refactor(webui): replace direct VOICES_INTERNAL/DEFAULT_SUPERTONIC_VOICES with get_metadata API

- Add get_default_voice() helper to tts_backend_registry
- Replace all VOICES_INTERNAL imports in WebUI with get_metadata().voices
- Replace all DEFAULT_SUPERTONIC_VOICES imports in conversion_runner with get_metadata().voices
- Remove unused VOICES_INTERNAL import from voices.py

Core modules (voice_profiles, voice_formulas, voice_cache) already used
get_metadata(). This completes the WebUI migration.

.github/workflows/test_pip.yml

@@ -2,6 +2,7 @@ name: pip install
 run-name: pip install
 on:
   push:
+    branches: [main]
     paths:
       - '**.py'
       - 'pyproject.toml'
@@ -15,18 +16,18 @@ jobs:
   install-and-run:
     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 }}
+          cache: pip
       - name: Install from repository
         run: python -m pip install .
       #- name: Run abogen

.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

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,
@@ -259,8 +260,7 @@ class ConversionThread(QThread):
         output_folder,
         subtitle_mode,
         output_format,
-        np_module,
-        kpipeline_class,
+        backend,
         start_time,
         total_char_count,
         use_gpu=True,
@@ -270,8 +270,7 @@ class ConversionThread(QThread):
         super().__init__()
         self._chapter_options_event = threading.Event()
         self._timestamp_response_event = threading.Event()
-        self.np = np_module
-        self.KPipeline = kpipeline_class
+        self.backend = backend
         self.file_name = file_name
         self.lang_code = lang_code
         self.speed = speed
@@ -490,19 +489,6 @@ class ConversionThread(QThread):
 
             self.log_updated.emit(("\nInitializing TTS pipeline...", "grey"))
 
-            # Set device based on use_gpu setting and platform
-            if self.use_gpu:
-                if platform.system() == "Darwin" and platform.processor() == "arm":
-                    device = "mps"  # Use MPS for Apple Silicon
-                else:
-                    device = "cuda"  # Use CUDA for other platforms
-            else:
-                device = "cpu"
-
-            tts = self.KPipeline(
-                lang_code=self.lang_code, repo_id="hexgrad/Kokoro-82M", device=device
-            )
-
             # Check if the input is a subtitle file or timestamp text file
             is_subtitle_file = False
             is_timestamp_text = False
@@ -538,7 +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:
@@ -1071,7 +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"))
@@ -1080,7 +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
@@ -1166,7 +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,
@@ -1368,7 +1354,7 @@ class ConversionThread(QThread):
                     silence_samples = int(
                         self.silence_duration * 24000
                     )  # Silence duration at 24,000 Hz
-                    silence_audio = self.np.zeros(silence_samples, dtype="float32")
+                    silence_audio = np.zeros(silence_samples, dtype="float32")
                     silence_bytes = silence_audio.tobytes()
 
                     if merged_out_file:
@@ -1707,7 +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"
             )
 
@@ -1771,7 +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,
@@ -1789,11 +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"
                     )
                 )
@@ -1827,8 +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)
                                 )
                             ),
                         )
@@ -1861,7 +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",
                         )
@@ -1875,7 +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,
@@ -1886,14 +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"
                             )
                         )
@@ -1910,10 +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"
                             ),
                         ]
@@ -1926,10 +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"
                             ),
                         ]
@@ -1971,7 +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})"
@@ -2440,8 +2426,7 @@ class VoicePreviewThread(QThread):
 
     def __init__(
         self,
-        np_module,
-        kpipeline_class,
+        backend,
         lang_code,
         voice,
         speed,
@@ -2449,8 +2434,7 @@ class VoicePreviewThread(QThread):
         parent=None,
     ):
         super().__init__(parent)
-        self.np_module = np_module
-        self.kpipeline_class = kpipeline_class
+        self.backend = backend
         self.lang_code = lang_code
         self.voice = voice
         self.speed = speed
@@ -2484,31 +2468,19 @@ class VoicePreviewThread(QThread):
         # Generate the preview and save to cache
         try:
 
-            # Set device based on use_gpu setting and platform
-            if self.use_gpu:
-                if platform.system() == "Darwin" and platform.processor() == "arm":
-                    device = "mps"  # Use MPS for Apple Silicon
-                else:
-                    device = "cuda"  # Use CUDA for other platforms
-            else:
-                device = "cpu"
-
-            tts = self.kpipeline_class(
-                lang_code=self.lang_code, repo_id="hexgrad/Kokoro-82M", device=device
-            )
             # Enable voice formula support for preview
             if "*" in self.voice:
-                loaded_voice = get_new_voice(tts, self.voice, self.use_gpu)
+                loaded_voice = get_new_voice(self.backend, self.voice, self.use_gpu)
             else:
                 loaded_voice = self.voice
             sample_text = get_sample_voice_text(self.lang_code)
             audio_segments = []
-            for result in tts(
+            for result in self.backend(
                 sample_text, voice=loaded_voice, speed=self.speed, split_pattern=None
             ):
                 audio_segments.append(result.audio)
             if audio_segments:
-                audio = self.np_module.concatenate(audio_segments)
+                audio = np.concatenate(audio_segments)
                 # Save directly to the cache path
                 sf.write(self.cache_path, audio, 24000)
                 self.temp_wav = self.cache_path

abogen/pyqt/gui.py

@@ -2316,9 +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
 
@@ -2341,8 +2341,7 @@ class abogen(QWidget):
                 self.selected_output_folder,
                 subtitle_mode=actual_subtitle_mode,
                 output_format=self.selected_format,
-                np_module=np_module,
-                kpipeline_class=kpipeline_class,
+                backend=backend,
                 start_time=self.start_time,
                 total_char_count=self.char_count,
                 use_gpu=self.gpu_ok,
@@ -2426,7 +2425,20 @@ class abogen(QWidget):
             self.gpu_ok = gpu_ok
             self.update_log((gpu_msg, gpu_ok))
             self.update_log("Loading modules...")
-            load_thread = LoadPipelineThread(pipeline_loaded_callback)
+
+            # Determine device based on GPU availability
+            if gpu_ok:
+                if platform.system() == "Darwin" and platform.processor() == "arm":
+                    device = "mps"
+                else:
+                    device = "cuda"
+            else:
+                device = "cpu"
+
+            lang_code = self.selected_lang or "a"
+            load_thread = LoadPipelineThread(
+                pipeline_loaded_callback, lang_code=lang_code, device=device
+            )
             load_thread.start()
 
         threading.Thread(target=gpu_and_load, daemon=True).start()
@@ -2863,18 +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)
@@ -2912,7 +2933,7 @@ class abogen(QWidget):
         gpu_msg, gpu_ok = get_gpu_acceleration(self.use_gpu)
 
         self.preview_thread = VoicePreviewThread(
-            np_module, kpipeline_class, lang, voice, speed, gpu_ok
+            backend, lang, voice, speed, gpu_ok
         )
         self.preview_thread.finished.connect(self._play_preview_audio)
         self.preview_thread.error.connect(self._preview_error)

abogen/tts_backend.py

@@ -0,0 +1,89 @@
+"""
+TTS Backend Interface
+
+This module defines the protocol for TTS backends and the
+metadata model that describes a backend implementation.
+"""
+
+from dataclasses import dataclass
+from typing import Protocol, List, Dict, Any
+
+
+@dataclass(frozen=True)
+class TTSBackendMetadata:
+    """
+    Immutable metadata describing a TTS backend implementation.
+
+    Attributes:
+        id: Unique backend identifier (e.g. ``"kokoro"``, ``"supertonic"``).
+        name: Human-readable display name.
+        description: Short description of the backend.
+        voices: Tuple of supported voice identifiers.
+    """
+
+    id: str
+    name: str
+    description: str
+    voices: tuple[str, ...] = ()
+
+
+class TTSBackend(Protocol):
+    """
+    Protocol for TTS backends.
+
+    All TTS backends must implement this interface to be compatible
+    with the application.
+    """
+
+    @property
+    def metadata(self) -> TTSBackendMetadata:
+        ...
+
+    def __init__(self, **kwargs) -> None:
+        """
+        Initialize the TTS backend.
+
+        Args:
+            **kwargs: Backend-specific configuration parameters
+        """
+        ...
+
+    def synthesize(self, text: str, **kwargs) -> bytes:
+        """
+        Synthesize speech from text.
+
+        Args:
+            text: Text to synthesize
+            **kwargs: Additional parameters for synthesis
+
+        Returns:
+            Audio data as bytes
+        """
+        ...
+
+    def get_available_voices(self) -> List[str]:
+        """
+        Get list of available voices.
+
+        Returns:
+            List of voice identifiers
+        """
+        ...
+
+    def get_supported_formats(self) -> List[str]:
+        """
+        Get list of supported audio formats.
+
+        Returns:
+            List of supported audio formats
+        """
+        ...
+
+    def get_info(self) -> Dict[str, Any]:
+        """
+        Get backend information.
+
+        Returns:
+            Dictionary with backend information
+        """
+        ...

abogen/tts_backend_registry.py

@@ -0,0 +1,90 @@
+"""
+TTS Backend Registry
+
+Provides a global registry for TTS backend factories.
+Backends register themselves with metadata and a factory callable.
+The registry is universal and does not know about backend constructors.
+"""
+
+from typing import Callable, Any
+
+from abogen.tts_backend import TTSBackend, TTSBackendMetadata
+
+
+class TTSBackendRegistry:
+    """Registry of TTS backend factories.
+
+    Stores metadata and factory callables for registered backends.
+    """
+
+    def __init__(self) -> None:
+        self._backends: dict[str, TTSBackendMetadata] = {}
+        self._factories: dict[str, Callable[..., TTSBackend]] = {}
+
+    def register(
+        self,
+        metadata: TTSBackendMetadata,
+        factory: Callable[..., TTSBackend],
+    ) -> None:
+        """Register a backend with its metadata and factory callable."""
+        self._backends[metadata.id] = metadata
+        self._factories[metadata.id] = factory
+
+    def list_backends(self) -> list[TTSBackendMetadata]:
+        """Return metadata for all registered backends."""
+        return list(self._backends.values())
+
+    def get_metadata(self, backend_id: str) -> TTSBackendMetadata:
+        """Get metadata for a specific backend.
+
+        Raises:
+            KeyError: If backend with given id is not registered.
+        """
+        if backend_id not in self._backends:
+            raise KeyError(f"Unknown backend: {backend_id}")
+        return self._backends[backend_id]
+
+    def create_backend(self, backend_id: str, **kwargs: Any) -> TTSBackend:
+        """Create a backend instance by id.
+
+        Raises:
+            KeyError: If backend with given id is not registered.
+        """
+        if backend_id not in self._factories:
+            raise KeyError(f"Unknown backend: {backend_id}")
+        return self._factories[backend_id](**kwargs)
+
+
+_registry = TTSBackendRegistry()
+
+
+def register_backend(
+    metadata: TTSBackendMetadata,
+    factory: Callable[..., TTSBackend],
+) -> None:
+    """Register a TTS backend in the global registry."""
+    _registry.register(metadata, factory)
+
+
+def get_metadata(backend_id: str) -> TTSBackendMetadata:
+    """Get metadata for a specific backend by id.
+
+    Ensures all backends are registered by importing the tts_backends
+    package on first access.
+
+    Raises:
+        KeyError: If backend with given id is not registered.
+    """
+    import abogen.tts_backends  # noqa: F401  — triggers backend registration
+    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)

abogen/tts_backends/__init__.py

@@ -0,0 +1,20 @@
+"""TTS backends package.
+
+Backend modules are auto-discovered and imported here.
+Each backend module registers itself with the global registry
+when imported.
+"""
+
+import importlib
+import pkgutil
+
+
+def _discover_backends():
+    """Import all modules in this package to trigger their registration."""
+    package = __name__
+    for _importer, modname, _ispkg in pkgutil.iter_modules(path=__path__):
+        importlib.import_module(f"{package}.{modname}")
+
+
+_discover_backends()
+

abogen/tts_backends/kokoro.py

@@ -0,0 +1,121 @@
+"""
+Kokoro TTS Backend
+
+Encapsulates the Kokoro KPipeline as a TTSBackend implementation.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Dict, Iterator, List, Optional
+
+import numpy as np
+
+from abogen.constants import VOICES_INTERNAL
+from abogen.tts_backend import TTSBackendMetadata
+
+_KOKORO_METADATA = TTSBackendMetadata(
+    id="kokoro",
+    name="Kokoro",
+    description="Kokoro TTS engine",
+    voices=tuple(VOICES_INTERNAL),
+)
+
+
+def _load_kpipeline():
+    """Lazy-load Kokoro dependencies."""
+    from kokoro import KPipeline  # type: ignore[import-not-found]
+
+    return KPipeline
+
+
+class KokoroBackend:
+    """TTSBackend implementation wrapping the Kokoro KPipeline.
+
+    All interaction with KPipeline is encapsulated here.
+    The rest of the project depends only on this class.
+    """
+
+    def __init__(self, **kwargs: Any) -> None:
+        lang_code = kwargs["lang_code"]
+        repo_id = kwargs.get("repo_id", "hexgrad/Kokoro-82M")
+        device = kwargs.get("device", "cpu")
+
+        KPipeline = _load_kpipeline()
+        self._pipeline = KPipeline(
+            lang_code=lang_code,
+            repo_id=repo_id,
+            device=device,
+        )
+        self._lang_code = lang_code
+
+    @property
+    def metadata(self) -> TTSBackendMetadata:
+        return _KOKORO_METADATA
+
+    def __call__(
+        self,
+        text: str,
+        *,
+        voice: Any,
+        speed: float = 1.0,
+        split_pattern: Optional[str] = None,
+    ) -> Iterator[Any]:
+        """Delegate to KPipeline's __call__."""
+        return self._pipeline(
+            text,
+            voice=voice,
+            speed=speed,
+            split_pattern=split_pattern,
+        )
+
+    def load_single_voice(self, voice_name: str) -> Any:
+        """Load a single voice tensor. Used by voice formula system."""
+        return self._pipeline.load_single_voice(voice_name)
+
+    def synthesize(self, text: str, **kwargs: Any) -> bytes:
+        """Synthesize speech from text. Returns raw audio bytes."""
+        voice = kwargs.get("voice", "")
+        speed = kwargs.get("speed", 1.0)
+        split_pattern = kwargs.get("split_pattern", None)
+
+        audio_parts: list[np.ndarray] = []
+        for segment in self(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 b""
+
+        combined = np.concatenate(audio_parts).astype("float32", copy=False)
+        return combined.tobytes()
+
+    def get_available_voices(self) -> List[str]:
+        """Return known Kokoro voice identifiers."""
+        return list(self.metadata.voices)
+
+    def get_supported_formats(self) -> List[str]:
+        """Kokoro outputs raw PCM float32 audio."""
+        return ["pcm_float32"]
+
+    def get_info(self) -> Dict[str, Any]:
+        return {
+            "id": "kokoro",
+            "name": "Kokoro",
+            "lang_code": self._lang_code,
+        }
+
+
+def create_kokoro_backend(**kwargs: Any) -> KokoroBackend:
+    """Factory callable registered with TTSBackendRegistry."""
+    return KokoroBackend(**kwargs)
+
+
+# --- Registration ---
+from abogen.tts_backend_registry import register_backend  # noqa: E402
+
+register_backend(
+    metadata=_KOKORO_METADATA,
+    factory=create_kokoro_backend,
+)

abogen/tts_backends/supertonic.py

@@ -5,7 +5,7 @@ from dataclasses import dataclass
 import logging
 import math
 import re
-from typing import Any, Iterable, Iterator, Optional
+from typing import Any, Dict, Iterable, Iterator, List, Optional
 
 import numpy as np
 
@@ -15,6 +15,15 @@ logger = logging.getLogger(__name__)
 
 DEFAULT_SUPERTONIC_VOICES = ("M1", "M2", "M3", "M4", "M5", "F1", "F2", "F3", "F4", "F5")
 
+from abogen.tts_backend import TTSBackendMetadata
+
+_SUPERTONIC_METADATA = TTSBackendMetadata(
+    id="supertonic",
+    name="SuperTonic",
+    description="SuperTonic TTS engine",
+    voices=DEFAULT_SUPERTONIC_VOICES,
+)
+
 
 @dataclass
 class SupertonicSegment:
@@ -273,3 +282,111 @@ class SupertonicPipeline:
                 audio = _resample_linear(audio, src_rate, self.sample_rate)
 
             yield SupertonicSegment(graphemes=chunk_to_speak, audio=audio)
+
+
+class SupertonicBackend:
+    """Supertonic TTS backend implementing the TTSBackend protocol.
+
+    Encapsulates ``SupertonicPipeline`` as an internal implementation detail.
+    """
+
+    @property
+    def metadata(self) -> TTSBackendMetadata:
+        return _SUPERTONIC_METADATA
+
+    def __init__(self, **kwargs: Any) -> None:
+        self._pipeline = SupertonicPipeline(
+            sample_rate=kwargs.get("sample_rate", 24000),
+            auto_download=kwargs.get("auto_download", True),
+            total_steps=kwargs.get("total_steps", 5),
+        )
+
+    def synthesize(self, text: str, **kwargs: Any) -> bytes:
+        """Synthesize speech and return raw audio bytes (WAV).
+
+        Delegates to the internal :class:`SupertonicPipeline` and concatenates
+        all produced segments into a single byte buffer.
+        """
+        import io
+
+        import soundfile as sf
+
+        voice = kwargs.get("voice", "M1")
+        speed = float(kwargs.get("speed", 1.0))
+        split_pattern = kwargs.get("split_pattern")
+        total_steps = kwargs.get("total_steps")
+
+        segments = self._pipeline(
+            text,
+            voice=voice,
+            speed=speed,
+            split_pattern=split_pattern,
+            total_steps=total_steps,
+        )
+
+        audio_parts: list[np.ndarray] = []
+        for seg in segments:
+            audio_parts.append(seg.audio)
+
+        if not audio_parts:
+            return b""
+
+        combined = np.concatenate(audio_parts)
+        buf = io.BytesIO()
+        sf.write(buf, combined, self._pipeline.sample_rate, format="WAV")
+        return buf.getvalue()
+
+    def get_available_voices(self) -> List[str]:
+        """Return the list of built-in SuperTonic voice identifiers."""
+        return list(self.metadata.voices)
+
+    def get_supported_formats(self) -> List[str]:
+        return ["wav"]
+
+    def get_info(self) -> Dict[str, Any]:
+        return {
+            "sample_rate": self._pipeline.sample_rate,
+            "total_steps": self._pipeline.total_steps,
+            "max_chunk_length": self._pipeline.max_chunk_length,
+            "voices": list(DEFAULT_SUPERTONIC_VOICES),
+        }
+
+    def __call__(
+        self,
+        text: str,
+        *,
+        voice: str,
+        speed: float,
+        split_pattern: Optional[str] = None,
+        total_steps: Optional[int] = None,
+    ) -> Iterator[SupertonicSegment]:
+        """Backward-compatible call interface, delegates to the pipeline."""
+        return self._pipeline(
+            text,
+            voice=voice,
+            speed=speed,
+            split_pattern=split_pattern,
+            total_steps=total_steps,
+        )
+
+
+def create_supertonic_backend(**kwargs: Any) -> SupertonicBackend:
+    """Create a SuperTonic TTS backend instance.
+
+    Args:
+        sample_rate: Audio sample rate. Defaults to 24000.
+        auto_download: Auto-download models. Defaults to True.
+        total_steps: Inference steps. Defaults to 5.
+
+    Returns:
+        SupertonicBackend instance.
+    """
+    return SupertonicBackend(**kwargs)
+
+
+from abogen.tts_backend_registry import register_backend  # noqa: E402
+
+register_backend(
+    metadata=_SUPERTONIC_METADATA,
+    factory=create_supertonic_backend,
+)

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_backend_registry import create_backend
+
+            backend = create_backend(
+                "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_backend_registry import get_metadata
 
 _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_metadata("kokoro").voices
     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
 

abogen/voice_formulas.py

@@ -1,7 +1,7 @@
 import re
 from typing import List, Tuple
 
-from abogen.constants import VOICES_INTERNAL
+from abogen.tts_backend_registry import get_metadata
 
 
 # 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_metadata("kokoro").voices
     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_backend_registry import get_metadata
 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_metadata("supertonic").voices
+    return raw if raw in supertonic_voices else "M1"
 
 
 def _coerce_supertonic_steps(value: Any) -> int:
@@ -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_metadata("kokoro").voices
     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

@@ -27,8 +27,6 @@ RUN python3 -m venv "$VIRTUAL_ENV"
 WORKDIR /app
 
 COPY pyproject.toml README.md ./
-COPY abogen ./abogen
-
 RUN pip install --upgrade pip \
   && if [ -n "$TORCH_VERSION" ]; then \
        pip install torch=="$TORCH_VERSION" torchvision=="$TORCH_VERSION" torchaudio=="$TORCH_VERSION" --index-url "$TORCH_INDEX_URL"; \
@@ -39,6 +37,8 @@ RUN pip install --upgrade pip \
     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"
 
+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 \

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
 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 +39,15 @@ from abogen.utils import (
     get_user_cache_path,
     get_user_output_path,
     load_config,
-    load_numpy_kpipeline,
 )
+from abogen.tts_backend_registry import create_backend
+from abogen.tts_backend import TTSBackend
 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
 
@@ -67,11 +68,11 @@ def _supertonic_voice_from_spec(spec: Any, fallback: str) -> str:
         raw = "M1"
 
     upper = raw.upper()
-    if upper in DEFAULT_SUPERTONIC_VOICES:
+    if upper in get_metadata("supertonic").voices:
         return upper
 
     fallback_upper = fallback_raw.upper() if fallback_raw else ""
-    if fallback_upper in DEFAULT_SUPERTONIC_VOICES:
+    if fallback_upper in get_metadata("supertonic").voices:
         return fallback_upper
 
     return "M1"
@@ -122,7 +123,7 @@ def _infer_provider_from_spec(value: Any, fallback: str = "kokoro") -> str:
     if not raw:
         return fallback
     upper = raw.upper()
-    if upper in DEFAULT_SUPERTONIC_VOICES:
+    if upper in get_metadata("supertonic").voices:
         return "supertonic"
     if "*" in raw or "+" in raw:
         return "kokoro"
@@ -575,7 +576,7 @@ 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()
 
@@ -639,7 +640,7 @@ 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
 
 
@@ -1581,7 +1582,8 @@ def run_conversion_job(job: Job) -> None:
                 return existing
 
             if provider_norm == "supertonic":
-                pipelines[provider_norm] = SupertonicPipeline(
+                pipelines[provider_norm] = create_backend(
+                    "supertonic",
                     sample_rate=SAMPLE_RATE,
                     auto_download=True,
                     total_steps=int(getattr(job, "supertonic_total_steps", 5) or 5),
@@ -1594,16 +1596,12 @@ 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 (conforms to TTSBackend protocol)
+            pipelines[provider_norm] = create_backend(
+                "kokoro",
+                lang_code=job.language,
+                device=device
+            )
             if not kokoro_cache_ready:
                 _initialize_voice_cache(job)
                 kokoro_cache_ready = True
@@ -1644,8 +1642,8 @@ 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 +1772,8 @@ 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 +1803,8 @@ def run_conversion_job(job: Job) -> None:
                     fallback_key = next(iter(voice_cache.keys()), "")
                     if fallback_key and fallback_key != "__custom_mix":
                         intro_voice_spec = fallback_key.split(":", 1)[-1]
-                if not intro_voice_spec and VOICES_INTERNAL:
-                    intro_voice_spec = VOICES_INTERNAL[0]
+                if not intro_voice_spec:
+                    intro_voice_spec = get_default_voice("kokoro")
 
                 if intro_voice_spec:
                     intro_provider, _, intro_voice_choice, intro_speed, intro_steps = resolve_voice_choice(
@@ -1860,8 +1858,8 @@ def run_conversion_job(job: Job) -> None:
                     total_steps=int(supertonic_steps_override if supertonic_steps_override is not None else getattr(job, "supertonic_total_steps", 5)),
                 )
             else:
-                kokoro_pipeline = get_pipeline("kokoro")
-                segment_iter = kokoro_pipeline(
+                kokoro_backend = get_pipeline("kokoro")
+                segment_iter = kokoro_backend(
                     normalized,
                     voice=voice_choice,
                     speed=float(speed_override if speed_override is not None else job.speed),
@@ -1950,8 +1948,8 @@ 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 +2093,9 @@ 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 +2237,8 @@ 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
@@ -2445,7 +2443,8 @@ 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(
+        return create_backend(
+            "supertonic",
             sample_rate=SAMPLE_RATE,
             auto_download=True,
             total_steps=int(getattr(job, "supertonic_total_steps", 5) or 5),
@@ -2454,8 +2453,7 @@ def _load_pipeline(job: Job):
     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_backend("kokoro", lang_code=job.language, device=device)
 
 
 def _select_device() -> str:

abogen/webui/debug_tts_runner.py

@@ -15,7 +15,7 @@ from abogen.normalization_settings import build_apostrophe_config
 from abogen.text_extractor import extract_from_path
 from abogen.voice_cache import ensure_voice_assets
 from abogen.webui.conversion_runner import SAMPLE_RATE, SPLIT_PATTERN, _select_device, _to_float32, _resolve_voice, _spec_to_voice_ids
-from abogen.utils import load_numpy_kpipeline
+from abogen.tts_backend_registry import create_backend
 
 
 _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_backend("kokoro", lang_code=language, device=device)
 
 
 def _extract_cases_from_text(text: str) -> List[Tuple[str, str]]:

abogen/webui/routes/utils/form.py

@@ -32,7 +32,7 @@ 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
@@ -616,8 +616,8 @@ 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 +796,8 @@ def build_pending_job_from_extraction(
         profile_selection = inferred_profile
 
     base_voice = base_voice_input or resolved_default_voice or str(default_voice_setting).strip()
-    if not base_voice and VOICES_INTERNAL:
-        base_voice = VOICES_INTERNAL[0]
+    if not base_voice:
+        base_voice = get_default_voice("kokoro")
     selected_speaker_config = (form.get("speaker_config") or "").strip()
     speaker_config_payload = get_config(selected_speaker_config) if selected_speaker_config else None
 

abogen/webui/routes/utils/preview.py

@@ -78,10 +78,9 @@ 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_backend_registry import create_backend
 
-        _, KPipeline = load_numpy_kpipeline()
-        pipeline = KPipeline(lang_code=language, repo_id="hexgrad/Kokoro-82M", device=device)
+        pipeline = create_backend("kokoro", lang_code=language, device=device)
         _preview_pipelines[key] = pipeline
         return pipeline
 
@@ -137,9 +136,9 @@ def generate_preview_audio(
             normalized_text = source_text
 
     if provider == "supertonic":
-        from abogen.tts_supertonic import SupertonicPipeline
+        from abogen.tts_backend_registry import create_backend
 
-        pipeline = SupertonicPipeline(sample_rate=SAMPLE_RATE, auto_download=True, total_steps=supertonic_total_steps)
+        pipeline = create_backend("supertonic", sample_rate=SAMPLE_RATE, auto_download=True, total_steps=supertonic_total_steps)
         segments = pipeline(
             normalized_text,
             voice=voice_spec,

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

@@ -17,10 +17,10 @@ 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.utils import load_numpy_kpipeline
+from abogen.tts_backend_registry import create_backend
 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_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,
@@ -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_backend("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__)
 

tests/test_kokoro_backend.py

@@ -0,0 +1,216 @@
+"""Tests for KokoroBackend class."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any, Iterator, List
+from unittest.mock import MagicMock, patch
+
+import numpy as np
+import pytest
+
+from abogen.tts_backend import TTSBackendMetadata
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+@dataclass
+class _FakeSegment:
+    graphemes: str
+    audio: Any  # np.ndarray or torch-like tensor
+
+
+class _FakePipeline:
+    """Minimal mock for kokoro.KPipeline."""
+
+    def __init__(self, *, lang_code: str, repo_id: str, device: str):
+        self.lang_code = lang_code
+        self.repo_id = repo_id
+        self.device = device
+        self._voices: dict[str, np.ndarray] = {}
+
+    def __call__(
+        self,
+        text: str,
+        *,
+        voice: Any = "",
+        speed: float = 1.0,
+        split_pattern: str | None = None,
+    ) -> Iterator[_FakeSegment]:
+        yield _FakeSegment(graphemes=text, audio=np.zeros(100, dtype="float32"))
+
+    def load_single_voice(self, name: str) -> np.ndarray:
+        if name not in self._voices:
+            self._voices[name] = np.ones((1, 256), dtype="float32") * 0.5
+        return self._voices[name]
+
+
+def _make_backend(**kwargs: Any):
+    """Create KokoroBackend with mocked KPipeline."""
+    with patch("abogen.tts_backends.kokoro._load_kpipeline") as load:
+        load.return_value = _FakePipeline
+        from abogen.tts_backends.kokoro import KokoroBackend
+
+        return KokoroBackend(**kwargs)
+
+
+# ---------------------------------------------------------------------------
+# Tests
+# ---------------------------------------------------------------------------
+
+class TestKokoroBackendMetadata:
+    def test_metadata_returns_tts_backend_metadata(self):
+        backend = _make_backend(lang_code="a")
+        meta = backend.metadata
+        assert isinstance(meta, TTSBackendMetadata)
+
+    def test_metadata_fields(self):
+        backend = _make_backend(lang_code="a")
+        meta = backend.metadata
+        assert meta.id == "kokoro"
+        assert meta.name == "Kokoro"
+        assert "Kokoro" in meta.description
+
+
+class TestKokoroBackendInit:
+    def test_stores_lang_code(self):
+        backend = _make_backend(lang_code="b")
+        assert backend._lang_code == "b"
+
+    def test_default_repo_id(self):
+        with patch("abogen.tts_backends.kokoro._load_kpipeline") as load:
+            load.return_value = _FakePipeline
+            from abogen.tts_backends.kokoro import KokoroBackend
+
+            b = KokoroBackend(lang_code="a")
+            assert b._pipeline.repo_id == "hexgrad/Kokoro-82M"
+
+    def test_custom_repo_id(self):
+        backend = _make_backend(lang_code="a", repo_id="custom/repo")
+        assert backend._pipeline.repo_id == "custom/repo"
+
+    def test_default_device(self):
+        backend = _make_backend(lang_code="a")
+        assert backend._pipeline.device == "cpu"
+
+    def test_custom_device(self):
+        backend = _make_backend(lang_code="a", device="cuda")
+        assert backend._pipeline.device == "cuda"
+
+
+class TestKokoroBackendCall:
+    def test_call_delegates_to_pipeline(self):
+        backend = _make_backend(lang_code="a")
+        results = list(backend("hello", voice="af_heart", speed=1.2, split_pattern=r"\n"))
+        assert len(results) == 1
+        assert results[0].graphemes == "hello"
+
+    def test_call_returns_iterator(self):
+        backend = _make_backend(lang_code="a")
+        result = backend("test", voice="af_heart")
+        assert hasattr(result, "__iter__")
+
+    def test_call_with_voice_tensor(self):
+        backend = _make_backend(lang_code="a")
+        voice_tensor = np.ones((1, 256), dtype="float32")
+        results = list(backend("test", voice=voice_tensor))
+        assert len(results) == 1
+
+    def test_call_default_speed(self):
+        backend = _make_backend(lang_code="a")
+        # Should not raise with default speed
+        list(backend("text", voice="af_heart"))
+
+    def test_call_default_split_pattern_is_none(self):
+        backend = _make_backend(lang_code="a")
+        # split_pattern defaults to None
+        list(backend("text", voice="af_heart"))
+
+
+class TestLoadSingleVoice:
+    def test_load_single_voice_delegates(self):
+        backend = _make_backend(lang_code="a")
+        tensor = backend.load_single_voice("af_heart")
+        assert isinstance(tensor, np.ndarray)
+        assert tensor.shape == (1, 256)
+
+    def test_load_single_voice_caches(self):
+        backend = _make_backend(lang_code="a")
+        t1 = backend.load_single_voice("af_heart")
+        t2 = backend.load_single_voice("af_heart")
+        assert t1 is t2  # same object
+
+
+class TestSynthesize:
+    def test_synthesize_returns_bytes(self):
+        backend = _make_backend(lang_code="a")
+        result = backend.synthesize("hello", voice="af_heart")
+        assert isinstance(result, bytes)
+
+    def test_synthesize_nonempty(self):
+        backend = _make_backend(lang_code="a")
+        result = backend.synthesize("hello", voice="af_heart")
+        assert len(result) > 0
+
+    def test_synthesize_with_speed(self):
+        backend = _make_backend(lang_code="a")
+        result = backend.synthesize("hello", voice="af_heart", speed=1.5)
+        assert isinstance(result, bytes)
+
+    def test_synthesize_empty_text(self):
+        backend = _make_backend(lang_code="a")
+        # Empty text produces no segments
+        result = backend.synthesize("", voice="af_heart")
+        assert isinstance(result, bytes)
+
+
+class TestProtocolMethods:
+    def test_get_available_voices(self):
+        backend = _make_backend(lang_code="a")
+        voices = backend.get_available_voices()
+        assert isinstance(voices, list)
+        assert len(voices) > 0
+        assert all(isinstance(v, str) for v in voices)
+
+    def test_get_supported_formats(self):
+        backend = _make_backend(lang_code="a")
+        formats = backend.get_supported_formats()
+        assert "pcm_float32" in formats
+
+    def test_get_info(self):
+        backend = _make_backend(lang_code="a")
+        info = backend.get_info()
+        assert info["id"] == "kokoro"
+        assert info["name"] == "Kokoro"
+        assert info["lang_code"] == "a"
+
+
+class TestRegistration:
+    def test_factory_creates_kokoro_backend(self):
+        from abogen.tts_backends.kokoro import create_kokoro_backend, KokoroBackend
+
+        with patch("abogen.tts_backends.kokoro._load_kpipeline") as load:
+            load.return_value = _FakePipeline
+            backend = create_kokoro_backend(lang_code="a")
+            assert isinstance(backend, KokoroBackend)
+
+    def test_registry_has_kokoro(self):
+        import abogen.tts_backends  # noqa: F401
+        from abogen.tts_backend_registry import _registry
+
+        meta = _registry.get_metadata("kokoro")
+        assert meta.id == "kokoro"
+        assert meta.name == "Kokoro"
+
+    def test_registry_factory_returns_kokoro_backend(self):
+        import abogen.tts_backends  # noqa: F401
+        from abogen.tts_backend_registry import _registry
+        from abogen.tts_backends.kokoro import KokoroBackend
+
+        factory = _registry._factories["kokoro"]
+        with patch("abogen.tts_backends.kokoro._load_kpipeline") as load:
+            load.return_value = _FakePipeline
+            backend = factory(lang_code="a")
+            assert isinstance(backend, KokoroBackend)

tests/test_preview_applies_manual_overrides.py

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

tests/test_tts_backend.py

@@ -0,0 +1,204 @@
+from dataclasses import dataclass
+
+from abogen.tts_backend import TTSBackendMetadata
+from abogen.tts_backend_registry import TTSBackendRegistry
+
+
+class TestTTSBackendMetadata:
+    def test_is_frozen_dataclass(self):
+        assert dataclass(TTSBackendMetadata)
+
+    def test_fields_are_present(self):
+        meta = TTSBackendMetadata(
+            id="test",
+            name="Test Backend",
+            description="A test backend",
+        )
+        assert meta.id == "test"
+        assert meta.name == "Test Backend"
+        assert meta.description == "A test backend"
+
+    def test_voices_field_default_empty(self):
+        meta = TTSBackendMetadata(
+            id="test",
+            name="Test",
+            description="Test backend",
+        )
+        assert meta.voices == ()
+
+    def test_voices_field_stored(self):
+        meta = TTSBackendMetadata(
+            id="test",
+            name="Test",
+            description="Test backend",
+            voices=("v1", "v2"),
+        )
+        assert meta.voices == ("v1", "v2")
+
+    def test_is_immutable(self):
+        import pytest
+
+        meta = TTSBackendMetadata(
+            id="kokoro",
+            name="Kokoro",
+            description="Test",
+        )
+        with pytest.raises(Exception):
+            meta.id = "changed"
+
+
+class TestTTSBackendRegistry:
+    def test_register_and_list(self):
+        registry = TTSBackendRegistry()
+        meta = TTSBackendMetadata(id="a", name="A", description="Backend A")
+        registry.register(metadata=meta, factory=lambda: None)
+
+        backends = registry.list_backends()
+        assert len(backends) == 1
+        assert backends[0].id == "a"
+
+    def test_list_multiple(self):
+        registry = TTSBackendRegistry()
+        meta_a = TTSBackendMetadata(id="a", name="A", description="A")
+        meta_b = TTSBackendMetadata(id="b", name="B", description="B")
+        registry.register(metadata=meta_a, factory=lambda: None)
+        registry.register(metadata=meta_b, factory=lambda: None)
+
+        backends = registry.list_backends()
+        ids = [b.id for b in backends]
+        assert "a" in ids
+        assert "b" in ids
+
+    def test_get_metadata(self):
+        registry = TTSBackendRegistry()
+        meta = TTSBackendMetadata(id="x", name="X", description="X backend")
+        registry.register(metadata=meta, factory=lambda: None)
+
+        result = registry.get_metadata("x")
+        assert result.id == "x"
+        assert result.name == "X"
+
+    def test_get_metadata_unknown_raises(self):
+        import pytest
+
+        registry = TTSBackendRegistry()
+        with pytest.raises(KeyError, match="Unknown backend: nope"):
+            registry.get_metadata("nope")
+
+    def test_create_backend(self):
+        registry = TTSBackendRegistry()
+        meta = TTSBackendMetadata(id="test", name="Test", description="Test backend")
+
+        def factory(**kwargs):
+            return {"created": True, "kwargs": kwargs}
+
+        registry.register(metadata=meta, factory=factory)
+        result = registry.create_backend("test", foo="bar")
+
+        assert result == {"created": True, "kwargs": {"foo": "bar"}}
+
+    def test_create_backend_unknown_raises(self):
+        import pytest
+
+        registry = TTSBackendRegistry()
+        with pytest.raises(KeyError, match="Unknown backend: missing"):
+            registry.create_backend("missing")
+
+    def test_register_overwrites(self):
+        registry = TTSBackendRegistry()
+        meta1 = TTSBackendMetadata(id="x", name="V1", description="First")
+        meta2 = TTSBackendMetadata(id="x", name="V2", description="Second")
+        registry.register(metadata=meta1, factory=lambda: "v1")
+        registry.register(metadata=meta2, factory=lambda: "v2")
+
+        result = registry.get_metadata("x")
+        assert result.name == "V2"
+        assert registry.create_backend("x") == "v2"
+
+
+class TestBackendRegistration:
+    """Tests that existing backends are auto-registered."""
+
+    def test_import_triggers_registration(self):
+        import abogen.tts_backends  # noqa: F401
+
+        from abogen.tts_backend_registry import _registry
+
+        backends = _registry.list_backends()
+        ids = [b.id for b in backends]
+        assert "kokoro" in ids
+        assert "supertonic" in ids
+
+    def test_kokoro_metadata(self):
+        import abogen.tts_backends  # noqa: F401
+
+        from abogen.tts_backend_registry import _registry
+
+        meta = _registry.get_metadata("kokoro")
+        assert meta.id == "kokoro"
+        assert meta.name == "Kokoro"
+        assert "Kokoro" in meta.description
+
+    def test_supertonic_metadata(self):
+        import abogen.tts_backends  # noqa: F401
+
+        from abogen.tts_backend_registry import _registry
+
+        meta = _registry.get_metadata("supertonic")
+        assert meta.id == "supertonic"
+        assert meta.name == "SuperTonic"
+        assert "SuperTonic" in meta.description
+
+    def test_kokoro_metadata_has_voices(self):
+        import abogen.tts_backends  # noqa: F401
+
+        from abogen.tts_backend_registry import _registry
+
+        meta = _registry.get_metadata("kokoro")
+        assert isinstance(meta.voices, tuple)
+        assert len(meta.voices) > 0
+        assert all(isinstance(v, str) for v in meta.voices)
+
+    def test_supertonic_metadata_has_voices(self):
+        import abogen.tts_backends  # noqa: F401
+
+        from abogen.tts_backend_registry import _registry
+
+        meta = _registry.get_metadata("supertonic")
+        assert isinstance(meta.voices, tuple)
+        assert len(meta.voices) == 10
+        assert meta.voices == ("M1", "M2", "M3", "M4", "M5", "F1", "F2", "F3", "F4", "F5")
+
+    def test_kokoro_factory_callable(self):
+        import abogen.tts_backends  # noqa: F401
+
+        from abogen.tts_backend_registry import _registry
+
+        factory = _registry._factories["kokoro"]
+        assert callable(factory)
+
+    def test_supertonic_factory_callable(self):
+        import abogen.tts_backends  # noqa: F401
+
+        from abogen.tts_backend_registry import _registry
+
+        factory = _registry._factories["supertonic"]
+        assert callable(factory)
+
+    def test_kokoro_metadata_voices_match_registry(self):
+        """Ensure the metadata property on the instance shares voices with registry."""
+        from abogen.tts_backends.kokoro import _KOKORO_METADATA
+        from abogen.tts_backend_registry import _registry
+
+        registry_meta = _registry.get_metadata("kokoro")
+        assert _KOKORO_METADATA is registry_meta
+        assert _KOKORO_METADATA.voices == registry_meta.voices
+
+    def test_supertonic_metadata_voices_match_registry(self):
+        """Ensure the metadata property on the instance shares voices with registry."""
+        from abogen.tts_backends.supertonic import _SUPERTONIC_METADATA
+        from abogen.tts_backend_registry import _registry
+
+        registry_meta = _registry.get_metadata("supertonic")
+        assert _SUPERTONIC_METADATA is registry_meta
+        assert _SUPERTONIC_METADATA.voices == registry_meta.voices

tests/test_tts_supertonic_unsupported_chars.py

@@ -1,6 +1,6 @@
 import numpy as np
 
-from abogen.tts_supertonic import SupertonicPipeline
+from abogen.tts_backends.supertonic import SupertonicBackend, SupertonicPipeline
 
 
 class _DummyTTS:
@@ -26,13 +26,23 @@ class _DummyTTS:
         return audio, 0.05
 
 
-def test_supertonic_pipeline_strips_unsupported_characters_and_retries():
-    # Avoid importing/initializing real supertonic by manually constructing the pipeline.
+def _make_pipeline() -> SupertonicPipeline:
     pipeline = SupertonicPipeline.__new__(SupertonicPipeline)
     pipeline.sample_rate = 24000
     pipeline.total_steps = 5
     pipeline.max_chunk_length = 1000
     pipeline._tts = _DummyTTS()
+    return pipeline
+
+
+def _make_backend() -> SupertonicBackend:
+    backend = SupertonicBackend.__new__(SupertonicBackend)
+    backend._pipeline = _make_pipeline()
+    return backend
+
+
+def test_supertonic_pipeline_strips_unsupported_characters_and_retries():
+    pipeline = _make_pipeline()
 
     segs = list(pipeline("Hello • world", voice="M1", speed=1.0))
     assert len(segs) == 1
@@ -43,11 +53,56 @@ def test_supertonic_pipeline_strips_unsupported_characters_and_retries():
 
 
 def test_supertonic_pipeline_drops_chunk_if_only_unsupported_characters():
-    pipeline = SupertonicPipeline.__new__(SupertonicPipeline)
-    pipeline.sample_rate = 24000
-    pipeline.total_steps = 5
-    pipeline.max_chunk_length = 1000
-    pipeline._tts = _DummyTTS()
+    pipeline = _make_pipeline()
 
     segs = list(pipeline("•", voice="M1", speed=1.0))
     assert segs == []
+
+
+# --- SupertonicBackend tests ---
+
+
+def test_backend_metadata():
+    backend = _make_backend()
+    meta = backend.metadata
+    assert meta.id == "supertonic"
+    assert meta.name == "SuperTonic"
+    assert "SuperTonic" in meta.description
+
+
+def test_backend_get_available_voices():
+    backend = _make_backend()
+    voices = backend.get_available_voices()
+    assert isinstance(voices, list)
+    assert "M1" in voices
+    assert "F1" in voices
+
+
+def test_backend_get_supported_formats():
+    backend = _make_backend()
+    formats = backend.get_supported_formats()
+    assert "wav" in formats
+
+
+def test_backend_get_info():
+    backend = _make_backend()
+    info = backend.get_info()
+    assert info["sample_rate"] == 24000
+    assert info["total_steps"] == 5
+    assert isinstance(info["voices"], list)
+
+
+def test_backend_call_delegates_to_pipeline():
+    backend = _make_backend()
+    segs = list(backend("Hello • world", voice="M1", speed=1.0))
+    assert len(segs) == 1
+    assert segs[0].audio.size > 0
+
+
+def test_backend_synthesize_returns_wav_bytes():
+    backend = _make_backend()
+    wav_bytes = backend.synthesize("Hello world", voice="M1", speed=1.0)
+    assert isinstance(wav_bytes, bytes)
+    assert len(wav_bytes) > 0
+    # WAV magic number
+    assert wav_bytes[:4] == b"RIFF"

tests/test_voice_formula_resolution.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 
 from abogen.webui.conversion_runner import _resolve_voice, _supertonic_voice_from_spec
-from abogen.tts_supertonic import DEFAULT_SUPERTONIC_VOICES
+from abogen.tts_backends.supertonic import DEFAULT_SUPERTONIC_VOICES
 
 
 def test_resolve_voice_formula_without_pipeline_does_not_crash() -> None:

tests/test_voice_metadata.py

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