stub the kokoro pipeline path so generate_preview_audio won't proceed.
     # We'll instead validate by calling the override logic through generate_preview_audio
-    # with provider=supertonic and stub SupertonicPipeline to capture input.
+    # with provider=supertonic and stub create_backend to capture input.
     captured = {}
 
     class DummyPipeline:
@@ -30,11 +30,16 @@ def test_preview_applies_manual_override_before_normalization(monkeypatch):
             captured["text"] = text
             return iter(())
 
-    monkeypatch.setitem(
-        __import__("sys").modules,
-        "abogen.tts_supertonic",
-        type("M", (), {"SupertonicPipeline": DummyPipeline}),
-    )
+    from abogen.tts_plugin import utils
+
+    original_create_pipeline = utils.create_pipeline
+
+    def _mock_create_pipeline(backend_id, **kwargs):
+        if backend_id == "supertonic":
+            return DummyPipeline(**kwargs)
+        return original_create_pipeline(backend_id, **kwargs)
+
+    monkeypatch.setattr(utils, "create_pipeline", _mock_create_pipeline)
 
     try:
         preview.generate_preview_audio(

tests/test_supertonic_plugin.py

@@ -0,0 +1,281 @@
+"""Tests for the SuperTonic TTS Plugin.
+
+These tests verify that the SuperTonic plugin:
+- Loads correctly through the Plugin Loader
+- Has a valid manifest
+- Creates a valid Engine
+- Satisfies the Engine/EngineSession contract (via EngineContractMixin)
+- Implements VoiceLister capability
+"""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Any
+
+import pytest
+
+from abogen.tts_plugin.engine import Engine, EngineSession
+from abogen.tts_plugin.host_context import HostContext
+from abogen.tts_plugin.loader import load_plugin_from_dir
+from abogen.tts_plugin.manifest import PluginManifest
+from abogen.tts_plugin.types import (
+    AudioFormat,
+    EngineConfig,
+    ParameterValues,
+    SynthesisRequest,
+    VoiceSelection,
+)
+
+from tests.contracts.engine_contract import EngineContractMixin
+
+
+# ──────────────────────────────────────────────────────────────
+# Helpers
+# ──────────────────────────────────────────────────────────────
+
+def _supertonic_available() -> bool:
+    try:
+        from supertonic import TTS  # type: ignore[import-not-found]
+        return True
+    except ImportError:
+        return False
+
+
+def _make_mock_engine() -> Any:
+    from plugins.supertonic.engine import SuperTonicEngine
+
+    class MockSegment:
+        def __init__(self):
+            import numpy as np
+            self.audio = np.zeros(24000, dtype="float32")
+
+    class MockPipeline:
+        sample_rate = 24000
+
+        def __call__(self, text, voice, speed, split_pattern=None, total_steps=None):
+            return [MockSegment()]
+
+    engine = SuperTonicEngine(MockPipeline())
+
+    # Override listVoices for testing (real engine reads from manifest)
+    from abogen.tts_plugin.manifest import VoiceManifest
+    original_list_voices = engine.listVoices
+    
+    def mock_list_voices(source_id):
+        if engine._disposed:
+            from abogen.tts_plugin.errors import EngineError
+            raise EngineError("Engine disposed")
+        return [
+            VoiceManifest(id="M1", name="Male 1", tags=("male",)),
+            VoiceManifest(id="M2", name="Male 2", tags=("male",)),
+            VoiceManifest(id="M3", name="Male 3", tags=("male",)),
+            VoiceManifest(id="M4", name="Male 4", tags=("male",)),
+            VoiceManifest(id="M5", name="Male 5", tags=("male",)),
+            VoiceManifest(id="F1", name="Female 1", tags=("female",)),
+            VoiceManifest(id="F2", name="Female 2", tags=("female",)),
+            VoiceManifest(id="F3", name="Female 3", tags=("female",)),
+            VoiceManifest(id="F4", name="Female 4", tags=("female",)),
+            VoiceManifest(id="F5", name="Female 5", tags=("female",)),
+        ]
+    
+    engine.listVoices = mock_list_voices
+    return engine
+
+
+# ──────────────────────────────────────────────────────────────
+# Fixtures
+# ──────────────────────────────────────────────────────────────
+
+@pytest.fixture
+def supertonic_plugin_dir() -> Path:
+    return Path(__file__).parent.parent / "plugins" / "supertonic"
+
+
+@pytest.fixture
+def host_context(tmp_path: Path) -> HostContext:
+    class FakeHttpClient:
+        def get(self, url: str, **kwargs: object) -> object:
+            return None
+        def post(self, url: str, **kwargs: object) -> object:
+            return None
+
+    return HostContext(
+        config_dir=tmp_path,
+        logger=logging.getLogger("test"),
+        http_client=FakeHttpClient(),
+    )
+
+
+@pytest.fixture
+def engine() -> Engine:
+    return _make_mock_engine()
+
+
+# ──────────────────────────────────────────────────────────────
+# Plugin Loading Tests
+# ──────────────────────────────────────────────────────────────
+
+class TestSuperTonicPluginLoading:
+
+    def test_plugin_loads_successfully(self, supertonic_plugin_dir: Path) -> None:
+        result = load_plugin_from_dir(supertonic_plugin_dir)
+        assert result.success is True
+        assert result.manifest is not None
+        assert result.create_engine is not None
+
+    def test_plugin_has_valid_manifest(self, supertonic_plugin_dir: Path) -> None:
+        result = load_plugin_from_dir(supertonic_plugin_dir)
+        assert result.success is True
+        manifest = result.manifest
+        assert isinstance(manifest, PluginManifest)
+        assert manifest.id == "supertonic"
+        assert manifest.name == "SuperTonic"
+        assert manifest.api_version == "1.0"
+
+    def test_plugin_has_model_requirements(self, supertonic_plugin_dir: Path) -> None:
+        result = load_plugin_from_dir(supertonic_plugin_dir)
+        assert result.success is True
+        assert result.model_requirements is not None
+        assert isinstance(result.model_requirements, tuple)
+
+    def test_plugin_manifest_capabilities(self, supertonic_plugin_dir: Path) -> None:
+        result = load_plugin_from_dir(supertonic_plugin_dir)
+        assert result.success is True
+        assert "voice_list" in result.manifest.capabilities
+
+    def test_plugin_manifest_engine(self, supertonic_plugin_dir: Path) -> None:
+        result = load_plugin_from_dir(supertonic_plugin_dir)
+        assert result.success is True
+        engine_manifest = result.manifest.engine
+        assert len(engine_manifest.voiceSources) > 0
+        assert len(engine_manifest.audioFormats) > 0
+        assert len(engine_manifest.parameters) > 0
+
+
+# ──────────────────────────────────────────────────────────────
+# Engine Creation (real backend, skipped if not installed)
+# ──────────────────────────────────────────────────────────────
+
+class TestSuperTonicEngineCreation:
+
+    @pytest.mark.skipif(not _supertonic_available(), reason="SuperTonic not installed")
+    def test_create_engine(self, supertonic_plugin_dir: Path, host_context: HostContext) -> None:
+        result = load_plugin_from_dir(supertonic_plugin_dir)
+        assert result.success is True
+        engine = result.create_engine(host_context, None, EngineConfig())
+        assert isinstance(engine, Engine)
+        engine.dispose()
+
+    @pytest.mark.skipif(not _supertonic_available(), reason="SuperTonic not installed")
+    def test_engine_satisfies_protocol(self, supertonic_plugin_dir: Path, host_context: HostContext) -> None:
+        result = load_plugin_from_dir(supertonic_plugin_dir)
+        assert result.success is True
+        engine = result.create_engine(host_context, None, EngineConfig())
+        assert isinstance(engine, Engine)
+        engine.dispose()
+
+
+# ──────────────────────────────────────────────────────────────
+# Engine / Session Contract (inherited from base)
+# ──────────────────────────────────────────────────────────────
+
+class TestSuperTonicEngineContract(EngineContractMixin):
+    """Every test from EngineContractMixin runs against SuperTonicEngine."""
+
+    @pytest.fixture
+    def default_voice(self) -> str:
+        return "M1"
+
+
+# ──────────────────────────────────────────────────────────────
+# VoiceLister Tests
+# ──────────────────────────────────────────────────────────────
+
+class TestSuperTonicVoiceLister:
+
+    def test_list_voices(self) -> None:
+        engine = _make_mock_engine()
+        voices = engine.listVoices("builtin")
+        assert len(voices) == 10
+        assert all(hasattr(v, "id") for v in voices)
+        assert all(hasattr(v, "name") for v in voices)
+        engine.dispose()
+
+    def test_voices_have_tags(self) -> None:
+        engine = _make_mock_engine()
+        for voice in engine.listVoices("builtin"):
+            assert isinstance(voice.tags, tuple)
+            assert len(voice.tags) > 0
+        engine.dispose()
+
+    def test_male_voices_have_male_tag(self) -> None:
+        engine = _make_mock_engine()
+        for v in engine.listVoices("builtin"):
+            if v.id.startswith("M"):
+                assert "male" in v.tags
+        engine.dispose()
+
+    def test_female_voices_have_female_tag(self) -> None:
+        engine = _make_mock_engine()
+        for v in engine.listVoices("builtin"):
+            if v.id.startswith("F"):
+                assert "female" in v.tags
+        engine.dispose()
+
+    def test_list_voices_after_dispose_raises(self) -> None:
+        from abogen.tts_plugin.errors import EngineError
+        engine = _make_mock_engine()
+        engine.dispose()
+        with pytest.raises(EngineError):
+            engine.listVoices("builtin")
+
+
+# ──────────────────────────────────────────────────────────────
+# SuperTonic-specific parameter tests
+# ──────────────────────────────────────────────────────────────
+
+class TestSuperTonicParameters:
+
+    def test_speed_parameter(self) -> None:
+        engine = _make_mock_engine()
+        session = engine.createSession()
+        request = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="M1"),
+            parameters=ParameterValues(values={"speed": 1.5}),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        result = session.synthesize(request)
+        assert isinstance(result.data, bytes)
+        session.dispose()
+        engine.dispose()
+
+    def test_total_steps_parameter(self) -> None:
+        engine = _make_mock_engine()
+        session = engine.createSession()
+        request = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="M1"),
+            parameters=ParameterValues(values={"total_steps": 10}),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        result = session.synthesize(request)
+        assert isinstance(result.data, bytes)
+        session.dispose()
+        engine.dispose()
+
+    def test_default_parameters(self) -> None:
+        engine = _make_mock_engine()
+        session = engine.createSession()
+        request = SynthesisRequest(
+            text="Hello",
+            voice=VoiceSelection(source="builtin", key="M1"),
+            parameters=ParameterValues(values={}),
+            format=AudioFormat(mime="audio/wav", extension="wav"),
+        )
+        result = session.synthesize(request)
+        assert isinstance(result.data, bytes)
+        session.dispose()
+        engine.dispose()

tests/test_tts_supertonic_unsupported_chars.py

@@ -1,53 +0,0 @@
-import numpy as np
-
-from abogen.tts_supertonic import SupertonicPipeline
-
-
-class _DummyTTS:
-    def get_voice_style(self, voice_name: str):
-        return {"voice": voice_name}
-
-    def synthesize(
-        self,
-        *,
-        text: str,
-        voice_style,
-        total_steps: int,
-        speed: float,
-        max_chunk_length: int,
-        silence_duration: float,
-        verbose: bool,
-    ):
-        if "•" in text:
-            raise ValueError("Found 1 unsupported character(s): ['•']")
-        # Return 50ms of audio at 24kHz.
-        sr = 24000
-        audio = np.zeros(int(0.05 * sr), dtype="float32")
-        return audio, 0.05
-
-
-def test_supertonic_pipeline_strips_unsupported_characters_and_retries():
-    # Avoid importing/initializing real supertonic by manually constructing the pipeline.
-    pipeline = SupertonicPipeline.__new__(SupertonicPipeline)
-    pipeline.sample_rate = 24000
-    pipeline.total_steps = 5
-    pipeline.max_chunk_length = 1000
-    pipeline._tts = _DummyTTS()
-
-    segs = list(pipeline("Hello • world", voice="M1", speed=1.0))
-    assert len(segs) == 1
-    assert segs[0].graphemes == "Hello  world" or segs[0].graphemes == "Hello world"
-    assert isinstance(segs[0].audio, np.ndarray)
-    assert segs[0].audio.dtype == np.float32
-    assert segs[0].audio.size > 0
-
-
-def test_supertonic_pipeline_drops_chunk_if_only_unsupported_characters():
-    pipeline = SupertonicPipeline.__new__(SupertonicPipeline)
-    pipeline.sample_rate = 24000
-    pipeline.total_steps = 5
-    pipeline.max_chunk_length = 1000
-    pipeline._tts = _DummyTTS()
-
-    segs = list(pipeline("•", voice="M1", speed=1.0))
-    assert segs == []

tests/test_voice_cache.py

@@ -3,7 +3,7 @@ from typing import cast
 
 import pytest
 
-from abogen.constants import VOICES_INTERNAL
+from abogen.tts_plugin.utils import get_voices
 from abogen.voice_cache import (
     LocalEntryNotFoundError,
     _CACHED_VOICES,
@@ -66,4 +66,4 @@ def test_collect_required_voice_ids_includes_all():
     voices = _collect_required_voice_ids(cast(Job, job))
 
     assert {"af_nova", "am_liam", "am_michael"}.issubset(voices)
-    assert voices.issuperset(VOICES_INTERNAL)
+    assert voices.issuperset(get_voices("kokoro"))

tests/test_voice_formula_resolution.py

@@ -1,18 +0,0 @@
-from __future__ import annotations
-
-from abogen.webui.conversion_runner import _resolve_voice, _supertonic_voice_from_spec
-from abogen.tts_supertonic import DEFAULT_SUPERTONIC_VOICES
-
-
-def test_resolve_voice_formula_without_pipeline_does_not_crash() -> None:
-    # This can happen when a previously-saved Kokoro mix formula is present
-    # but the active provider is SuperTonic (no Kokoro pipeline object).
-    formula = "af_heart*0.5+af_sky*0.5"
-    resolved = _resolve_voice(None, formula, use_gpu=False)
-    assert resolved == formula
-
-
-def test_supertonic_voice_from_formula_falls_back_to_valid_voice() -> None:
-    # When a stale Kokoro mix formula is present, SuperTonic should not receive it.
-    chosen = _supertonic_voice_from_spec("af_heart*0.5+af_sky*0.5", "af_heart*1.0")
-    assert chosen in DEFAULT_SUPERTONIC_VOICES

tests/test_voice_metadata.py

@@ -0,0 +1,233 @@
+import pytest
+
+from abogen.voice_metadata import VoiceMetadata
+
+
+class TestVoiceMetadataCreation:
+    def test_create_with_all_fields(self):
+        voice = VoiceMetadata(
+            id="af_alloy",
+            display_name="Alloy",
+            language="a",
+            gender="female",
+            backend_id="kokoro",
+        )
+        assert voice.id == "af_alloy"
+        assert voice.display_name == "Alloy"
+        assert voice.language == "a"
+        assert voice.gender == "female"
+        assert voice.backend_id == "kokoro"
+
+    def test_create_supertonic_voice(self):
+        voice = VoiceMetadata(
+            id="M1",
+            display_name="Male 1",
+            language="en",
+            gender="male",
+            backend_id="supertonic",
+        )
+        assert voice.id == "M1"
+        assert voice.backend_id == "supertonic"
+
+    def test_create_with_unknown_gender(self):
+        voice = VoiceMetadata(
+            id="custom_voice",
+            display_name="Custom",
+            language="en",
+            gender="unknown",
+            backend_id="custom_backend",
+        )
+        assert voice.gender == "unknown"
+
+
+class TestVoiceMetadataImmutability:
+    def test_frozen_dataclass(self):
+        voice = VoiceMetadata(
+            id="af_alloy",
+            display_name="Alloy",
+            language="a",
+            gender="female",
+            backend_id="kokoro",
+        )
+        with pytest.raises(AttributeError):
+            voice.id = "new_id"
+
+    def test_cannot_modify_display_name(self):
+        voice = VoiceMetadata(
+            id="af_alloy",
+            display_name="Alloy",
+            language="a",
+            gender="female",
+            backend_id="kokoro",
+        )
+        with pytest.raises(AttributeError):
+            voice.display_name = "New Name"
+
+    def test_cannot_modify_backend_id(self):
+        voice = VoiceMetadata(
+            id="af_alloy",
+            display_name="Alloy",
+            language="a",
+            gender="female",
+            backend_id="kokoro",
+        )
+        with pytest.raises(AttributeError):
+            voice.backend_id = "new_backend"
+
+
+class TestVoiceMetadataEquality:
+    def test_equal_voices_are_equal(self):
+        voice1 = VoiceMetadata(
+            id="af_alloy",
+            display_name="Alloy",
+            language="a",
+            gender="female",
+            backend_id="kokoro",
+        )
+        voice2 = VoiceMetadata(
+            id="af_alloy",
+            display_name="Alloy",
+            language="a",
+            gender="female",
+            backend_id="kokoro",
+        )
+        assert voice1 == voice2
+
+    def test_different_voices_are_not_equal(self):
+        voice1 = VoiceMetadata(
+            id="af_alloy",
+            display_name="Alloy",
+            language="a",
+            gender="female",
+            backend_id="kokoro",
+        )
+        voice2 = VoiceMetadata(
+            id="am_adam",
+            display_name="Adam",
+            language="a",
+            gender="male",
+            backend_id="kokoro",
+        )
+        assert voice1 != voice2
+
+    def test_different_backend_id_not_equal(self):
+        voice1 = VoiceMetadata(
+            id="custom",
+            display_name="Custom",
+            language="en",
+            gender="unknown",
+            backend_id="backend_a",
+        )
+        voice2 = VoiceMetadata(
+            id="custom",
+            display_name="Custom",
+            language="en",
+            gender="unknown",
+            backend_id="backend_b",
+        )
+        assert voice1 != voice2
+
+
+class TestVoiceMetadataHashing:
+    def test_hashable(self):
+        voice = VoiceMetadata(
+            id="af_alloy",
+            display_name="Alloy",
+            language="a",
+            gender="female",
+            backend_id="kokoro",
+        )
+        assert hash(voice) is not None
+
+    def test_equal_voices_same_hash(self):
+        voice1 = VoiceMetadata(
+            id="af_alloy",
+            display_name="Alloy",
+            language="a",
+            gender="female",
+            backend_id="kokoro",
+        )
+        voice2 = VoiceMetadata(
+            id="af_alloy",
+            display_name="Alloy",
+            language="a",
+            gender="female",
+            backend_id="kokoro",
+        )
+        assert hash(voice1) == hash(voice2)
+
+    def test_usable_in_set(self):
+        voice1 = VoiceMetadata(
+            id="af_alloy",
+            display_name="Alloy",
+            language="a",
+            gender="female",
+            backend_id="kokoro",
+        )
+        voice2 = VoiceMetadata(
+            id="af_alloy",
+            display_name="Alloy",
+            language="a",
+            gender="female",
+            backend_id="kokoro",
+        )
+        voice3 = VoiceMetadata(
+            id="am_adam",
+            display_name="Adam",
+            language="a",
+            gender="male",
+            backend_id="kokoro",
+        )
+        voice_set = {voice1, voice2, voice3}
+        assert len(voice_set) == 2
+
+
+class TestVoiceMetadataUseCases:
+    def test_backend_populates_backend_id(self):
+        """Simulate how a backend would populate backend_id automatically."""
+
+        class MockBackend:
+            def __init__(self):
+                self._backend_id = "kokoro"
+
+            def get_voices(self):
+                return [
+                    VoiceMetadata(
+                        id="af_alloy",
+                        display_name="Alloy",
+                        language="a",
+                        gender="female",
+                        backend_id=self._backend_id,
+                    ),
+                ]
+
+        backend = MockBackend()
+        voices = backend.get_voices()
+        assert voices[0].backend_id == "kokoro"
+
+    def test_filter_by_language(self):
+        voices = [
+            VoiceMetadata(id="af_alloy", display_name="Alloy", language="a", gender="female", backend_id="kokoro"),
+            VoiceMetadata(id="jf_alpha", display_name="Alpha", language="j", gender="female", backend_id="kokoro"),
+            VoiceMetadata(id="am_adam", display_name="Adam", language="a", gender="male", backend_id="kokoro"),
+        ]
+        english_voices = [v for v in voices if v.language == "a"]
+        assert len(english_voices) == 2
+
+    def test_filter_by_gender(self):
+        voices = [
+            VoiceMetadata(id="af_alloy", display_name="Alloy", language="a", gender="female", backend_id="kokoro"),
+            VoiceMetadata(id="am_adam", display_name="Adam", language="a", gender="male", backend_id="kokoro"),
+            VoiceMetadata(id="am_puck", display_name="Puck", language="a", gender="male", backend_id="kokoro"),
+        ]
+        male_voices = [v for v in voices if v.gender == "male"]
+        assert len(male_voices) == 2
+
+    def test_filter_by_backend(self):
+        voices = [
+            VoiceMetadata(id="af_alloy", display_name="Alloy", language="a", gender="female", backend_id="kokoro"),
+            VoiceMetadata(id="M1", display_name="Male 1", language="en", gender="male", backend_id="supertonic"),
+        ]
+        kokoro_voices = [v for v in voices if v.backend_id == "kokoro"]
+        assert len(kokoro_voices) == 1
+        assert kokoro_voices[0].id == "af_alloy"