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/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

README.md

@@ -38,8 +38,6 @@ This method handles everything automatically - installing all dependencies inclu
 #### <b>OPTION 2: Install using uv</b>
 First, [install uv](https://docs.astral.sh/uv/getting-started/installation/) if you haven't already.
 
-The CUDA extras install both GPU-accelerated Kokoro (via PyTorch) and Supertonic (via onnxruntime-gpu).
-
 ```bash
 # For NVIDIA GPUs (CUDA 12.8) - Recommended
 uv tool install --python 3.12 abogen[cuda] --extra-index-url https://download.pytorch.org/whl/cu128 --index-strategy unsafe-best-match
@@ -67,9 +65,6 @@ venv\Scripts\activate
 # We need to use an older version of PyTorch (2.8.0) until this issue is fixed: https://github.com/pytorch/pytorch/issues/166628
 pip install torch==2.8.0+cu128 torchvision==0.23.0+cu128 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128
 
-# Also install onnxruntime-gpu for Supertonic GPU acceleration:
-pip install onnxruntime-gpu
-
 # For AMD GPUs:
 # Not supported yet, because ROCm is not available on Windows. Use Linux if you have AMD GPU.
 
@@ -178,7 +173,7 @@ Abogen offers **two interfaces**, but currently they have different feature sets
 
 | Command | Interface | Features |
 |---------|-----------|----------|
-| `abogen` | PyQt6 Desktop GUI | Stable core features + **Supertonic TTS**|
+| `abogen` | PyQt6 Desktop GUI | Stable core features |
 | `abogen-web` | Flask Web UI | Core features + **Supertonic TTS**, **LLM Normalization**, **Audiobookshelf Integration** and more! |
 
 > **Note:** The Web UI is under active development. We are working to integrate these new features into the PyQt desktop app. until then, the Web UI provides the most feature-rich experience.
@@ -412,18 +407,18 @@ When Audiobookshelf sits behind Nginx Proxy Manager (NPM), make sure the API pat
 1. Create a **Proxy Host** that points to your ABS container or host (default forward port `13378`).
 2. Under the **SSL** tab, enable your certificate and tick **Force SSL** if you want HTTPS only.
 3. In the **Advanced** tab, append the snippet below so bearer tokens, client IPs, and large uploads survive the proxy hop:
-```nginx
-proxy_set_header Host $host;
-proxy_set_header X-Real-IP $remote_addr;
-proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
-proxy_set_header X-Forwarded-Proto $scheme;
-proxy_set_header X-Forwarded-Host $host;
-proxy_set_header X-Forwarded-Port $server_port;
-proxy_set_header Authorization $http_authorization;
-client_max_body_size 5g;
-proxy_read_timeout 300s;
-proxy_connect_timeout 300s;
-  ```
+   ```nginx
+  proxy_set_header Host $host;
+  proxy_set_header X-Real-IP $remote_addr;
+  proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+  proxy_set_header X-Forwarded-Proto $scheme;
+  proxy_set_header X-Forwarded-Host $host;
+  proxy_set_header X-Forwarded-Port $server_port;
+  proxy_set_header Authorization $http_authorization;
+  client_max_body_size 5g;
+  proxy_read_timeout 300s;
+  proxy_connect_timeout 300s;
+   ```
 4. Disable **Block Common Exploits** (it strips Authorization headers in some NPM builds).
 5. Enable **Websockets Support** on the main proxy screen (Audiobookshelf uses it for the web UI, and it keeps the reverse proxy configuration consistent).
 6. If you publish Audiobookshelf under a path prefix (for example `/abs`), add a **Custom Location** with `Location: /abs/` and set the **Forward Path** to `/`. That rewrite strips the `/abs` prefix before traffic reaches Audiobookshelf so `/abs/api/...` on the internet becomes `/api/...` on the backend. Use the same prefixed URL in Abogen’s “Base URL” field.

WINDOWS_INSTALL.bat

@@ -323,13 +323,6 @@ if /I "%IS_NVIDIA%"=="true" (
             pause
             exit /b
         )
-        echo Installing onnxruntime-gpu for Supertonic GPU acceleration...
-        %PYTHON_CONSOLE_PATH% -m uv pip install --system onnxruntime-gpu
-        if errorlevel 1 (
-            echo Failed to install onnxruntime-gpu.
-            pause
-            exit /b
-        )
     ) else (
         echo CUDA is available on NVIDIA GPU.
     )
@@ -355,13 +348,6 @@ if /I "%IS_NVIDIA%"=="true" (
             pause
             exit /b
         )
-        echo Installing onnxruntime-gpu for Supertonic GPU acceleration...
-        %PYTHON_CONSOLE_PATH% -m uv pip install --system onnxruntime-gpu
-        if errorlevel 1 (
-            echo Failed to install onnxruntime-gpu.
-            pause
-            exit /b
-        )
     )
 )
 

abogen/constants.py

@@ -29,57 +29,6 @@ LANGUAGE_DESCRIPTIONS = {
     "z": "Mandarin Chinese",
 }
 
-# Mapping from Kokoro single-letter language codes to ISO 3166-1 alpha-2 country codes
-# Used for loading flag icons
-KOKORO_LANG_TO_COUNTRY = {
-    "a": "us",  # American English -> United States
-    "b": "gb",  # British English -> United Kingdom
-    "e": "es",  # Spanish -> Spain
-    "f": "fr",  # French -> France
-    "h": "in",  # Hindi -> India
-    "i": "it",  # Italian -> Italy
-    "j": "jp",  # Japanese -> Japan
-    "p": "br",  # Brazilian Portuguese -> Brazil
-    "z": "cn",  # Mandarin Chinese -> China
-}
-
-# Mapping from Supertonic ISO 639-1 language codes to ISO 3166-1 alpha-2 country codes
-# Used for loading flag icons in the Supertonic language picker
-SUPERTONIC_LANG_TO_COUNTRY = {
-    "en": "gb",
-    "ko": "kr",
-    "ja": "jp",
-    "ar": "ae",
-    "bg": "bg",
-    "cs": "cz",
-    "da": "dk",
-    "de": "de",
-    "el": "gr",
-    "es": "es",
-    "et": "ee",
-    "fi": "fi",
-    "fr": "fr",
-    "hi": "in",
-    "hr": "hr",
-    "hu": "hu",
-    "id": "id",
-    "it": "it",
-    "lt": "lt",
-    "lv": "lv",
-    "nl": "nl",
-    "pl": "pl",
-    "pt": "pt",
-    "ro": "ro",
-    "ru": "ru",
-    "sk": "sk",
-    "sl": "si",
-    "sv": "se",
-    "tr": "tr",
-    "uk": "ua",
-    "vi": "vn",
-    "na": "na",
-}
-
 # Supported sound formats
 SUPPORTED_SOUND_FORMATS = [
     "wav",
@@ -114,64 +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/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,
@@ -20,7 +21,6 @@ from abogen.constants import (
     SUPPORTED_SOUND_FORMATS,
     SUPPORTED_SUBTITLE_FORMATS,
 )
-from abogen.tts_supertonic import SupertonicPipeline, SUPERTONIC_AVAILABLE_LANGS, DEFAULT_SUPERTONIC_VOICES
 from abogen.voice_formulas import get_new_voice
 import abogen.hf_tracker as hf_tracker
 import static_ffmpeg
@@ -260,30 +260,21 @@ class ConversionThread(QThread):
         output_folder,
         subtitle_mode,
         output_format,
-        np_module,
-        kpipeline_class,
+        backend,
         start_time,
         total_char_count,
         use_gpu=True,
         from_queue=False,
         save_base_path=None,
-        tts_provider="kokoro",
-        supertonic_language="en",
-        supertonic_total_steps=8,
-    ):
+    ):  # Add use_gpu parameter
         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
         self.voice = voice
-        self.tts_provider = tts_provider
-        self.supertonic_language = supertonic_language
-        self.supertonic_total_steps = supertonic_total_steps
-        self.sample_rate = 44100 if tts_provider == "supertonic" else 24000
         self.save_option = save_option
         self.output_folder = output_folder
         self.subtitle_mode = subtitle_mode
@@ -435,10 +426,6 @@ class ConversionThread(QThread):
             )
             self.log_updated.emit(f"- Voice: {self.voice}")
             self.log_updated.emit(f"- Speed: {self.speed}")
-            tts_provider_label = self.tts_provider.capitalize()
-            if self.tts_provider == "supertonic":
-                tts_provider_label += f" (lang={self.supertonic_language}, steps={self.supertonic_total_steps})"
-            self.log_updated.emit(f"- TTS Engine: {tts_provider_label}")
             self.log_updated.emit(f"- Subtitle mode: {self.subtitle_mode}")
             self.log_updated.emit(f"- Output format: {self.output_format}")
             self.log_updated.emit(
@@ -502,26 +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"
-
-            if self.tts_provider == "supertonic":
-                tts = SupertonicPipeline(
-                    sample_rate=self.sample_rate,
-                    lang=self.supertonic_language,
-                    total_steps=self.supertonic_total_steps,
-                )
-            else:
-                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
@@ -557,7 +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:
@@ -766,7 +733,7 @@ class ConversionThread(QThread):
                 merged_out_path = f"{base_filepath_no_ext}.{self.output_format}"
                 subtitle_entries = []
                 current_time = 0.0
-                rate = self.sample_rate
+                rate = 24000
                 subtitle_mode = self.subtitle_mode
                 self.etr_start_time = time.time()
                 self.processed_char_count = 0
@@ -782,7 +749,7 @@ class ConversionThread(QThread):
                     merged_out_file = sf.SoundFile(
                         merged_out_path,
                         "w",
-                        samplerate=self.sample_rate,
+                        samplerate=24000,
                         channels=1,
                         format=self.output_format,
                     )
@@ -798,18 +765,62 @@ class ConversionThread(QThread):
                     # Prepare ffmpeg command for m4b output
                     cmd = [
                         "ffmpeg",
-                            "-y",
-                            "-thread_queue_size",
-                            "32768",
-                            "-f",
-                            "f32le",
-                            "-ar",
-                            str(self.sample_rate),
-                            "-ac",
-                            "1",
-                            "-i",
-                            "pipe:0",
+                        "-y",
+                        "-thread_queue_size",
+                        "32768",
+                        "-f",
+                        "f32le",
+                        "-ar",
+                        "24000",
+                        "-ac",
+                        "1",
+                        "-i",
+                        "pipe:0",
+                    ]
+                    if cover_path and os.path.exists(cover_path):
+                        cmd.extend(
+                            [
+                                "-i",
+                                cover_path,
+                                "-map",
+                                "0:a",
+                                "-map",
+                                "1",
+                                "-c:v",
+                                "copy",
+                                "-disposition:v",
+                                "attached_pic",
+                            ]
+                        )
+                    cmd.extend(
+                        [
+                            "-c:a",
+                            "aac",
+                            "-q:a",
+                            "2",
+                            "-movflags",
+                            "+faststart+use_metadata_tags",
                         ]
+                    )
+                    cmd += metadata_options
+                    cmd.append(merged_out_path)
+                    ffmpeg_proc = create_process(cmd, stdin=subprocess.PIPE, text=False)
+                elif self.output_format == "opus":
+                    static_ffmpeg.add_paths()
+                    cmd = [
+                        "ffmpeg",
+                        "-y",
+                        "-thread_queue_size",
+                        "32768",
+                        "-f",
+                        "f32le",
+                        "-ar",
+                        "24000",
+                        "-ac",
+                        "1",
+                        "-i",
+                        "pipe:0",
+                    ]
                     cmd.extend(["-c:a", "libopus", "-b:a", "24000"])
                     cmd.append(merged_out_path)
                     ffmpeg_proc = create_process(cmd, stdin=subprocess.PIPE, text=False)
@@ -891,7 +902,7 @@ class ConversionThread(QThread):
                 merged_out_path = None
                 subtitle_entries = []
                 current_time = 0.0
-                rate = self.sample_rate
+                rate = 24000
                 subtitle_mode = self.subtitle_mode
                 self.etr_start_time = time.time()
                 self.processed_char_count = 0
@@ -944,7 +955,7 @@ class ConversionThread(QThread):
                         chapter_out_file = sf.SoundFile(
                             chapter_out_path,
                             "w",
-                            samplerate=self.sample_rate,
+                            samplerate=24000,
                             channels=1,
                             format=separate_chapters_format,
                         )
@@ -959,7 +970,7 @@ class ConversionThread(QThread):
                             "-f",
                             "f32le",
                             "-ar",
-                            str(self.sample_rate),
+                            "24000",
                             "-ac",
                             "1",
                             "-i",
@@ -1046,7 +1057,7 @@ class ConversionThread(QThread):
                 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, tts)
+                        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"))
@@ -1055,7 +1066,7 @@ class ConversionThread(QThread):
                             (f"⚠ Voice loading error for '{voice_name}', continuing with previous", "orange")
                         )
                         if segment_idx == 0:
-                            loaded_voice = self.load_voice_cached(self.voice, tts)
+                            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
@@ -1141,7 +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,
@@ -1341,9 +1352,9 @@ class ConversionThread(QThread):
                 # Add silence between chapters for merged output (except after the last chapter)
                 if merge_chapters_at_end and chapter_idx < total_chapters:
                     silence_samples = int(
-                        self.silence_duration * self.sample_rate
+                        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:
@@ -1572,7 +1583,7 @@ class ConversionThread(QThread):
                 parent_dir, f"{sanitized_base_name}{suffix}"
             )
             merged_out_path = f"{base_filepath_no_ext}.{self.output_format}"
-            rate = self.sample_rate
+            rate = 24000
 
             # Setup audio output
             merged_out_file, ffmpeg_proc = None, None
@@ -1682,7 +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"
             )
 
@@ -1746,7 +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,
@@ -1764,11 +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"
                     )
                 )
@@ -1802,8 +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)
                                 )
                             ),
                         )
@@ -1836,7 +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",
                         )
@@ -1850,7 +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,
@@ -1861,14 +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"
                             )
                         )
@@ -1885,10 +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"
                             ),
                         ]
@@ -1901,10 +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"
                             ),
                         ]
@@ -1946,7 +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})"
@@ -2386,6 +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:
@@ -2415,28 +2430,19 @@ class VoicePreviewThread(QThread):
 
     def __init__(
         self,
-        np_module,
-        kpipeline_class,
+        backend,
         lang_code,
         voice,
         speed,
         use_gpu=False,
         parent=None,
-        tts_provider="kokoro",
-        supertonic_language="en",
-        supertonic_total_steps=8,
     ):
         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
         self.use_gpu = use_gpu
-        self.tts_provider = tts_provider
-        self.supertonic_language = supertonic_language
-        self.supertonic_total_steps = supertonic_total_steps
-        self.sample_rate = 44100 if tts_provider == "supertonic" else 24000
 
         # Cache location for preview audio
         self.cache_dir = get_user_cache_path("preview_cache")
@@ -2446,11 +2452,6 @@ class VoicePreviewThread(QThread):
 
     def _get_cache_path(self):
         """Generate a unique filename for the voice with its parameters"""
-        if self.tts_provider == "supertonic":
-            voice_id = self.voice or "M1"
-            filename = f"st_{voice_id}_{self.supertonic_language}_steps{self.supertonic_total_steps}_{self.speed:.2f}.wav"
-            return os.path.join(self.cache_dir, filename)
-
         # For a voice formula, use a hash of the formula
         if "*" in self.voice:
             voice_id = (
@@ -2465,56 +2466,27 @@ class VoicePreviewThread(QThread):
 
     def run(self):
         print(
-            f"\nVoice: {self.voice}\nLanguage: {self.lang_code}\nSpeed: {self.speed}\nGPU: {self.use_gpu}\nTTS Provider: {self.tts_provider}\n"
+            f"\nVoice: {self.voice}\nLanguage: {self.lang_code}\nSpeed: {self.speed}\nGPU: {self.use_gpu}\n"
         )
 
         # Generate the preview and save to cache
         try:
-            if self.tts_provider == "supertonic":
-                from abogen.tts_supertonic import SupertonicPipeline
 
-                tts = SupertonicPipeline(
-                    sample_rate=self.sample_rate,
-                    lang=self.supertonic_language,
-                    total_steps=self.supertonic_total_steps,
-                )
-                loaded_voice = self.voice or "M1"
-                sample_text = "Hello, this is a sample of the selected voice."
-                audio_segments = []
-                for result in tts(
-                    sample_text,
-                    voice=loaded_voice,
-                    speed=self.speed,
-                    split_pattern=None,
-                ):
-                    audio_segments.append(result.audio)
+            # Enable voice formula support for preview
+            if "*" in self.voice:
+                loaded_voice = get_new_voice(self.backend, self.voice, self.use_gpu)
             else:
-                # Set device based on use_gpu setting and platform
-                if self.use_gpu:
-                    if platform.system() == "Darwin" and platform.processor() == "arm":
-                        device = "mps"
-                    else:
-                        device = "cuda"
-                else:
-                    device = "cpu"
-
-                tts = self.kpipeline_class(
-                    lang_code=self.lang_code, repo_id="hexgrad/Kokoro-82M", device=device
-                )
-                if "*" in self.voice:
-                    loaded_voice = get_new_voice(tts, 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(
-                    sample_text, voice=loaded_voice, speed=self.speed, split_pattern=None
-                ):
-                    audio_segments.append(result.audio)
-
+                loaded_voice = self.voice
+            sample_text = get_sample_voice_text(self.lang_code)
+            audio_segments = []
+            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)
-                sf.write(self.cache_path, audio, self.sample_rate)
+                audio = np.concatenate(audio_segments)
+                # Save directly to the cache path
+                sf.write(self.cache_path, audio, 24000)
                 self.temp_wav = self.cache_path
             self.finished.emit()
         except Exception as e:

abogen/pyqt/gui.py

@@ -9,7 +9,6 @@ from abogen.pyqt.queue_manager_gui import QueueManager
 from abogen.pyqt.queued_item import QueuedItem
 import abogen.hf_tracker as hf_tracker
 import hashlib  # Added for cache path generation
-from abogen.tts_supertonic import SUPERTONIC_AVAILABLE_LANGS, DEFAULT_SUPERTONIC_VOICES
 from PyQt6.QtWidgets import (
     QApplication,
     QWidget,
@@ -83,13 +82,11 @@ from abogen.constants import (
     GITHUB_URL,
     PROGRAM_DESCRIPTION,
     LANGUAGE_DESCRIPTIONS,
-    VOICES_INTERNAL,
-    KOKORO_LANG_TO_COUNTRY,
-    SUPERTONIC_LANG_TO_COUNTRY,
     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
@@ -973,9 +970,6 @@ class abogen(QWidget):
         self.fix_nonstandard_punctuation = self.config.get(
             "fix_nonstandard_punctuation", False
         )
-        self.tts_provider_config = self.config.get("tts_provider", "kokoro")
-        self.supertonic_language_config = self.config.get("supertonic_language", "en")
-        self.supertonic_total_steps_config = self.config.get("supertonic_total_steps", 8)
         self._pending_close_event = None
         self.gpu_ok = False  # Initialize GPU availability status
 
@@ -1024,16 +1018,6 @@ class abogen(QWidget):
                 else:
                     self.mixed_voice_state = entry
                     self.selected_lang = entry[0][0] if entry and entry[0] else None
-        # Restore TTS provider and supertonic settings from config
-        provider_text = "Supertonic" if self.tts_provider_config == "supertonic" else "Kokoro"
-        idx_st = self.provider_combo.findText(provider_text)
-        if idx_st >= 0:
-            self.provider_combo.setCurrentIndex(idx_st)
-        self.st_lang_combo.setCurrentText(self.supertonic_language_config)
-        idx_steps = self.st_steps_combo.findData(self.supertonic_total_steps_config)
-        if idx_steps >= 0:
-            self.st_steps_combo.setCurrentIndex(idx_steps)
-
         if self.save_option == "Choose output folder" and self.selected_output_folder:
             self.save_path_label.setText(self.selected_output_folder)
             self.save_path_row_widget.show()
@@ -1123,53 +1107,6 @@ class abogen(QWidget):
         speed_layout.addWidget(self.speed_label)
         controls_layout.addLayout(speed_layout)
         self.speed_slider.valueChanged.connect(self.update_speed_label)
-
-        # TTS Provider selection
-        provider_layout = QHBoxLayout()
-        provider_layout.setSpacing(7)
-        provider_label = QLabel("TTS Engine:", self)
-        provider_layout.addWidget(provider_label)
-        self.provider_combo = QComboBox(self)
-        self.provider_combo.addItem("Kokoro", "kokoro")
-        self.provider_combo.addItem("Supertonic", "supertonic")
-        self.provider_combo.setStyleSheet("QComboBox { min-height: 20px; padding: 6px 12px; }")
-        self.provider_combo.currentIndexChanged.connect(self.on_provider_changed)
-        provider_layout.addWidget(self.provider_combo)
-        controls_layout.addLayout(provider_layout)
-
-        # Supertonic-specific controls (language + steps), hidden by default
-        self.supertonic_row = QWidget()
-        supertonic_row_layout = QHBoxLayout(self.supertonic_row)
-        supertonic_row_layout.setContentsMargins(0, 0, 0, 0)
-        supertonic_row_layout.setSpacing(7)
-
-        st_lang_label = QLabel("Language:", self)
-        supertonic_row_layout.addWidget(st_lang_label)
-        self.st_lang_combo = QComboBox(self)
-        for code in SUPERTONIC_AVAILABLE_LANGS:
-            country_code = SUPERTONIC_LANG_TO_COUNTRY.get(code, code)
-            flag = get_resource_path("abogen.assets.flags", f"{country_code}.png")
-            icon_st = QIcon(flag) if flag and os.path.exists(flag) else QIcon()
-            self.st_lang_combo.addItem(icon_st, code, code)
-        self.st_lang_combo.setCurrentText("en")
-        self.st_lang_combo.setStyleSheet("QComboBox { min-height: 20px; padding: 6px 12px; }")
-        self.st_lang_combo.currentTextChanged.connect(self._on_st_lang_changed)
-        supertonic_row_layout.addWidget(self.st_lang_combo)
-
-        st_steps_label = QLabel("Steps:", self)
-        supertonic_row_layout.addWidget(st_steps_label)
-        self.st_steps_combo = QComboBox(self)
-        for val in range(2, 16):
-            self.st_steps_combo.addItem(str(val), val)
-        self.st_steps_combo.setCurrentIndex(self.st_steps_combo.findData(8))
-        self.st_steps_combo.setStyleSheet("QComboBox { min-height: 20px; padding: 6px 12px; }")
-        self.st_steps_combo.currentIndexChanged.connect(self._on_st_steps_changed)
-        supertonic_row_layout.addWidget(self.st_steps_combo)
-
-        supertonic_row_layout.addStretch()
-        self.supertonic_row.hide()
-        controls_layout.addWidget(self.supertonic_row)
-
         # Voice selection
         voice_layout = QHBoxLayout()
         voice_layout.setSpacing(7)
@@ -1827,12 +1764,6 @@ class abogen(QWidget):
         Update the enabled state of subtitle options based on the selected language.
         For non-English languages, only sentence-based and line-based modes are supported.
         """
-        provider = self.provider_combo.currentData()
-        if provider == "supertonic":
-            self.subtitle_combo.setEnabled(False)
-            self.subtitle_format_combo.setEnabled(False)
-            return
-
         # Check if current file is a subtitle file
         is_subtitle_input = False
         if self.selected_file and self.selected_file.lower().endswith(
@@ -1892,48 +1823,6 @@ class abogen(QWidget):
         # Enable/disable subtitle options based on language
         self.update_subtitle_options_availability()
 
-    def on_provider_changed(self, index):
-        provider = self.provider_combo.itemData(index)
-        self.config["tts_provider"] = provider
-        save_config(self.config)
-        is_supertonic = provider == "supertonic"
-
-        # Show/hide Supertonic controls
-        self.supertonic_row.setVisible(is_supertonic)
-
-        # Update subtitles availability
-        self.update_subtitle_options_availability()
-
-        # Repopulate voice list
-        self.populate_profiles_in_voice_combo()
-
-        # Clear/reset mixed voice state when switching provider
-        if is_supertonic:
-            self.mixed_voice_state = None
-            self.btn_voice_formula_mixer.setEnabled(False)
-            self.voice_combo.setToolTip(
-                "Supertonic voices:\n"
-                "M1-M5 = Male voices\n"
-                "F1-F5 = Female voices"
-            )
-        else:
-            self.btn_voice_formula_mixer.setEnabled(True)
-            self.voice_combo.setToolTip(
-                "The first character represents the language:\n"
-                '"a" => American English\n"b" => British English\n"e" => Spanish\n"f" => French\n"h" => Hindi\n"i" => Italian\n"j" => Japanese\n"p" => Brazilian Portuguese\n"z" => Mandarin Chinese\nThe second character represents the gender:\n"m" => Male\n"f" => Female'
-            )
-
-    def _on_st_lang_changed(self, lang):
-        self.config["supertonic_language"] = lang
-        save_config(self.config)
-        if self.provider_combo.currentData() == "supertonic":
-            self.selected_lang = lang
-            self.update_subtitle_options_availability()
-
-    def _on_st_steps_changed(self):
-        self.config["supertonic_total_steps"] = self.st_steps_combo.currentData()
-        save_config(self.config)
-
     def on_voice_combo_changed(self, index):
         data = self.voice_combo.itemData(index)
         if isinstance(data, str) and data.startswith("profile:"):
@@ -1942,26 +1831,10 @@ class abogen(QWidget):
             from abogen.voice_profiles import load_profiles
 
             entry = load_profiles().get(pname, {})
+            # set mixed voices and language
             if isinstance(entry, dict):
-                entry_provider = str(entry.get("provider", "")).strip().lower()
-                if entry_provider == "supertonic":
-                    # Switch provider to Supertonic if not already
-                    if self.provider_combo.currentData() != "supertonic":
-                        self.provider_combo.setCurrentIndex(1)
-                    self.mixed_voice_state = None
-                    self.selected_lang = entry.get("language", self.st_lang_combo.currentText())
-                    # Sync supertonic controls from profile
-                    profile_steps = entry.get("total_steps")
-                    if profile_steps is not None:
-                        idx_steps = self.st_steps_combo.findData(int(profile_steps))
-                        if idx_steps >= 0:
-                            self.st_steps_combo.setCurrentIndex(idx_steps)
-                    profile_lang = entry.get("language")
-                    if profile_lang and profile_lang in SUPERTONIC_AVAILABLE_LANGS:
-                        self.st_lang_combo.setCurrentText(profile_lang)
-                else:
-                    self.mixed_voice_state = entry.get("voices", [])
-                    self.selected_lang = entry.get("language")
+                self.mixed_voice_state = entry.get("voices", [])
+                self.selected_lang = entry.get("language")
             else:
                 self.mixed_voice_state = entry
                 self.selected_lang = entry[0][0] if entry and entry[0] else None
@@ -1974,12 +1847,7 @@ class abogen(QWidget):
         else:
             self.mixed_voice_state = None
             self.selected_profile_name = None
-            self.selected_voice = data
-            provider = self.provider_combo.currentData()
-            if provider == "supertonic":
-                self.selected_lang = self.st_lang_combo.currentText()
-            else:
-                self.selected_lang = data[0] if data else ""
+            self.selected_voice, self.selected_lang = data, data[0]
             self.config["selected_voice"] = data
             if "selected_profile_name" in self.config:
                 del self.config["selected_profile_name"]
@@ -1998,40 +1866,19 @@ class abogen(QWidget):
     def populate_profiles_in_voice_combo(self):
         # preserve current voice or profile
         current = self.voice_combo.currentData()
-        provider = self.provider_combo.currentData()
         self.voice_combo.blockSignals(True)
         self.voice_combo.clear()
-        # re-add profiles matching current provider
+        # re-add profiles
         profile_icon = QIcon(get_resource_path("abogen.assets", "profile.png"))
-        for pname, entry in load_profiles().items():
-            entry_provider = ""
-            if isinstance(entry, dict):
-                entry_provider = str(entry.get("provider", "")).strip().lower()
-            if provider == "supertonic":
-                if entry_provider == "supertonic":
-                    self.voice_combo.addItem(profile_icon, pname, f"profile:{pname}")
-            else:
-                if entry_provider != "supertonic":
-                    self.voice_combo.addItem(profile_icon, pname, f"profile:{pname}")
+        for pname in load_profiles().keys():
+            self.voice_combo.addItem(profile_icon, pname, f"profile:{pname}")
         # re-add voices
-        if provider == "supertonic":
-            for v in DEFAULT_SUPERTONIC_VOICES:
-                icon = QIcon()
-                if v.startswith("F"):
-                    icon_path = get_resource_path("abogen.assets", "female.png")
-                else:
-                    icon_path = get_resource_path("abogen.assets", "male.png")
-                if icon_path and os.path.exists(icon_path):
-                    icon = QIcon(icon_path)
-                self.voice_combo.addItem(icon, f"{v}", v)
-        else:
-            for v in VOICES_INTERNAL:
-                icon = QIcon()
-                country_code = KOKORO_LANG_TO_COUNTRY.get(v[0], v[0])
-                flag_path = get_resource_path("abogen.assets.flags", f"{country_code}.png")
-                if flag_path and os.path.exists(flag_path):
-                    icon = QIcon(flag_path)
-                self.voice_combo.addItem(icon, f"{v}", v)
+        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):
+                icon = QIcon(flag_path)
+            self.voice_combo.addItem(icon, f"{v}", v)
         # restore selection
         idx = -1
         if self.selected_profile_name:
@@ -2222,9 +2069,6 @@ class abogen(QWidget):
             save_base_path=save_base_path,
             save_chapters_separately=getattr(self, "save_chapters_separately", None),
             merge_chapters_at_end=getattr(self, "merge_chapters_at_end", None),
-            tts_provider=self.provider_combo.currentData(),
-            supertonic_language=self.st_lang_combo.currentText(),
-            supertonic_total_steps=self.st_steps_combo.currentData(),
         )
 
         # Prevent adding duplicate items to the queue
@@ -2368,15 +2212,6 @@ class abogen(QWidget):
                 self.config["replace_numerals"] = self.replace_numerals
                 self.config["fix_nonstandard_punctuation"] = self.fix_nonstandard_punctuation
 
-                # TTS provider settings
-                tts_provider = getattr(queued_item, "tts_provider", "kokoro")
-                self.provider_combo.setCurrentText("Supertonic" if tts_provider == "supertonic" else "Kokoro")
-                self.st_lang_combo.setCurrentText(getattr(queued_item, "supertonic_language", "en"))
-                steps_val = getattr(queued_item, "supertonic_total_steps", 8)
-                idx_steps = self.st_steps_combo.findData(steps_val)
-                if idx_steps >= 0:
-                    self.st_steps_combo.setCurrentIndex(idx_steps)
-
                 # Sync Voice/Profile in config
                 self.config["selected_voice"] = self.selected_voice
                 if "selected_profile_name" in self.config:
@@ -2399,8 +2234,6 @@ class abogen(QWidget):
             self.current_queue_index = 0  # Reset for next time
 
     def get_voice_formula(self) -> str:
-        if self.provider_combo.currentData() == "supertonic":
-            return self._get_supertonic_voice()
         if self.mixed_voice_state:
             formula_components = [
                 f"{name}*{weight}" for name, weight in self.mixed_voice_state
@@ -2410,8 +2243,6 @@ class abogen(QWidget):
             return self.selected_voice
 
     def get_selected_lang(self, voice_formula) -> str:
-        if self.provider_combo.currentData() == "supertonic":
-            return self.st_lang_combo.currentText()
         if self.selected_profile_name:
             from abogen.voice_profiles import load_profiles
 
@@ -2485,9 +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
 
@@ -2501,10 +2332,6 @@ class abogen(QWidget):
             # determine selected language: use profile setting if profile selected, else voice code
             selected_lang = self.get_selected_lang(voice_formula)
 
-            tts_provider = self.provider_combo.currentData()
-            supertonic_language = self.st_lang_combo.currentText()
-            supertonic_total_steps = self.st_steps_combo.currentData()
-
             self.conversion_thread = ConversionThread(
                 self.selected_file,
                 selected_lang,
@@ -2514,17 +2341,13 @@ 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,
                 from_queue=from_queue,
-                save_base_path=self.displayed_file_path,
-                tts_provider=tts_provider,
-                supertonic_language=supertonic_language,
-                supertonic_total_steps=supertonic_total_steps,
-            )
+                save_base_path=self.displayed_file_path,  # Pass the save base path (original file for EPUB)
+            )  # Use gpu_ok status
             # Pass the displayed file path to the log_updated signal handler in ConversionThread
             self.conversion_thread.display_path = display_path
             # Pass the file size string
@@ -2601,15 +2424,22 @@ class abogen(QWidget):
             # Store gpu_ok status to use when creating the conversion thread
             self.gpu_ok = gpu_ok
             self.update_log((gpu_msg, gpu_ok))
-            tts_provider = self.provider_combo.currentData()
-            if tts_provider == "supertonic":
-                # Supertonic doesn't need KPipeline, call callback directly
-                import numpy as np
-                pipeline_loaded_callback(np, None, None)
+            self.update_log("Loading modules...")
+
+            # Determine device based on GPU availability
+            if gpu_ok:
+                if platform.system() == "Darwin" and platform.processor() == "arm":
+                    device = "mps"
+                else:
+                    device = "cuda"
             else:
-                self.update_log("Loading modules...")
-                load_thread = LoadPipelineThread(pipeline_loaded_callback)
-                load_thread.start()
+                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()
 
@@ -2922,32 +2752,9 @@ class abogen(QWidget):
                 "Open File Error", f"Could not open file:\n{e}"
             )
 
-    def _get_supertonic_voice(self) -> str:
-        """Resolve the effective Supertonic voice from the current combo selection."""
-        if self.selected_profile_name:
-            from abogen.voice_profiles import load_profiles
-            entry = load_profiles().get(self.selected_profile_name, {})
-            if isinstance(entry, dict):
-                return str(entry.get("voice", "M1"))
-            return "M1"
-        current_data = self.voice_combo.currentData()
-        if current_data and isinstance(current_data, str) and not current_data.startswith("profile:"):
-            return current_data
-        return "M1"
-
     def _get_preview_cache_path(self):
         """Generate the expected cache path for the current voice settings."""
         speed = self.speed_slider.value() / 100.0
-        provider = self.provider_combo.currentData()
-
-        if provider == "supertonic":
-            voice_to_cache = self._get_supertonic_voice()
-            lang_to_cache = self.st_lang_combo.currentText()
-            steps = self.st_steps_combo.currentData()
-            cache_dir = get_user_cache_path("preview_cache")
-            filename = f"st_{voice_to_cache}_{lang_to_cache}_steps{steps}_{speed:.2f}.wav"
-            return os.path.join(cache_dir, filename)
-
         voice_to_cache = ""
         lang_to_cache = ""
 
@@ -3052,13 +2859,6 @@ class abogen(QWidget):
         self.btn_voice_formula_mixer.setEnabled(False)  # Disable mixer button
         self.btn_start.setEnabled(False)  # Disable start button during preview
 
-        # For Supertonic, skip KPipeline loading and use SupertonicPipeline directly
-        if self.provider_combo.currentData() == "supertonic":
-            import numpy as np
-            self.loading_movie.start()
-            self._on_pipeline_loaded_for_preview(np, None, None)
-            return
-
         # Start loading animation - ensure signal connection is always active
         if hasattr(self, "loading_movie"):
             # Disconnect previous connections to avoid multiple connections
@@ -3075,18 +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)
@@ -3117,28 +2926,14 @@ class abogen(QWidget):
                     else None
                 )
         else:
-            if self.provider_combo.currentData() == "supertonic":
-                voice = self._get_supertonic_voice()
-            else:
-                voice = self.selected_voice or ""
-
-        tts_provider = self.provider_combo.currentData()
-        supertonic_language = self.st_lang_combo.currentText()
-        supertonic_total_steps = self.st_steps_combo.currentData()
-
-        if tts_provider == "supertonic":
-            lang = supertonic_language
-        else:
-            lang = self.selected_voice[0] if self.selected_voice else ""
+            lang = self.selected_voice[0]
+            voice = self.selected_voice
 
         # use same gpu/cpu logic as in conversion
         gpu_msg, gpu_ok = get_gpu_acceleration(self.use_gpu)
 
         self.preview_thread = VoicePreviewThread(
-            np_module, kpipeline_class, lang, voice, speed, gpu_ok,
-            tts_provider=tts_provider,
-            supertonic_language=supertonic_language,
-            supertonic_total_steps=supertonic_total_steps,
+            backend, lang, voice, speed, gpu_ok
         )
         self.preview_thread.finished.connect(self._play_preview_audio)
         self.preview_thread.error.connect(self._preview_error)
@@ -3441,12 +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()

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/queued_item.py

@@ -26,7 +26,3 @@ class QueuedItem:
     replace_all_caps: bool = False
     replace_numerals: bool = False
     fix_nonstandard_punctuation: bool = False
-    # TTS Provider fields
-    tts_provider: str = "kokoro"
-    supertonic_language: str = "en"
-    supertonic_total_steps: int = 8

abogen/pyqt/voice_formula_gui.py

@@ -28,12 +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,
-    KOKORO_LANG_TO_COUNTRY,
     COLORS,
 )
+from abogen.tts_plugin.utils import get_voices
 import re
 import platform
 from abogen.utils import get_resource_path
@@ -180,7 +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()
@@ -190,9 +189,8 @@ class VoiceMixer(QWidget):
         )  # Center the icons horizontally
 
         # Flag icon
-        country_code = KOKORO_LANG_TO_COUNTRY.get(language_code, language_code)
         flag_icon_path = get_resource_path(
-            "abogen.assets.flags", f"{country_code}.png"
+            "abogen.assets.flags", f"{language_code}.png"
         )
         gender_icon_path = get_resource_path(
             "abogen.assets", "female.png" if is_female else "male.png"
@@ -514,8 +512,7 @@ class VoiceFormulaDialog(QDialog):
         header_row.addWidget(QLabel("Language:"))
         self.language_combo = QComboBox()
         for code, desc in LANGUAGE_OPTIONS:
-            country_code = KOKORO_LANG_TO_COUNTRY.get(code, code)
-            flag = get_resource_path("abogen.assets.flags", f"{country_code}.png")
+            flag = get_resource_path("abogen.assets.flags", f"{code}.png")
             if flag and os.path.exists(flag):
                 self.language_combo.addItem(QIcon(flag), desc, code)
             else:
@@ -775,7 +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

@@ -466,7 +466,7 @@ def sanitize_name_for_os(name, is_folder=True):
 
 
 def validate_voice_name(voice_name):
-    """Validate voice name against VOICES_INTERNAL list (case-insensitive).
+    """Validate voice name against available voices (case-insensitive).
     Handles both single voices and formulas like 'af_heart*0.5 + am_echo*0.5'.
 
     Args:
@@ -477,10 +477,10 @@ def validate_voice_name(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.constants import VOICES_INTERNAL
+    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 VOICES_INTERNAL}
+    voice_lookup_lower = {v.lower() for v in get_voices("kokoro")}
     voice_name = voice_name.strip()
 
     # Check if it's a formula (contains *)
@@ -505,7 +505,7 @@ 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 VOICES_INTERNAL.
+    Voice names are normalized to lowercase to match canonical voice names.
 
     Args:
         text: Text potentially containing <<VOICE:name>> markers
@@ -518,7 +518,7 @@ def split_text_by_voice_markers(text, default_voice):
             - valid_count: Number of valid voice markers processed
             - invalid_count: Number of invalid voice markers skipped
     """
-    from abogen.constants import VOICES_INTERNAL
+    from abogen.tts_plugin.utils import get_voices
 
     voice_splits = list(_VOICE_MARKER_SEARCH_PATTERN.finditer(text))
 
@@ -560,7 +560,7 @@ def split_text_by_voice_markers(text, default_voice):
                         # Find the canonical (lowercase) voice name
                         voice_part_lower = voice_part.strip().lower()
                         canonical_voice = next(
-                            (v for v in VOICES_INTERNAL if v.lower() == voice_part_lower),
+                            (v for v in get_voices("kokoro") if v.lower() == voice_part_lower),
                             voice_part.strip()
                         )
                         normalized_parts.append(f"{canonical_voice}*{weight.strip()}")
@@ -569,7 +569,7 @@ def split_text_by_voice_markers(text, default_voice):
                 # Find the canonical (lowercase) voice name
                 voice_name_lower = voice_name.lower()
                 current_voice = next(
-                    (v for v in VOICES_INTERNAL if v.lower() == voice_name_lower),
+                    (v for v in get_voices("kokoro") if v.lower() == voice_name_lower),
                     voice_name
                 )
             valid_markers += 1

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
 
@@ -56,25 +56,26 @@ 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 ""
 
-    upper = raw.upper()
-    if upper in DEFAULT_SUPERTONIC_VOICES:
-        return upper
+    # 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 ""
 
-    fallback_upper = fallback_raw.upper() if fallback_raw else ""
-    if fallback_upper in DEFAULT_SUPERTONIC_VOICES:
-        return fallback_upper
+    # If still empty, use default Supertonic voice
+    if not upper or "*" in upper or "+" in upper:
+        upper = "M1"
 
-    return "M1"
+    return upper
 
 
 def _split_speaker_reference(value: Any) -> tuple[Optional[str], str]:
@@ -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", 8) or 8),
+                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
@@ -1618,7 +1605,7 @@ def run_conversion_job(job: Job) -> None:
                 provider = str(entry.get("provider") or "kokoro").strip().lower() or "kokoro"
                 if provider == "supertonic":
                     voice = str(entry.get("voice") or getattr(job, "voice", "M1") or "M1").strip() or "M1"
-                    steps = int(entry.get("total_steps") or getattr(job, "supertonic_total_steps", 8) or 8)
+                    steps = int(entry.get("total_steps") or getattr(job, "supertonic_total_steps", 5) or 5)
                     speed = float(entry.get("speed") or getattr(job, "speed", 1.0) or 1.0)
                     return "supertonic", _supertonic_voice_from_spec(voice, getattr(job, "voice", "M1")), speed, steps
                 formula = _formula_from_kokoro_entry(entry)
@@ -1634,7 +1621,7 @@ def run_conversion_job(job: Job) -> None:
             """Resolve a raw voice spec into (provider, resolved_spec, choice, speed, steps).
 
             For Kokoro formulas, `choice` will be a resolved voice tensor (via `voice_formulas`).
-            For Supertonic, `choice` will be a valid Supertonic voice id.
+            For SuperTonic, `choice` will be a valid SuperTonic voice id.
             """
 
             provider, resolved, speed, steps = resolve_voice_target(raw_spec)
@@ -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(
@@ -1857,56 +1844,62 @@ def run_conversion_job(job: Job) -> None:
                     voice=voice_name,
                     speed=float(speed_override if speed_override is not None else job.speed),
                     split_pattern=split_pattern,
-                    total_steps=int(supertonic_steps_override if supertonic_steps_override is not None else getattr(job, "supertonic_total_steps", 8)),
+                    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,
                 )
 
-            for segment in segment_iter:
-                canceller()
-                graphemes_raw = getattr(segment, "graphemes", "") or ""
-                graphemes = graphemes_raw.strip()
+            try:
+                for segment in segment_iter:
+                    canceller()
+                    graphemes_raw = getattr(segment, "graphemes", "") or ""
+                    graphemes = graphemes_raw.strip()
 
-                audio = _to_float32(getattr(segment, "audio", None))
-                if audio.size == 0:
-                    continue
+                    audio = _to_float32(getattr(segment, "audio", None))
+                    if audio.size == 0:
+                        continue
 
-                local_segments += 1
-                if chapter_sink:
-                    chapter_sink.write(audio)
-                if audio_sink:
-                    audio_sink.write(audio)
+                    local_segments += 1
+                    if chapter_sink:
+                        chapter_sink.write(audio)
+                    if audio_sink:
+                        audio_sink.write(audio)
 
-                duration = len(audio) / SAMPLE_RATE
-                processed_chars += len(graphemes)
-                job.processed_characters = processed_chars
-                if job.total_characters:
-                    job.progress = min(processed_chars / job.total_characters, 0.999)
-                else:
-                    job.progress = 0.0 if processed_chars == 0 else 0.999
+                    duration = len(audio) / SAMPLE_RATE
+                    processed_chars += len(graphemes)
+                    job.processed_characters = processed_chars
+                    if job.total_characters:
+                        job.progress = min(processed_chars / job.total_characters, 0.999)
+                    else:
+                        job.progress = 0.0 if processed_chars == 0 else 0.999
 
-                preview_text = graphemes or (graphemes_raw[:80] if graphemes_raw else "[silence]")
-                prefix = f"{preview_prefix} · " if preview_prefix else ""
-                job.add_log(f"{prefix}{processed_chars:,}/{job.total_characters or '—'}: {preview_text[:80]}")
+                    preview_text = graphemes or (graphemes_raw[:80] if graphemes_raw else "[silence]")
+                    prefix = f"{preview_prefix} · " if preview_prefix else ""
+                    job.add_log(f"{prefix}{processed_chars:,}/{job.total_characters or '—'}: {preview_text[:80]}")
 
-                if subtitle_writer and audio_sink and graphemes:
-                    subtitle_writer.write_segment(
-                        index=subtitle_index,
-                        text=graphemes,
-                        start=current_time,
-                        end=current_time + duration,
-                    )
-                    subtitle_index += 1
+                    if subtitle_writer and audio_sink and graphemes:
+                        subtitle_writer.write_segment(
+                            index=subtitle_index,
+                            text=graphemes,
+                            start=current_time,
+                            end=current_time + duration,
+                        )
+                        subtitle_index += 1
 
-                if audio_sink:
-                    current_time += duration
+                    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", 8) or 8),
+        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:
@@ -2610,7 +2605,7 @@ def _build_ffmpeg_command(path: Path, fmt: str, metadata: Optional[Dict[str, str
 def _resolve_voice(pipeline, voice_spec: str, use_gpu: bool):
     if "*" in voice_spec:
         # Voice formulas are a Kokoro-only feature (they require a pipeline that can
-        # load individual Kokoro voices). When running with Supertonic (or when the
+        # load individual Kokoro voices). When running with SuperTonic (or when the
         # pipeline is otherwise unavailable), treat the spec as a plain string and
         # allow downstream provider-specific resolution to choose a safe fallback.
         if pipeline is None or not hasattr(pipeline, "load_single_voice"):

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 = {
@@ -162,7 +163,7 @@ def api_voice_profiles_preview() -> ResponseReturnValue:
     formula = str(payload.get("formula") or "").strip()
     profile_name = str(payload.get("profile") or "").strip()
     provider = str(payload.get("tts_provider") or payload.get("provider") or "").strip().lower() or None
-    supertonic_total_steps = int(payload.get("supertonic_total_steps") or payload.get("total_steps") or settings.get("supertonic_total_steps") or 8)
+    supertonic_total_steps = int(payload.get("supertonic_total_steps") or payload.get("total_steps") or settings.get("supertonic_total_steps") or 5)
 
     voice_spec = ""
     resolved_provider = provider or "kokoro"
@@ -224,13 +225,13 @@ def api_speaker_preview() -> ResponseReturnValue:
     speed_value = payload.get("speed")
     speed = coerce_float(speed_value, 1.0)
     tts_provider = str(payload.get("tts_provider") or "").strip().lower()
-    supertonic_total_steps = int(payload.get("supertonic_total_steps") or 8)
+    supertonic_total_steps = int(payload.get("supertonic_total_steps") or 5)
     
     settings = load_settings()
     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))
@@ -269,7 +270,7 @@ def api_speaker_preview() -> ResponseReturnValue:
             use_gpu=use_gpu
             ,
             tts_provider=resolved_provider,
-            supertonic_total_steps=supertonic_total_steps or int(settings.get("supertonic_total_steps") or 8),
+            supertonic_total_steps=supertonic_total_steps or int(settings.get("supertonic_total_steps") or 5),
             pronunciation_overrides=pronunciation_overrides,
             manual_overrides=manual_overrides,
             speakers=speakers,

abogen/webui/routes/settings.py

@@ -43,12 +43,8 @@ def update_settings() -> ResponseReturnValue:
     current["language"] = (form.get("language") or "en").strip()
     current["default_speaker"] = (form.get("default_speaker") or "").strip()
     current["default_voice"] = (form.get("default_voice") or "").strip()
-    provider = str(form.get("tts_provider") or "kokoro").strip().lower()
-    if provider in {"kokoro", "supertonic"}:
-        current["tts_provider"] = provider
     try:
-        total_steps = int(form.get("supertonic_total_steps", current.get("supertonic_total_steps", 8)))
-        current["supertonic_total_steps"] = max(2, min(15, total_steps))
+        current["supertonic_total_steps"] = max(2, min(15, int(form.get("supertonic_total_steps", current.get("supertonic_total_steps", 5)))))
     except (TypeError, ValueError):
         pass
     try:

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
@@ -577,9 +578,9 @@ def apply_book_step_form(
     # NOTE: Do not auto-set a global TTS provider at the book level based on the
     # narrator defaults. Provider is resolved per-speaker/per-chunk from the voice
     # spec (e.g. "speaker:Name" for saved speakers, or a Kokoro mix formula).
-    # This enables mixed-provider conversions (e.g. narrator=Supertonic, characters=Kokoro).
+    # 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
 
@@ -913,15 +914,6 @@ def build_pending_job_from_extraction(
         else:
             normalization_overrides[key] = default_val
 
-    provider_value = str(form.get("tts_provider") or "").strip().lower()
-    if provider_value not in {"kokoro", "supertonic"}:
-        provider_value = settings.get("tts_provider", "kokoro")
-    try:
-        total_steps = int(form.get("supertonic_total_steps", settings.get("supertonic_total_steps", 8)))
-        supertonic_steps = max(2, min(15, total_steps))
-    except (TypeError, ValueError):
-        supertonic_steps = int(settings.get("supertonic_total_steps", 8))
-
     pending = PendingJob(
         id=uuid.uuid4().hex,
         original_filename=original_name,
@@ -937,8 +929,6 @@ def build_pending_job_from_extraction(
         replace_single_newlines=replace_single_newlines,
         subtitle_format=subtitle_format,
         total_characters=total_chars,
-        tts_provider=provider_value,
-        supertonic_total_steps=supertonic_steps,
         save_chapters_separately=save_chapters_separately,
         merge_chapters_at_end=merge_chapters_at_end,
         separate_chapters_format=separate_chapters_format,

abogen/webui/routes/utils/preview.py

@@ -14,6 +14,17 @@ _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
 
@@ -78,10 +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
 
@@ -92,7 +102,7 @@ def generate_preview_audio(
     speed: float,
     use_gpu: bool,
     tts_provider: str = "kokoro",
-    supertonic_total_steps: int = 8,
+    supertonic_total_steps: int = 5,
     max_seconds: float = 8.0,
     pronunciation_overrides: Optional[Iterable[Mapping[str, Any]]] = None,
     manual_overrides: Optional[Iterable[Mapping[str, Any]]] = None,
@@ -137,9 +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,
@@ -201,7 +211,7 @@ def synthesize_preview(
     speed: float,
     use_gpu: bool,
     tts_provider: str = "kokoro",
-    supertonic_total_steps: int = 8,
+    supertonic_total_steps: int = 5,
     max_seconds: float = 8.0,
     pronunciation_overrides: Optional[Iterable[Mapping[str, Any]]] = None,
     manual_overrides: Optional[Iterable[Mapping[str, Any]]] = None,

abogen/webui/routes/utils/service.py

@@ -25,7 +25,7 @@ def submit_job(pending: PendingJob) -> str:
         tts_provider=getattr(pending, "tts_provider", "kokoro"),
         voice=pending.voice,
         speed=pending.speed,
-        supertonic_total_steps=getattr(pending, "supertonic_total_steps", 8),
+        supertonic_total_steps=getattr(pending, "supertonic_total_steps", 5),
         use_gpu=pending.use_gpu,
         subtitle_mode=pending.subtitle_mode,
         output_format=pending.output_format,

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,
@@ -174,9 +174,8 @@ 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 "",
-        "tts_provider": "kokoro",
-        "supertonic_total_steps": 8,
+        "default_voice": get_default_voice("kokoro"),
+        "supertonic_total_steps": 5,
         "supertonic_speed": 1.0,
         "replace_single_newlines": False,
         "use_gpu": True,

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,
@@ -666,7 +666,7 @@ def resolve_voice_choice(
 
         # Provider-aware behavior:
         # - Kokoro profiles typically represent mixes (formula strings).
-        # - Supertonic profiles represent a discrete voice id + settings.
+        # - SuperTonic profiles represent a discrete voice id + settings.
         #   In that case, we return a speaker reference so downstream can
         #   resolve provider per-speaker and allow mixed-provider casting.
         if provider == "supertonic":
@@ -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

@@ -111,7 +111,7 @@ class Job:
     subtitle_format: str
     created_at: float
     tts_provider: str = "kokoro"
-    supertonic_total_steps: int = 8
+    supertonic_total_steps: int = 5
     save_chapters_separately: bool = False
     merge_chapters_at_end: bool = True
     separate_chapters_format: str = "wav"
@@ -204,7 +204,7 @@ class Job:
             "queue_position": self.queue_position,
             "options": {
                 "tts_provider": getattr(self, "tts_provider", "kokoro"),
-                "supertonic_total_steps": getattr(self, "supertonic_total_steps", 8),
+                "supertonic_total_steps": getattr(self, "supertonic_total_steps", 5),
                 "save_chapters_separately": self.save_chapters_separately,
                 "merge_chapters_at_end": self.merge_chapters_at_end,
                 "separate_chapters_format": self.separate_chapters_format,
@@ -552,7 +552,7 @@ class PendingJob:
     normalization_overrides: Dict[str, Any]
     created_at: float
     tts_provider: str = "kokoro"
-    supertonic_total_steps: int = 8
+    supertonic_total_steps: int = 5
     cover_image_path: Optional[Path] = None
     cover_image_mime: Optional[str] = None
     chapter_intro_delay: float = 0.5
@@ -621,7 +621,7 @@ class ConversionService:
         voice: str,
         speed: float,
         tts_provider: str = "kokoro",
-        supertonic_total_steps: int = 8,
+        supertonic_total_steps: int = 5,
         use_gpu: bool,
         subtitle_mode: str,
         output_format: str,
@@ -674,7 +674,7 @@ class ConversionService:
             voice=voice,
             speed=speed,
             tts_provider=tts_provider,
-            supertonic_total_steps=int(supertonic_total_steps or 8),
+            supertonic_total_steps=int(supertonic_total_steps or 5),
             use_gpu=use_gpu,
             subtitle_mode=subtitle_mode,
             output_format=output_format,
@@ -1147,7 +1147,7 @@ class ConversionService:
             "tts_provider": getattr(job, "tts_provider", "kokoro"),
             "voice": job.voice,
             "speed": job.speed,
-            "supertonic_total_steps": getattr(job, "supertonic_total_steps", 8),
+            "supertonic_total_steps": getattr(job, "supertonic_total_steps", 5),
             "use_gpu": job.use_gpu,
             "subtitle_mode": job.subtitle_mode,
             "output_format": job.output_format,
@@ -1275,7 +1275,7 @@ class ConversionService:
             replace_single_newlines=bool(payload.get("replace_single_newlines", False)),
             subtitle_format=payload.get("subtitle_format", "srt"),
             created_at=float(payload.get("created_at", time.time())),
-            supertonic_total_steps=int(payload.get("supertonic_total_steps", 8)),
+            supertonic_total_steps=int(payload.get("supertonic_total_steps", 5)),
             save_chapters_separately=bool(payload.get("save_chapters_separately", False)),
             merge_chapters_at_end=bool(payload.get("merge_chapters_at_end", True)),
             separate_chapters_format=payload.get("separate_chapters_format", "wav"),
@@ -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/webui/templates/partials/new_job_step_book.html

@@ -26,26 +26,6 @@
     {% set subtitle_value = settings_dict.get('subtitle_mode', 'Disabled') %}
   {% endif %}
 {% endif %}
-{% set tts_provider_value = form_values.get('tts_provider') if form_values else None %}
-{% if not tts_provider_value %}
-  {% if pending and pending.tts_provider %}
-    {% set tts_provider_value = pending.tts_provider %}
-  {% else %}
-    {% set tts_provider_value = settings_dict.get('tts_provider', 'kokoro') %}
-  {% endif %}
-{% endif %}
-{% set supertonic_steps_value = form_values.get('supertonic_total_steps') if form_values else None %}
-{% if supertonic_steps_value is none %}
-  {% if pending and pending.supertonic_total_steps is not none %}
-    {% set supertonic_steps_value = pending.supertonic_total_steps %}
-  {% else %}
-    {% set supertonic_steps_value = settings_dict.get('supertonic_total_steps', 8) %}
-  {% endif %}
-{% endif %}
-{% if supertonic_steps_value is not none and supertonic_steps_value is string %}
-  {% set supertonic_steps_value = supertonic_steps_value|int %}
-{% endif %}
-{% set is_supertonic = tts_provider_value == 'supertonic' %}
 {% set generate_flag = form_values.get('generate_epub3') if form_values else None %}
 {% if generate_flag is not none %}
   {% set generate_epub3 = True %}
@@ -293,13 +273,6 @@
       <div class="form-section__layout form-section__layout--split">
         <div class="form-section__group">
           <div class="field">
-            <label for="tts_provider">TTS Engine</label>
-            <select id="tts_provider" name="tts_provider" data-role="tts-provider" {{ 'disabled' if readonly else '' }}>
-              <option value="kokoro" {% if tts_provider_value == 'kokoro' %}selected{% endif %}>Kokoro</option>
-              <option value="supertonic" {% if tts_provider_value == 'supertonic' %}selected{% endif %}>Supertonic</option>
-            </select>
-          </div>
-          <div class="field" data-role="voice-profile-field">
             <label for="voice_profile">Voice profile</label>
             <select id="voice_profile" name="voice_profile" data-role="voice-profile" {{ 'disabled' if readonly else '' }}>
               <option value="__standard" {% if profile_value == '__standard' %}selected{% endif %}>Standard voice</option>
@@ -307,13 +280,13 @@
               {% if options.voice_profile_options %}
               <optgroup label="Saved mixes">
                 {% for profile in options.voice_profile_options %}
-                <option value="{{ profile.name }}" data-language="{{ profile.language }}" data-formula="{{ profile.formula|e }}" data-provider="{{ profile.provider|default('kokoro')|lower }}" {% if profile_value == profile.name %}selected{% endif %}>{{ profile.name }}{% if profile.language %} ({{ profile.language|upper }}){% endif %}{% if profile.provider and profile.provider|lower != 'kokoro' %} · {{ profile.provider|capitalize }}{% endif %}</option>
+                <option value="{{ profile.name }}" data-language="{{ profile.language }}" data-formula="{{ profile.formula|e }}" {% if profile_value == profile.name %}selected{% endif %}>{{ profile.name }}{% if profile.language %} ({{ profile.language|upper }}){% endif %}</option>
                 {% endfor %}
               </optgroup>
               {% endif %}
             </select>
           </div>
-          <div class="field" data-role="voice-field" data-provider="kokoro" {% if profile_value != '__standard' or is_supertonic %}hidden aria-hidden="true"{% endif %}>
+          <div class="field" data-role="voice-field" {% if profile_value != '__standard' %}hidden aria-hidden="true"{% endif %}>
             <label for="voice">Voice</label>
             <select id="voice" name="voice" data-role="voice-select" data-default="{{ narrator_voice or settings_dict.get('default_voice', '') }}" {{ 'disabled' if readonly else '' }}>
               {% for voice in options.voices %}
@@ -321,23 +294,10 @@
               {% endfor %}
             </select>
           </div>
-          <div class="field" data-role="voice-field" data-provider="supertonic" {% if profile_value != '__standard' or not is_supertonic %}hidden aria-hidden="true"{% endif %}>
-            <label for="voice_st">Supertonic voice</label>
-            <select id="voice_st" name="voice" data-role="voice-select" data-default="{{ narrator_voice or 'M1' }}" {{ 'disabled' if readonly else '' }}>
-              {% for voice in ['M1','M2','M3','M4','M5','F1','F2','F3','F4','F5'] %}
-              <option value="{{ voice }}" {% if narrator_voice == voice and profile_value == '__standard' %}selected{% endif %}>{{ voice }}</option>
-              {% endfor %}
-            </select>
-          </div>
-          <div class="field" data-conditional="formula" data-role="formula-field" data-provider="kokoro" {% if profile_value != '__formula' or is_supertonic %}hidden aria-hidden="true"{% endif %}>
+          <div class="field" data-conditional="formula" data-role="formula-field" {% if profile_value != '__formula' %}hidden aria-hidden="true"{% endif %}>
             <label for="voice_formula">Custom voice formula</label>
             <input type="text" id="voice_formula" name="voice_formula" placeholder="af_nova*0.4+am_liam*0.6" data-role="voice-formula" value="{{ voice_formula_value }}" {{ 'disabled' if readonly else '' }}>
           </div>
-          <div class="field" data-role="supertonic-steps-field" {% if not is_supertonic %}hidden aria-hidden="true"{% endif %}>
-            <label for="supertonic_total_steps">Supertonic quality (total steps)</label>
-            <input type="number" id="supertonic_total_steps" name="supertonic_total_steps" min="2" max="15" value="{{ supertonic_steps_value }}" {{ 'disabled' if readonly else '' }}>
-            <p class="hint">2 = fastest/lowest quality, 15 = slowest/highest quality.</p>
-          </div>
         </div>
         <div class="form-section__group">
           <div class="field field--slider">
@@ -463,54 +423,3 @@
     </div>
   </footer>
 </form>
-
-<script nonce="{{ csp_nonce() if csp_nonce else '' }}">
-(function() {
-  const form = document.querySelector('[data-wizard-form="true"][data-step="book"]');
-  if (!form) return;
-  const providerSelect = form.querySelector('[data-role="tts-provider"]');
-  if (!providerSelect) return;
-
-  function filterProfilesByProvider(provider) {
-    const profileSelect = form.querySelector('[data-role="voice-profile"]');
-    if (!profileSelect) return;
-    const options = profileSelect.querySelectorAll('option[data-provider]');
-    options.forEach(function(opt) {
-      const matches = !opt.dataset.provider || opt.dataset.provider === provider;
-      opt.hidden = !matches;
-      if (opt.selected && opt.hidden) {
-        opt.selected = false;
-      }
-    });
-    if (!profileSelect.value || profileSelect.selectedOptions[0]?.hidden) {
-      const firstVisible = profileSelect.querySelector('option:not([hidden])');
-      if (firstVisible) profileSelect.value = firstVisible.value;
-    }
-    profileSelect.dispatchEvent(new Event('change', { bubbles: true }));
-  }
-
-  function syncProviderUI(provider) {
-    var isSupertonic = provider === 'supertonic';
-    form.querySelectorAll('[data-role="voice-field"]').forEach(function(el) {
-      el.hidden = el.dataset.provider !== provider;
-      el.setAttribute('aria-hidden', el.hidden ? 'true' : 'false');
-    });
-    var formulaField = form.querySelector('[data-role="formula-field"]');
-    if (formulaField) {
-      formulaField.hidden = isSupertonic;
-      formulaField.setAttribute('aria-hidden', isSupertonic ? 'true' : 'false');
-    }
-    var stepsField = form.querySelector('[data-role="supertonic-steps-field"]');
-    if (stepsField) {
-      stepsField.hidden = !isSupertonic;
-      stepsField.setAttribute('aria-hidden', isSupertonic ? 'false' : 'true');
-    }
-    filterProfilesByProvider(provider);
-  }
-
-  providerSelect.addEventListener('change', function() {
-    syncProviderUI(providerSelect.value);
-  });
-  syncProviderUI(providerSelect.value);
-})();
-</script>

abogen/webui/templates/settings.html

@@ -61,15 +61,6 @@
               <p class="hint">Pick a saved speaker from Speaker Studio to use by default for new jobs.</p>
             </div>
 
-            <div class="field">
-              <label for="tts_provider">Default TTS Engine</label>
-              <select id="tts_provider" name="tts_provider">
-                <option value="kokoro" {% if settings.tts_provider == 'kokoro' %}selected{% endif %}>Kokoro</option>
-                <option value="supertonic" {% if settings.tts_provider == 'supertonic' %}selected{% endif %}>Supertonic</option>
-              </select>
-              <p class="hint">Select the default TTS engine for new jobs.</p>
-            </div>
-
             <div class="field field--wide">
               <p class="tag">Kokoro settings</p>
             </div>

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 []