ci: add uv cache prune to optimize cache size

- Add system dependency step for libegl1 to fix PyQt6 import on headless CI
- Normalize CRLF to LF in epub exporter whitespace test for Windows CI

.github/workflows/test_pip.yml

@@ -1,5 +1,6 @@
-name: pip install
-run-name: pip install
+name: CI
+run-name: CI
+
 on:
   push:
     branches: [main]
@@ -12,8 +13,9 @@ on:
       - 'pyproject.toml'
       - '.github/workflows/**'
   workflow_dispatch:
+
 jobs:
-  install-and-run:
+  test:
     strategy:
       matrix:
         os: [ubuntu-latest, macos-14, windows-latest]
@@ -23,12 +25,30 @@ jobs:
     steps:
       - name: Checkout repository
         uses: actions/checkout@v7
+
       - name: Set up Python
         uses: actions/setup-python@v6
         with:
           python-version: ${{ matrix.python-version }}
-          cache: pip
-      - 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
+          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]
+
+      - name: Run tests
+        env:
+          QT_QPA_PLATFORM: offscreen
+        run: pytest tests/ -v --tb=short
+
+      - name: Minimize uv cache
+        if: always()
+        run: uv cache prune --ci

abogen/constants.py

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

abogen/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_backend_registry import get_metadata
 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_metadata("kokoro").voices
         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_metadata("kokoro").voices:
                 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_metadata("kokoro").voices)
         return (len(missing) == 0), missing
 
     def _check_kokoro_model(self) -> bool:

abogen/pyqt/gui.py

@@ -82,11 +82,11 @@ from abogen.constants import (
     GITHUB_URL,
     PROGRAM_DESCRIPTION,
     LANGUAGE_DESCRIPTIONS,
-    VOICES_INTERNAL,
     SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION,
     COLORS,
     SUBTITLE_FORMATS,
 )
+from abogen.tts_backend_registry import get_metadata
 import threading
 from abogen.pyqt.voice_formula_gui import VoiceFormulaDialog
 from abogen.voice_profiles import load_profiles
@@ -1873,7 +1873,7 @@ class abogen(QWidget):
         for pname in load_profiles().keys():
             self.voice_combo.addItem(profile_icon, pname, f"profile:{pname}")
         # re-add voices
-        for v in VOICES_INTERNAL:
+        for v in get_metadata("kokoro").voices:
             icon = QIcon()
             flag_path = get_resource_path("abogen.assets.flags", f"{v[0]}.png")
             if flag_path and os.path.exists(flag_path):

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_backend_registry import get_metadata
 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_metadata("kokoro").voices
         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_metadata("kokoro").voices:
                 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_metadata("kokoro").voices)
         return (len(missing) == 0), missing
 
     def _check_kokoro_model(self) -> bool:

abogen/pyqt/voice_formula_gui.py

@@ -28,11 +28,11 @@ from PyQt6.QtWidgets import (
 from PyQt6.QtCore import Qt, QTimer, QPoint, QRect, QSize
 from PyQt6.QtGui import QPixmap, QIcon, QAction
 from abogen.constants import (
-    VOICES_INTERNAL,
     SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION,
     LANGUAGE_DESCRIPTIONS,
     COLORS,
 )
+from abogen.tts_backend_registry import get_metadata
 import re
 import platform
 from abogen.utils import get_resource_path
@@ -179,7 +179,7 @@ class VoiceMixer(QWidget):
         layout.addWidget(QLabel(name), alignment=Qt.AlignmentFlag.AlignCenter)
 
         # Voice name label with gender icon
-        is_female = self.voice_name in VOICES_INTERNAL and self.voice_name[1] == "f"
+        is_female = self.voice_name in get_metadata("kokoro").voices and self.voice_name[1] == "f"
 
         # Icons layout (flag and gender)
         icons_layout = QHBoxLayout()
@@ -772,7 +772,7 @@ class VoiceFormulaDialog(QDialog):
 
     def add_voices(self, initial_state):
         first_enabled_voice = None
-        for voice in VOICES_INTERNAL:
+        for voice in get_metadata("kokoro").voices:
             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/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_backend_registry import get_metadata
 
     # 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_metadata("kokoro").voices}
     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_backend_registry import get_metadata
 
     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_metadata("kokoro").voices 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_metadata("kokoro").voices if v.lower() == voice_name_lower),
                     voice_name
                 )
             valid_markers += 1

abogen/tts_backend_registry.py

@@ -30,6 +30,10 @@ class TTSBackendRegistry:
         self._backends[metadata.id] = metadata
         self._factories[metadata.id] = factory
 
+    def is_registered(self, backend_id: str) -> bool:
+        """Return True if a backend with the given id is registered."""
+        return backend_id in self._backends
+
     def list_backends(self) -> list[TTSBackendMetadata]:
         """Return metadata for all registered backends."""
         return list(self._backends.values())
@@ -54,6 +58,33 @@ class TTSBackendRegistry:
             raise KeyError(f"Unknown backend: {backend_id}")
         return self._factories[backend_id](**kwargs)
 
+    def resolve_backend_for_voice(
+        self,
+        spec: str,
+        fallback: str = "kokoro",
+    ) -> str:
+        """Determine which backend owns the given voice specification.
+
+        Resolution rules:
+        1. Empty spec -> fallback
+        2. Kokoro formula (contains '*' or '+') -> "kokoro"
+        3. Exact voice ID match against registered backends -> backend 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()
+        for metadata in self._backends.values():
+            if upper in metadata.voices:
+                return metadata.id
+
+        return fallback
+
 
 _registry = TTSBackendRegistry()
 
@@ -79,6 +110,37 @@ def get_metadata(backend_id: str) -> TTSBackendMetadata:
     return _registry.get_metadata(backend_id)
 
 
+def get_default_voice(backend_id: str, fallback: str = "") -> str:
+    """Return the first voice of a backend, or *fallback* if none."""
+    voices = get_metadata(backend_id).voices
+    return voices[0] if voices else fallback
+
+
 def create_backend(backend_id: str, **kwargs: Any) -> TTSBackend:
     """Create a TTS backend instance by provider id."""
     return _registry.create_backend(backend_id, **kwargs)
+
+
+def is_registered_backend(backend_id: str) -> bool:
+    """Return True if *backend_id* is a registered TTS backend."""
+    import abogen.tts_backends  # noqa: F401  — triggers backend registration
+    return _registry.is_registered(backend_id)
+
+
+def resolve_backend_for_voice(
+    spec: str,
+    fallback: str = "kokoro",
+) -> str:
+    """Determine which backend owns the given voice specification.
+
+    Ensures all backends are registered by importing the tts_backends
+    package on first access.
+
+    Resolution rules:
+    1. Empty spec -> fallback
+    2. Kokoro formula (contains '*' or '+') -> "kokoro"
+    3. Exact voice ID match against registered backends -> backend id
+    4. Unknown voice -> fallback
+    """
+    import abogen.tts_backends  # noqa: F401  — triggers backend registration
+    return _registry.resolve_backend_for_voice(spec, fallback=fallback)

abogen/tts_backends/kokoro.py

@@ -10,14 +10,72 @@ from typing import Any, Dict, Iterator, List, Optional
 
 import numpy as np
 
-from abogen.constants import VOICES_INTERNAL
 from abogen.tts_backend import TTSBackendMetadata
 
+# Internal voice list — source of truth for Kokoro voices.
+# The rest of the project accesses voices via get_metadata("kokoro").voices.
+_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",
+]
+
 _KOKORO_METADATA = TTSBackendMetadata(
     id="kokoro",
     name="Kokoro",
     description="Kokoro TTS engine",
-    voices=tuple(VOICES_INTERNAL),
+    voices=tuple(_VOICES_INTERNAL),
 )
 
 

abogen/voice_profiles.py

@@ -2,7 +2,7 @@ import json
 import os
 from typing import Any, Dict, Iterable, List, Tuple
 
-from abogen.tts_backend_registry import get_metadata
+from abogen.tts_backend_registry import get_metadata, is_registered_backend
 from abogen.utils import get_user_config_path
 
 
@@ -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_registered_backend(provider):
         provider = "kokoro"
 
     language = str(entry.get("language") or "a").strip().lower() or "a"

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 ./
-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/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_backend_registry import get_metadata, is_registered_backend, resolve_backend_for_voice
 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 (
@@ -47,7 +47,7 @@ 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_backends.supertonic import DEFAULT_SUPERTONIC_VOICES
+
 
 from .service import Job, JobStatus
 
@@ -57,26 +57,27 @@ SAMPLE_RATE = 24000
 
 
 def _supertonic_voice_from_spec(spec: Any, fallback: str) -> str:
+    """Normalize a voice specification for Supertonic.
+
+    This function only performs Supertonic-specific normalization (uppercase conversion
+    and fallback handling). Backend resolution is handled by the registry.
+    """
     raw = str(spec or "").strip()
     fallback_raw = str(fallback or "").strip()
 
-    # SuperTonic voices are discrete IDs (M1/F3/...). If we see a Kokoro mix
-    # formula (contains '*' or '+'), ignore it and fall back to a safe voice.
-    if not raw or "*" in raw or "+" in raw:
-        raw = fallback_raw
-    if not raw or "*" in raw or "+" in raw:
-        raw = "M1"
+    # Normalize to uppercase for Supertonic voice IDs
+    upper = raw.upper() if raw else ""
+
+    # If empty or contains formula characters, use fallback
+    if not upper or "*" in upper or "+" in upper:
+        upper = fallback_raw.upper() if fallback_raw else ""
+
+    # If still empty, use default Supertonic voice
+    if not upper or "*" in upper or "+" in upper:
+        upper = "M1"
 
-    upper = raw.upper()
-    if upper in DEFAULT_SUPERTONIC_VOICES:
     return upper
 
-    fallback_upper = fallback_raw.upper() if fallback_raw else ""
-    if fallback_upper in DEFAULT_SUPERTONIC_VOICES:
-        return fallback_upper
-
-    return "M1"
-
 
 def _split_speaker_reference(value: Any) -> tuple[Optional[str], str]:
     raw = str(value or "").strip()
@@ -119,15 +120,7 @@ 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_backend_for_voice(str(value or ""), fallback=fallback)
 
 
 class _JobCancelled(Exception):
@@ -576,7 +569,7 @@ 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_metadata("kokoro").voices:
         return {text}
     return set()
 
@@ -640,7 +633,7 @@ 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_metadata("kokoro").voices)
     return voices
 
 
@@ -1574,7 +1567,7 @@ 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_registered_backend(provider_norm):
                 provider_norm = "kokoro"
 
             existing = pipelines.get(provider_norm)
@@ -1803,8 +1796,8 @@ 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(
@@ -2237,8 +2230,8 @@ 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

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_backend_registry import is_registered_backend
 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_registered_backend(provider):
             provider = "kokoro"
         if provider == "supertonic":
             profile = {
@@ -230,7 +231,7 @@ def api_speaker_preview() -> ResponseReturnValue:
     use_gpu = settings.get("use_gpu", False)
 
     base_spec, speaker_name = split_profile_spec(voice)
-    resolved_provider = tts_provider if tts_provider in {"kokoro", "supertonic"} else ""
+    resolved_provider = tts_provider if is_registered_backend(tts_provider) else ""
 
     if speaker_name:
         entry = normalize_profile_entry(load_profiles().get(speaker_name))

abogen/webui/routes/utils/form.py

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

abogen/webui/routes/utils/settings.py

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

abogen/webui/routes/utils/voice.py

@@ -17,8 +17,8 @@ from abogen.constants import (
     SUPPORTED_SOUND_FORMATS,
     SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION,
     SAMPLE_VOICE_TEXTS,
-    VOICES_INTERNAL,
 )
+from abogen.tts_backend_registry import get_metadata
 from abogen.speaker_configs import list_configs
 from abogen.tts_backend_registry import create_backend
 from abogen.webui.conversion_runner import _select_device, _to_float32, SAMPLE_RATE, SPLIT_PATTERN
@@ -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_metadata("kokoro").voices:
         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_metadata("kokoro").voices,
         "subtitle_formats": SUBTITLE_FORMATS,
         "supported_langs_for_subs": SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION,
         "output_formats": SUPPORTED_SOUND_FORMATS,

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__)
 

tests/test_conversion_voice_resolution.py

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

tests/test_epub_exporter.py

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

tests/test_tts_backend.py

@@ -202,3 +202,113 @@ class TestBackendRegistration:
         registry_meta = _registry.get_metadata("supertonic")
         assert _SUPERTONIC_METADATA is registry_meta
         assert _SUPERTONIC_METADATA.voices == registry_meta.voices
+
+
+class TestResolveBackendForVoice:
+    """Tests for the resolve_backend_for_voice method."""
+
+    def test_empty_spec_returns_fallback(self):
+        registry = TTSBackendRegistry()
+        assert registry.resolve_backend_for_voice("", fallback="kokoro") == "kokoro"
+        assert registry.resolve_backend_for_voice("", fallback="supertonic") == "supertonic"
+
+    def test_none_spec_returns_fallback(self):
+        registry = TTSBackendRegistry()
+        assert registry.resolve_backend_for_voice(None, fallback="kokoro") == "kokoro"
+
+    def test_kokoro_formula_with_star_returns_kokoro(self):
+        registry = TTSBackendRegistry()
+        assert registry.resolve_backend_for_voice("af_nova*0.7") == "kokoro"
+
+    def test_kokoro_formula_with_plus_returns_kokoro(self):
+        registry = TTSBackendRegistry()
+        assert registry.resolve_backend_for_voice("af_nova*0.7+am_liam*0.3") == "kokoro"
+
+    def test_kokoro_voice_id_resolves_to_kokoro(self):
+        registry = TTSBackendRegistry()
+        meta = TTSBackendMetadata(
+            id="kokoro",
+            name="Kokoro",
+            description="Kokoro TTS",
+            voices=("af_nova", "am_liam"),
+        )
+        registry.register(metadata=meta, factory=lambda: None)
+
+        assert registry.resolve_backend_for_voice("af_nova") == "kokoro"
+        assert registry.resolve_backend_for_voice("am_liam") == "kokoro"
+
+    def test_supertonic_voice_id_resolves_to_supertonic(self):
+        registry = TTSBackendRegistry()
+        meta = TTSBackendMetadata(
+            id="supertonic",
+            name="SuperTonic",
+            description="SuperTonic TTS",
+            voices=("M1", "M2", "F1", "F2"),
+        )
+        registry.register(metadata=meta, factory=lambda: None)
+
+        assert registry.resolve_backend_for_voice("M1") == "supertonic"
+        assert registry.resolve_backend_for_voice("F2") == "supertonic"
+
+    def test_unknown_voice_returns_fallback(self):
+        registry = TTSBackendRegistry()
+        meta = TTSBackendMetadata(
+            id="kokoro",
+            name="Kokoro",
+            description="Kokoro TTS",
+            voices=("af_nova",),
+        )
+        registry.register(metadata=meta, factory=lambda: None)
+
+        assert registry.resolve_backend_for_voice("unknown_voice") == "kokoro"
+        assert registry.resolve_backend_for_voice("unknown_voice", fallback="supertonic") == "supertonic"
+
+    def test_case_insensitive_matching(self):
+        registry = TTSBackendRegistry()
+        meta = TTSBackendMetadata(
+            id="supertonic",
+            name="SuperTonic",
+            description="SuperTonic TTS",
+            voices=("M1", "F1"),
+        )
+        registry.register(metadata=meta, factory=lambda: None)
+
+        assert registry.resolve_backend_for_voice("m1") == "supertonic"
+        assert registry.resolve_backend_for_voice("f1") == "supertonic"
+
+    def test_default_fallback_is_kokoro(self):
+        registry = TTSBackendRegistry()
+        assert registry.resolve_backend_for_voice("unknown") == "kokoro"
+
+    def test_multiple_backends_resolution(self):
+        registry = TTSBackendRegistry()
+        kokoro_meta = TTSBackendMetadata(
+            id="kokoro",
+            name="Kokoro",
+            description="Kokoro TTS",
+            voices=("af_nova",),
+        )
+        supertonic_meta = TTSBackendMetadata(
+            id="supertonic",
+            name="SuperTonic",
+            description="SuperTonic TTS",
+            voices=("M1",),
+        )
+        registry.register(metadata=kokoro_meta, factory=lambda: None)
+        registry.register(metadata=supertonic_meta, factory=lambda: None)
+
+        assert registry.resolve_backend_for_voice("af_nova") == "kokoro"
+        assert registry.resolve_backend_for_voice("M1") == "supertonic"
+
+    def test_global_wrapper_resolve_backend_for_voice(self):
+        from abogen.tts_backend_registry import resolve_backend_for_voice
+
+        # Test with empty spec
+        assert resolve_backend_for_voice("") == "kokoro"
+
+        # Test with formula
+        assert resolve_backend_for_voice("af_nova*0.7") == "kokoro"
+
+        # Test with a registered voice
+        assert resolve_backend_for_voice("af_nova") == "kokoro"
+        assert resolve_backend_for_voice("M1") == "supertonic"

tests/test_voice_cache.py

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