Merge pull request #182 from denizsafak/refactor/add-kokoro-backend

feat: add KokoroBackend implementing TTSBackend protocol

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

@@ -1,117 +1,87 @@
 """
-Minimal TTS Backend Interface
+TTS Backend Interface
 
-This module defines a minimal interface for TTS backends to enable future
-extensibility while maintaining backward compatibility with existing Kokoro
-implementation.
+This module defines the protocol for TTS backends and the
+metadata model that describes a backend implementation.
 """
 
-from abc import ABC, abstractmethod
-from typing import Any, Iterator, Optional, Union
+from dataclasses import dataclass
+from typing import Protocol, List, Dict, Any
 
 
-class TTSBackend(ABC):
+@dataclass(frozen=True)
+class TTSBackendMetadata:
     """
-    Minimal interface for TTS backends.
-    
-    This interface is designed to be minimal and focused on the essential
-    operations needed for text-to-speech conversion.
+    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.
     """
 
-    @abstractmethod
-    def __call__(
-        self,
-        text: str,
-        voice: Union[str, Any],
-        speed: float = 1.0,
-        **kwargs: Any
-    ) -> Iterator[Any]:
+    id: str
+    name: str
+    description: 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:
         """
-        Generate speech segments from text.
-        
+        Initialize the TTS backend.
+
         Args:
-            text: Text to convert to speech
-            voice: Voice specification or object
-            speed: Speed multiplier for speech
-            **kwargs: Additional backend-specific parameters
-            
-        Yields:
-            Speech segments (audio data, timing info, etc.)
+            **kwargs: Backend-specific configuration parameters
         """
-        pass
+        ...
 
-
-class KokoroTTSBackend(TTSBackend):
-    """
-    Implementation of TTSBackend using Kokoro.
-    
-    This class provides the concrete implementation that maintains
-    the existing behavior while conforming to the TTSBackend interface.
-    """
-
-    def __init__(self, lang_code: str, repo_id: str = "hexgrad/Kokoro-82M", device: str = "cpu"):
+    def synthesize(self, text: str, **kwargs) -> bytes:
         """
-        Initialize Kokoro backend.
-        
+        Synthesize speech from text.
+
         Args:
-            lang_code: Language code for the model
-            repo_id: Repository ID for the Kokoro model
-            device: Device to run the model on (cpu, cuda, etc.)
-        """
-        self.lang_code = lang_code
-        self.repo_id = repo_id
-        self.device = device
-        self._pipeline = None
+            text: Text to synthesize
+            **kwargs: Additional parameters for synthesis
 
-    def _get_pipeline(self):
-        """Lazy initialization of the Kokoro pipeline."""
-        if self._pipeline is None:
-            from abogen.utils import load_numpy_kpipeline
-            _, KPipeline = load_numpy_kpipeline()
-            try:
-                self._pipeline = KPipeline(
-                    lang_code=self.lang_code,
-                    repo_id=self.repo_id,
-                    device=self.device
-                )
-            except RuntimeError as e:
-                if "CUDA" in str(e) and self.device != "cpu":
-                    # Fall back to CPU if CUDA fails
-                    self._pipeline = KPipeline(
-                        lang_code=self.lang_code,
-                        repo_id=self.repo_id,
-                        device="cpu"
-                    )
-                else:
-                    raise
-        return self._pipeline
+        Returns:
+            Audio data as bytes
+        """
+        ...
 
-    def __call__(
-        self,
-        text: str,
-        voice: Union[str, Any],
-        speed: float = 1.0,
-        split_pattern: str = r"\n+",
-        **kwargs: Any
-    ) -> Iterator[Any]:
+    def get_available_voices(self) -> List[str]:
         """
-        Generate speech segments from text using Kokoro.
-        
-        Args:
-            text: Text to convert to speech
-            voice: Voice specification or object
-            speed: Speed multiplier for speech
-            split_pattern: Pattern to split text into segments
-            **kwargs: Additional parameters passed to the pipeline
-            
-        Yields:
-            Speech segments
+        Get list of available voices.
+
+        Returns:
+            List of voice identifiers
         """
-        pipeline = self._get_pipeline()
-        return pipeline(
-            text,
-            voice=voice,
-            speed=speed,
-            split_pattern=split_pattern,
-            **kwargs
-        )
+        ...
+
+    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 +1,71 @@
+"""
+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 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 +1,124 @@
+"""
+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
+
+
+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):
+        from abogen.tts_backend import TTSBackendMetadata
+
+        return TTSBackendMetadata(
+            id="kokoro",
+            name="Kokoro",
+            description="Kokoro TTS engine",
+        )
+
+    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."""
+        from abogen.constants import VOICES_INTERNAL
+
+        return list(VOICES_INTERNAL)
+
+    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 import TTSBackendMetadata  # noqa: E402
+from abogen.tts_backend_registry import register_backend  # noqa: E402
+
+register_backend(
+    metadata=TTSBackendMetadata(
+        id="kokoro",
+        name="Kokoro",
+        description="Kokoro TTS engine",
+    ),
+    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
 
@@ -273,3 +273,120 @@ 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 TTSBackendMetadata(
+            id="supertonic",
+            name="SuperTonic",
+            description="SuperTonic TTS engine",
+        )
+
+    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(DEFAULT_SUPERTONIC_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 import TTSBackendMetadata
+from abogen.tts_backend_registry import register_backend
+
+register_backend(
+    metadata=TTSBackendMetadata(
+        id="supertonic",
+        name="SuperTonic",
+        description="SuperTonic TTS engine",
+    ),
+    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_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

@@ -3,7 +3,7 @@ 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_backends.supertonic import DEFAULT_SUPERTONIC_VOICES
 from abogen.utils import get_user_config_path
 
 

abogen/webui/conversion_runner.py

@@ -39,15 +39,15 @@ from abogen.utils import (
     get_user_cache_path,
     get_user_output_path,
     load_config,
-    load_numpy_kpipeline,
 )
-from abogen.tts_backend import KokoroTTSBackend
+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 abogen.tts_backends.supertonic import DEFAULT_SUPERTONIC_VOICES
 
 from .service import Job, JobStatus
 
@@ -1582,7 +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),
@@ -1595,10 +1596,10 @@ def run_conversion_job(job: Job) -> None:
             device = "cpu"
             if not disable_gpu:
                 device = _select_device()
-            # Create KokoroTTSBackend instance instead of directly instantiating KPipeline
-            pipelines[provider_norm] = KokoroTTSBackend(
+            # Create KPipeline instance directly (conforms to TTSBackend protocol)
+            pipelines[provider_norm] = create_backend(
+                "kokoro",
                 lang_code=job.language,
-                repo_id="hexgrad/Kokoro-82M",
                 device=device
             )
             if not kokoro_cache_ready:
@@ -2442,7 +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),
@@ -2451,8 +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/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/voice.py

@@ -20,7 +20,7 @@ from abogen.constants import (
     VOICES_INTERNAL,
 )
 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()
@@ -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
 

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 +1,149 @@
+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_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_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)

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"