Update README.md to include Supertonic TTS in PyQt6 Desktop GUI features - Supertonic 3 support

Fixes #163

.github/workflows/test_pip.yml

@@ -1,9 +1,7 @@
-name: CI
-run-name: CI
-
+name: pip install
+run-name: pip install
 on: 
   push:
-    branches: [main]
     paths:
       - '**.py'
       - 'pyproject.toml'
@@ -13,42 +11,23 @@ on:
       - 'pyproject.toml'
       - '.github/workflows/**'
   workflow_dispatch:
-
 jobs:
-  test:
+  install-and-run:
     strategy:
       matrix:
-        os: [ubuntu-latest, macos-14, windows-latest]
+        os: [ubuntu-latest, macos-latest, windows-latest]
         python-version: ['3.12']
       fail-fast: false
+    continue-on-error: true
     runs-on: ${{ matrix.os }}
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v7
-
+        uses: actions/checkout@v4
       - name: Set up Python
-        uses: actions/setup-python@v6
+        uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
-
-      - name: Install uv
-        uses: astral-sh/setup-uv@v8.3.1
-        with:
-          enable-cache: true
-          cache-dependency-glob: pyproject.toml
-
-      - name: Install system dependencies (Ubuntu)
-        if: runner.os == 'Linux'
-        run: sudo apt-get update && sudo apt-get install -y libegl1
-
-      - name: Install dependencies
-        run: uv pip install --system .[dev]
-
-      - name: Run tests
-        env:
-          QT_QPA_PLATFORM: offscreen
-        run: pytest tests/ -v --tb=short
-
-      - name: Minimize uv cache
-        if: always()
-        run: uv cache prune --ci
+      - name: Install from repository
+        run: python -m pip install .
+      #- name: Run abogen
+      #  run: abogen

.github/workflows/test_publish.yml

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

README.md

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

WINDOWS_INSTALL.bat

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

abogen/constants.py

@@ -29,6 +29,57 @@ LANGUAGE_DESCRIPTIONS = {
     "z": "Mandarin Chinese",
 }
 
+# Mapping from Kokoro single-letter language codes to ISO 3166-1 alpha-2 country codes
+# Used for loading flag icons
+KOKORO_LANG_TO_COUNTRY = {
+    "a": "us",  # American English -> United States
+    "b": "gb",  # British English -> United Kingdom
+    "e": "es",  # Spanish -> Spain
+    "f": "fr",  # French -> France
+    "h": "in",  # Hindi -> India
+    "i": "it",  # Italian -> Italy
+    "j": "jp",  # Japanese -> Japan
+    "p": "br",  # Brazilian Portuguese -> Brazil
+    "z": "cn",  # Mandarin Chinese -> China
+}
+
+# Mapping from Supertonic ISO 639-1 language codes to ISO 3166-1 alpha-2 country codes
+# Used for loading flag icons in the Supertonic language picker
+SUPERTONIC_LANG_TO_COUNTRY = {
+    "en": "gb",
+    "ko": "kr",
+    "ja": "jp",
+    "ar": "ae",
+    "bg": "bg",
+    "cs": "cz",
+    "da": "dk",
+    "de": "de",
+    "el": "gr",
+    "es": "es",
+    "et": "ee",
+    "fi": "fi",
+    "fr": "fr",
+    "hi": "in",
+    "hr": "hr",
+    "hu": "hu",
+    "id": "id",
+    "it": "it",
+    "lt": "lt",
+    "lv": "lv",
+    "nl": "nl",
+    "pl": "pl",
+    "pt": "pt",
+    "ro": "ro",
+    "ru": "ru",
+    "sk": "sk",
+    "sl": "si",
+    "sv": "se",
+    "tr": "tr",
+    "uk": "ua",
+    "vi": "vn",
+    "na": "na",
+}
+
 # Supported sound formats
 SUPPORTED_SOUND_FORMATS = [
     "wav",
@@ -63,6 +114,64 @@ SUPPORTED_INPUT_FORMATS = [
 # 384 if self.lang_code in 'ab':
 SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION = list(LANGUAGE_DESCRIPTIONS.keys())
 
+# Voice and sample text constants
+VOICES_INTERNAL = [
+    "af_alloy",
+    "af_aoede",
+    "af_bella",
+    "af_heart",
+    "af_jessica",
+    "af_kore",
+    "af_nicole",
+    "af_nova",
+    "af_river",
+    "af_sarah",
+    "af_sky",
+    "am_adam",
+    "am_echo",
+    "am_eric",
+    "am_fenrir",
+    "am_liam",
+    "am_michael",
+    "am_onyx",
+    "am_puck",
+    "am_santa",
+    "bf_alice",
+    "bf_emma",
+    "bf_isabella",
+    "bf_lily",
+    "bm_daniel",
+    "bm_fable",
+    "bm_george",
+    "bm_lewis",
+    "ef_dora",
+    "em_alex",
+    "em_santa",
+    "ff_siwis",
+    "hf_alpha",
+    "hf_beta",
+    "hm_omega",
+    "hm_psi",
+    "if_sara",
+    "im_nicola",
+    "jf_alpha",
+    "jf_gongitsune",
+    "jf_nezumi",
+    "jf_tebukuro",
+    "jm_kumo",
+    "pf_dora",
+    "pm_alex",
+    "pm_santa",
+    "zf_xiaobei",
+    "zf_xiaoni",
+    "zf_xiaoxiao",
+    "zf_xiaoyi",
+    "zm_yunjian",
+    "zm_yunxi",
+    "zm_yunxia",
+    "zm_yunyang",
+]
+
 # Voice and sample text mapping
 SAMPLE_VOICE_TEXTS = {
     "a": "This is a sample of the selected voice.",

abogen/predownload_gui.py

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

abogen/pyqt/conversion.py

@@ -5,7 +5,6 @@ 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,
@@ -21,6 +20,7 @@ from abogen.constants import (
     SUPPORTED_SOUND_FORMATS,
     SUPPORTED_SUBTITLE_FORMATS,
 )
+from abogen.tts_supertonic import SupertonicPipeline, SUPERTONIC_AVAILABLE_LANGS, DEFAULT_SUPERTONIC_VOICES
 from abogen.voice_formulas import get_new_voice
 import abogen.hf_tracker as hf_tracker
 import static_ffmpeg
@@ -260,21 +260,30 @@ class ConversionThread(QThread):
         output_folder,
         subtitle_mode,
         output_format,
-        backend,
+        np_module,
+        kpipeline_class,
         start_time,
         total_char_count,
         use_gpu=True,
         from_queue=False,
         save_base_path=None,
-    ):  # Add use_gpu parameter
+        tts_provider="kokoro",
+        supertonic_language="en",
+        supertonic_total_steps=8,
+    ):
         super().__init__()
         self._chapter_options_event = threading.Event()
         self._timestamp_response_event = threading.Event()
-        self.backend = backend
+        self.np = np_module
+        self.KPipeline = kpipeline_class
         self.file_name = file_name
         self.lang_code = lang_code
         self.speed = speed
         self.voice = voice
+        self.tts_provider = tts_provider
+        self.supertonic_language = supertonic_language
+        self.supertonic_total_steps = supertonic_total_steps
+        self.sample_rate = 44100 if tts_provider == "supertonic" else 24000
         self.save_option = save_option
         self.output_folder = output_folder
         self.subtitle_mode = subtitle_mode
@@ -426,6 +435,10 @@ class ConversionThread(QThread):
             )
             self.log_updated.emit(f"- Voice: {self.voice}")
             self.log_updated.emit(f"- Speed: {self.speed}")
+            tts_provider_label = self.tts_provider.capitalize()
+            if self.tts_provider == "supertonic":
+                tts_provider_label += f" (lang={self.supertonic_language}, steps={self.supertonic_total_steps})"
+            self.log_updated.emit(f"- TTS Engine: {tts_provider_label}")
             self.log_updated.emit(f"- Subtitle mode: {self.subtitle_mode}")
             self.log_updated.emit(f"- Output format: {self.output_format}")
             self.log_updated.emit(
@@ -489,6 +502,26 @@ class ConversionThread(QThread):
 
             self.log_updated.emit(("\nInitializing TTS pipeline...", "grey"))
 
+            # Set device based on use_gpu setting and platform
+            if self.use_gpu:
+                if platform.system() == "Darwin" and platform.processor() == "arm":
+                    device = "mps"  # Use MPS for Apple Silicon
+                else:
+                    device = "cuda"  # Use CUDA for other platforms
+            else:
+                device = "cpu"
+
+            if self.tts_provider == "supertonic":
+                tts = SupertonicPipeline(
+                    sample_rate=self.sample_rate,
+                    lang=self.supertonic_language,
+                    total_steps=self.supertonic_total_steps,
+                )
+            else:
+                tts = self.KPipeline(
+                    lang_code=self.lang_code, repo_id="hexgrad/Kokoro-82M", device=device
+                )
+
             # Check if the input is a subtitle file or timestamp text file
             is_subtitle_file = False
             is_timestamp_text = False
@@ -524,7 +557,7 @@ class ConversionThread(QThread):
 
             # Process subtitle files separately
             if is_subtitle_file or is_timestamp_text:
-                self._process_subtitle_file(self.backend, base_path, is_timestamp_text)
+                self._process_subtitle_file(tts, base_path, is_timestamp_text)
                 return
 
             if self.is_direct_text:
@@ -733,7 +766,7 @@ class ConversionThread(QThread):
                 merged_out_path = f"{base_filepath_no_ext}.{self.output_format}"
                 subtitle_entries = []
                 current_time = 0.0
-                rate = 24000
+                rate = self.sample_rate
                 subtitle_mode = self.subtitle_mode
                 self.etr_start_time = time.time()
                 self.processed_char_count = 0
@@ -749,7 +782,7 @@ class ConversionThread(QThread):
                     merged_out_file = sf.SoundFile(
                         merged_out_path,
                         "w",
-                        samplerate=24000,
+                        samplerate=self.sample_rate,
                         channels=1,
                         format=self.output_format,
                     )
@@ -765,62 +798,18 @@ class ConversionThread(QThread):
                     # Prepare ffmpeg command for m4b output
                     cmd = [
                         "ffmpeg",
-                        "-y",
-                        "-thread_queue_size",
-                        "32768",
-                        "-f",
-                        "f32le",
-                        "-ar",
-                        "24000",
-                        "-ac",
-                        "1",
-                        "-i",
-                        "pipe:0",
-                    ]
-                    if cover_path and os.path.exists(cover_path):
-                        cmd.extend(
-                            [
-                                "-i",
-                                cover_path,
-                                "-map",
-                                "0:a",
-                                "-map",
-                                "1",
-                                "-c:v",
-                                "copy",
-                                "-disposition:v",
-                                "attached_pic",
-                            ]
-                        )
-                    cmd.extend(
-                        [
-                            "-c:a",
-                            "aac",
-                            "-q:a",
-                            "2",
-                            "-movflags",
-                            "+faststart+use_metadata_tags",
+                            "-y",
+                            "-thread_queue_size",
+                            "32768",
+                            "-f",
+                            "f32le",
+                            "-ar",
+                            str(self.sample_rate),
+                            "-ac",
+                            "1",
+                            "-i",
+                            "pipe:0",
                         ]
-                    )
-                    cmd += metadata_options
-                    cmd.append(merged_out_path)
-                    ffmpeg_proc = create_process(cmd, stdin=subprocess.PIPE, text=False)
-                elif self.output_format == "opus":
-                    static_ffmpeg.add_paths()
-                    cmd = [
-                        "ffmpeg",
-                        "-y",
-                        "-thread_queue_size",
-                        "32768",
-                        "-f",
-                        "f32le",
-                        "-ar",
-                        "24000",
-                        "-ac",
-                        "1",
-                        "-i",
-                        "pipe:0",
-                    ]
                     cmd.extend(["-c:a", "libopus", "-b:a", "24000"])
                     cmd.append(merged_out_path)
                     ffmpeg_proc = create_process(cmd, stdin=subprocess.PIPE, text=False)
@@ -902,7 +891,7 @@ class ConversionThread(QThread):
                 merged_out_path = None
                 subtitle_entries = []
                 current_time = 0.0
-                rate = 24000
+                rate = self.sample_rate
                 subtitle_mode = self.subtitle_mode
                 self.etr_start_time = time.time()
                 self.processed_char_count = 0
@@ -955,7 +944,7 @@ class ConversionThread(QThread):
                         chapter_out_file = sf.SoundFile(
                             chapter_out_path,
                             "w",
-                            samplerate=24000,
+                            samplerate=self.sample_rate,
                             channels=1,
                             format=separate_chapters_format,
                         )
@@ -970,7 +959,7 @@ class ConversionThread(QThread):
                             "-f",
                             "f32le",
                             "-ar",
-                            "24000",
+                            str(self.sample_rate),
                             "-ac",
                             "1",
                             "-i",
@@ -1057,7 +1046,7 @@ 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, self.backend)
+                        loaded_voice = self.load_voice_cached(voice_name, tts)
                         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"))
@@ -1066,7 +1055,7 @@ 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, self.backend)
+                            loaded_voice = self.load_voice_cached(self.voice, tts)
 
                     # Determine if spaCy segmentation should be used for PRE-TTS segmentation
                     # Only non-English languages use spaCy for pre-segmentation
@@ -1152,7 +1141,7 @@ class ConversionThread(QThread):
                         print("Using split pattern: (unprintable)")
 
                     for text_segment in text_segments:
-                        for result in self.backend(
+                        for result in tts(
                             text_segment,
                             voice=loaded_voice,
                             speed=self.speed,
@@ -1352,9 +1341,9 @@ class ConversionThread(QThread):
                 # Add silence between chapters for merged output (except after the last chapter)
                 if merge_chapters_at_end and chapter_idx < total_chapters:
                     silence_samples = int(
-                        self.silence_duration * 24000
+                        self.silence_duration * self.sample_rate
                     )  # Silence duration at 24,000 Hz
-                    silence_audio = np.zeros(silence_samples, dtype="float32")
+                    silence_audio = self.np.zeros(silence_samples, dtype="float32")
                     silence_bytes = silence_audio.tobytes()
 
                     if merged_out_file:
@@ -1583,7 +1572,7 @@ class ConversionThread(QThread):
                 parent_dir, f"{sanitized_base_name}{suffix}"
             )
             merged_out_path = f"{base_filepath_no_ext}.{self.output_format}"
-            rate = 24000
+            rate = self.sample_rate
 
             # Setup audio output
             merged_out_file, ffmpeg_proc = None, None
@@ -1693,7 +1682,7 @@ class ConversionThread(QThread):
             max_end_time = max(
                 (end for _, end, _ in subtitles if end is not None), default=0
             )
-            audio_buffer = np.zeros(
+            audio_buffer = self.np.zeros(
                 int(max_end_time * rate) + rate, dtype="float32"
             )
 
@@ -1757,7 +1746,7 @@ class ConversionThread(QThread):
                 # Generate TTS audio
                 tts_results = [
                     r
-                    for r in self.backend(
+                    for r in tts(
                         processed_text,
                         voice=loaded_voice,
                         speed=self.speed,
@@ -1775,11 +1764,11 @@ class ConversionThread(QThread):
 
                 # Concatenate audio and determine duration
                 full_audio = (
-                    np.concatenate(
+                    self.np.concatenate(
                         [a.numpy() if hasattr(a, "numpy") else a for a in audio_chunks]
                     )
                     if audio_chunks
-                    else np.zeros(
+                    else self.np.zeros(
                         int((subtitle_duration or 0) * rate), dtype="float32"
                     )
                 )
@@ -1813,8 +1802,8 @@ class ConversionThread(QThread):
                         num_stages = max(
                             1,
                             int(
-                                np.ceil(
-                                    np.log(speed_factor) / np.log(2.0)
+                                self.np.ceil(
+                                    self.np.log(speed_factor) / self.np.log(2.0)
                                 )
                             ),
                         )
@@ -1847,7 +1836,7 @@ class ConversionThread(QThread):
                             stdout=subprocess.PIPE,
                             stderr=subprocess.PIPE,
                         )
-                        full_audio = np.frombuffer(
+                        full_audio = self.np.frombuffer(
                             speed_proc.communicate(input=full_audio.tobytes())[0],
                             dtype="float32",
                         )
@@ -1861,7 +1850,7 @@ class ConversionThread(QThread):
 
                         tts_results = [
                             r
-                            for r in self.backend(
+                            for r in tts(
                                 processed_text,
                                 voice=loaded_voice,
                                 speed=new_speed,
@@ -1872,14 +1861,14 @@ class ConversionThread(QThread):
                         audio_chunks = [r.audio for r in tts_results]
 
                         full_audio = (
-                            np.concatenate(
+                            self.np.concatenate(
                                 [
                                     a.numpy() if hasattr(a, "numpy") else a
                                     for a in audio_chunks
                                 ]
                             )
                             if audio_chunks
-                            else np.zeros(
+                            else self.np.zeros(
                                 int(subtitle_duration * rate), dtype="float32"
                             )
                         )
@@ -1896,10 +1885,10 @@ class ConversionThread(QThread):
                 # Pad or trim to subtitle duration
                 target_samples = int(subtitle_duration * rate)
                 if len(full_audio) < target_samples:
-                    full_audio = np.concatenate(
+                    full_audio = self.np.concatenate(
                         [
                             full_audio,
-                            np.zeros(
+                            self.np.zeros(
                                 target_samples - len(full_audio), dtype="float32"
                             ),
                         ]
@@ -1912,10 +1901,10 @@ class ConversionThread(QThread):
                 end_sample = start_sample + len(full_audio)
                 if end_sample > len(audio_buffer):
                     # Extend buffer if needed
-                    audio_buffer = np.concatenate(
+                    audio_buffer = self.np.concatenate(
                         [
                             audio_buffer,
-                            np.zeros(
+                            self.np.zeros(
                                 end_sample - len(audio_buffer), dtype="float32"
                             ),
                         ]
@@ -1957,7 +1946,7 @@ class ConversionThread(QThread):
                 self.progress_updated.emit(percent, etr_str)
 
             # Normalize audio buffer to prevent clipping from mixed overlaps
-            max_amplitude = np.abs(audio_buffer).max()
+            max_amplitude = self.np.abs(audio_buffer).max()
             if max_amplitude > 1.0:
                 self.log_updated.emit(
                     f"\n  -> Normalizing audio (peak: {max_amplitude:.2f})"
@@ -2426,19 +2415,28 @@ class VoicePreviewThread(QThread):
 
     def __init__(
         self,
-        backend,
+        np_module,
+        kpipeline_class,
         lang_code,
         voice,
         speed,
         use_gpu=False,
         parent=None,
+        tts_provider="kokoro",
+        supertonic_language="en",
+        supertonic_total_steps=8,
     ):
         super().__init__(parent)
-        self.backend = backend
+        self.np_module = np_module
+        self.kpipeline_class = kpipeline_class
         self.lang_code = lang_code
         self.voice = voice
         self.speed = speed
         self.use_gpu = use_gpu
+        self.tts_provider = tts_provider
+        self.supertonic_language = supertonic_language
+        self.supertonic_total_steps = supertonic_total_steps
+        self.sample_rate = 44100 if tts_provider == "supertonic" else 24000
 
         # Cache location for preview audio
         self.cache_dir = get_user_cache_path("preview_cache")
@@ -2448,6 +2446,11 @@ class VoicePreviewThread(QThread):
 
     def _get_cache_path(self):
         """Generate a unique filename for the voice with its parameters"""
+        if self.tts_provider == "supertonic":
+            voice_id = self.voice or "M1"
+            filename = f"st_{voice_id}_{self.supertonic_language}_steps{self.supertonic_total_steps}_{self.speed:.2f}.wav"
+            return os.path.join(self.cache_dir, filename)
+
         # For a voice formula, use a hash of the formula
         if "*" in self.voice:
             voice_id = (
@@ -2462,27 +2465,56 @@ class VoicePreviewThread(QThread):
 
     def run(self):
         print(
-            f"\nVoice: {self.voice}\nLanguage: {self.lang_code}\nSpeed: {self.speed}\nGPU: {self.use_gpu}\n"
+            f"\nVoice: {self.voice}\nLanguage: {self.lang_code}\nSpeed: {self.speed}\nGPU: {self.use_gpu}\nTTS Provider: {self.tts_provider}\n"
         )
 
         # Generate the preview and save to cache
         try:
+            if self.tts_provider == "supertonic":
+                from abogen.tts_supertonic import SupertonicPipeline
 
-            # Enable voice formula support for preview
-            if "*" in self.voice:
-                loaded_voice = get_new_voice(self.backend, self.voice, self.use_gpu)
+                tts = SupertonicPipeline(
+                    sample_rate=self.sample_rate,
+                    lang=self.supertonic_language,
+                    total_steps=self.supertonic_total_steps,
+                )
+                loaded_voice = self.voice or "M1"
+                sample_text = "Hello, this is a sample of the selected voice."
+                audio_segments = []
+                for result in tts(
+                    sample_text,
+                    voice=loaded_voice,
+                    speed=self.speed,
+                    split_pattern=None,
+                ):
+                    audio_segments.append(result.audio)
             else:
-                loaded_voice = self.voice
-            sample_text = get_sample_voice_text(self.lang_code)
-            audio_segments = []
-            for result in self.backend(
-                sample_text, voice=loaded_voice, speed=self.speed, split_pattern=None
-            ):
-                audio_segments.append(result.audio)
+                # Set device based on use_gpu setting and platform
+                if self.use_gpu:
+                    if platform.system() == "Darwin" and platform.processor() == "arm":
+                        device = "mps"
+                    else:
+                        device = "cuda"
+                else:
+                    device = "cpu"
+
+                tts = self.kpipeline_class(
+                    lang_code=self.lang_code, repo_id="hexgrad/Kokoro-82M", device=device
+                )
+                if "*" in self.voice:
+                    loaded_voice = get_new_voice(tts, self.voice, self.use_gpu)
+                else:
+                    loaded_voice = self.voice
+                sample_text = get_sample_voice_text(self.lang_code)
+                audio_segments = []
+                for result in tts(
+                    sample_text, voice=loaded_voice, speed=self.speed, split_pattern=None
+                ):
+                    audio_segments.append(result.audio)
+
             if audio_segments:
-                audio = np.concatenate(audio_segments)
-                # Save directly to the cache path
-                sf.write(self.cache_path, audio, 24000)
+                audio = self.np_module.concatenate(audio_segments)
+                sf.write(self.cache_path, audio, self.sample_rate)
                 self.temp_wav = self.cache_path
             self.finished.emit()
         except Exception as e:

abogen/pyqt/gui.py

@@ -9,6 +9,7 @@ from abogen.pyqt.queue_manager_gui import QueueManager
 from abogen.pyqt.queued_item import QueuedItem
 import abogen.hf_tracker as hf_tracker
 import hashlib  # Added for cache path generation
+from abogen.tts_supertonic import SUPERTONIC_AVAILABLE_LANGS, DEFAULT_SUPERTONIC_VOICES
 from PyQt6.QtWidgets import (
     QApplication,
     QWidget,
@@ -82,11 +83,13 @@ from abogen.constants import (
     GITHUB_URL,
     PROGRAM_DESCRIPTION,
     LANGUAGE_DESCRIPTIONS,
+    VOICES_INTERNAL,
+    KOKORO_LANG_TO_COUNTRY,
+    SUPERTONIC_LANG_TO_COUNTRY,
     SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION,
     COLORS,
     SUBTITLE_FORMATS,
 )
-from abogen.tts_backend_registry import get_metadata
 import threading
 from abogen.pyqt.voice_formula_gui import VoiceFormulaDialog
 from abogen.voice_profiles import load_profiles
@@ -970,6 +973,9 @@ class abogen(QWidget):
         self.fix_nonstandard_punctuation = self.config.get(
             "fix_nonstandard_punctuation", False
         )
+        self.tts_provider_config = self.config.get("tts_provider", "kokoro")
+        self.supertonic_language_config = self.config.get("supertonic_language", "en")
+        self.supertonic_total_steps_config = self.config.get("supertonic_total_steps", 8)
         self._pending_close_event = None
         self.gpu_ok = False  # Initialize GPU availability status
 
@@ -1018,6 +1024,16 @@ class abogen(QWidget):
                 else:
                     self.mixed_voice_state = entry
                     self.selected_lang = entry[0][0] if entry and entry[0] else None
+        # Restore TTS provider and supertonic settings from config
+        provider_text = "Supertonic" if self.tts_provider_config == "supertonic" else "Kokoro"
+        idx_st = self.provider_combo.findText(provider_text)
+        if idx_st >= 0:
+            self.provider_combo.setCurrentIndex(idx_st)
+        self.st_lang_combo.setCurrentText(self.supertonic_language_config)
+        idx_steps = self.st_steps_combo.findData(self.supertonic_total_steps_config)
+        if idx_steps >= 0:
+            self.st_steps_combo.setCurrentIndex(idx_steps)
+
         if self.save_option == "Choose output folder" and self.selected_output_folder:
             self.save_path_label.setText(self.selected_output_folder)
             self.save_path_row_widget.show()
@@ -1107,6 +1123,53 @@ class abogen(QWidget):
         speed_layout.addWidget(self.speed_label)
         controls_layout.addLayout(speed_layout)
         self.speed_slider.valueChanged.connect(self.update_speed_label)
+
+        # TTS Provider selection
+        provider_layout = QHBoxLayout()
+        provider_layout.setSpacing(7)
+        provider_label = QLabel("TTS Engine:", self)
+        provider_layout.addWidget(provider_label)
+        self.provider_combo = QComboBox(self)
+        self.provider_combo.addItem("Kokoro", "kokoro")
+        self.provider_combo.addItem("Supertonic", "supertonic")
+        self.provider_combo.setStyleSheet("QComboBox { min-height: 20px; padding: 6px 12px; }")
+        self.provider_combo.currentIndexChanged.connect(self.on_provider_changed)
+        provider_layout.addWidget(self.provider_combo)
+        controls_layout.addLayout(provider_layout)
+
+        # Supertonic-specific controls (language + steps), hidden by default
+        self.supertonic_row = QWidget()
+        supertonic_row_layout = QHBoxLayout(self.supertonic_row)
+        supertonic_row_layout.setContentsMargins(0, 0, 0, 0)
+        supertonic_row_layout.setSpacing(7)
+
+        st_lang_label = QLabel("Language:", self)
+        supertonic_row_layout.addWidget(st_lang_label)
+        self.st_lang_combo = QComboBox(self)
+        for code in SUPERTONIC_AVAILABLE_LANGS:
+            country_code = SUPERTONIC_LANG_TO_COUNTRY.get(code, code)
+            flag = get_resource_path("abogen.assets.flags", f"{country_code}.png")
+            icon_st = QIcon(flag) if flag and os.path.exists(flag) else QIcon()
+            self.st_lang_combo.addItem(icon_st, code, code)
+        self.st_lang_combo.setCurrentText("en")
+        self.st_lang_combo.setStyleSheet("QComboBox { min-height: 20px; padding: 6px 12px; }")
+        self.st_lang_combo.currentTextChanged.connect(self._on_st_lang_changed)
+        supertonic_row_layout.addWidget(self.st_lang_combo)
+
+        st_steps_label = QLabel("Steps:", self)
+        supertonic_row_layout.addWidget(st_steps_label)
+        self.st_steps_combo = QComboBox(self)
+        for val in range(2, 16):
+            self.st_steps_combo.addItem(str(val), val)
+        self.st_steps_combo.setCurrentIndex(self.st_steps_combo.findData(8))
+        self.st_steps_combo.setStyleSheet("QComboBox { min-height: 20px; padding: 6px 12px; }")
+        self.st_steps_combo.currentIndexChanged.connect(self._on_st_steps_changed)
+        supertonic_row_layout.addWidget(self.st_steps_combo)
+
+        supertonic_row_layout.addStretch()
+        self.supertonic_row.hide()
+        controls_layout.addWidget(self.supertonic_row)
+
         # Voice selection
         voice_layout = QHBoxLayout()
         voice_layout.setSpacing(7)
@@ -1764,6 +1827,12 @@ class abogen(QWidget):
         Update the enabled state of subtitle options based on the selected language.
         For non-English languages, only sentence-based and line-based modes are supported.
         """
+        provider = self.provider_combo.currentData()
+        if provider == "supertonic":
+            self.subtitle_combo.setEnabled(False)
+            self.subtitle_format_combo.setEnabled(False)
+            return
+
         # Check if current file is a subtitle file
         is_subtitle_input = False
         if self.selected_file and self.selected_file.lower().endswith(
@@ -1823,6 +1892,48 @@ class abogen(QWidget):
         # Enable/disable subtitle options based on language
         self.update_subtitle_options_availability()
 
+    def on_provider_changed(self, index):
+        provider = self.provider_combo.itemData(index)
+        self.config["tts_provider"] = provider
+        save_config(self.config)
+        is_supertonic = provider == "supertonic"
+
+        # Show/hide Supertonic controls
+        self.supertonic_row.setVisible(is_supertonic)
+
+        # Update subtitles availability
+        self.update_subtitle_options_availability()
+
+        # Repopulate voice list
+        self.populate_profiles_in_voice_combo()
+
+        # Clear/reset mixed voice state when switching provider
+        if is_supertonic:
+            self.mixed_voice_state = None
+            self.btn_voice_formula_mixer.setEnabled(False)
+            self.voice_combo.setToolTip(
+                "Supertonic voices:\n"
+                "M1-M5 = Male voices\n"
+                "F1-F5 = Female voices"
+            )
+        else:
+            self.btn_voice_formula_mixer.setEnabled(True)
+            self.voice_combo.setToolTip(
+                "The first character represents the language:\n"
+                '"a" => American English\n"b" => British English\n"e" => Spanish\n"f" => French\n"h" => Hindi\n"i" => Italian\n"j" => Japanese\n"p" => Brazilian Portuguese\n"z" => Mandarin Chinese\nThe second character represents the gender:\n"m" => Male\n"f" => Female'
+            )
+
+    def _on_st_lang_changed(self, lang):
+        self.config["supertonic_language"] = lang
+        save_config(self.config)
+        if self.provider_combo.currentData() == "supertonic":
+            self.selected_lang = lang
+            self.update_subtitle_options_availability()
+
+    def _on_st_steps_changed(self):
+        self.config["supertonic_total_steps"] = self.st_steps_combo.currentData()
+        save_config(self.config)
+
     def on_voice_combo_changed(self, index):
         data = self.voice_combo.itemData(index)
         if isinstance(data, str) and data.startswith("profile:"):
@@ -1831,10 +1942,26 @@ class abogen(QWidget):
             from abogen.voice_profiles import load_profiles
 
             entry = load_profiles().get(pname, {})
-            # set mixed voices and language
             if isinstance(entry, dict):
-                self.mixed_voice_state = entry.get("voices", [])
-                self.selected_lang = entry.get("language")
+                entry_provider = str(entry.get("provider", "")).strip().lower()
+                if entry_provider == "supertonic":
+                    # Switch provider to Supertonic if not already
+                    if self.provider_combo.currentData() != "supertonic":
+                        self.provider_combo.setCurrentIndex(1)
+                    self.mixed_voice_state = None
+                    self.selected_lang = entry.get("language", self.st_lang_combo.currentText())
+                    # Sync supertonic controls from profile
+                    profile_steps = entry.get("total_steps")
+                    if profile_steps is not None:
+                        idx_steps = self.st_steps_combo.findData(int(profile_steps))
+                        if idx_steps >= 0:
+                            self.st_steps_combo.setCurrentIndex(idx_steps)
+                    profile_lang = entry.get("language")
+                    if profile_lang and profile_lang in SUPERTONIC_AVAILABLE_LANGS:
+                        self.st_lang_combo.setCurrentText(profile_lang)
+                else:
+                    self.mixed_voice_state = entry.get("voices", [])
+                    self.selected_lang = entry.get("language")
             else:
                 self.mixed_voice_state = entry
                 self.selected_lang = entry[0][0] if entry and entry[0] else None
@@ -1847,7 +1974,12 @@ class abogen(QWidget):
         else:
             self.mixed_voice_state = None
             self.selected_profile_name = None
-            self.selected_voice, self.selected_lang = data, data[0]
+            self.selected_voice = data
+            provider = self.provider_combo.currentData()
+            if provider == "supertonic":
+                self.selected_lang = self.st_lang_combo.currentText()
+            else:
+                self.selected_lang = data[0] if data else ""
             self.config["selected_voice"] = data
             if "selected_profile_name" in self.config:
                 del self.config["selected_profile_name"]
@@ -1866,19 +1998,40 @@ class abogen(QWidget):
     def populate_profiles_in_voice_combo(self):
         # preserve current voice or profile
         current = self.voice_combo.currentData()
+        provider = self.provider_combo.currentData()
         self.voice_combo.blockSignals(True)
         self.voice_combo.clear()
-        # re-add profiles
+        # re-add profiles matching current provider
         profile_icon = QIcon(get_resource_path("abogen.assets", "profile.png"))
-        for pname in load_profiles().keys():
-            self.voice_combo.addItem(profile_icon, pname, f"profile:{pname}")
+        for pname, entry in load_profiles().items():
+            entry_provider = ""
+            if isinstance(entry, dict):
+                entry_provider = str(entry.get("provider", "")).strip().lower()
+            if provider == "supertonic":
+                if entry_provider == "supertonic":
+                    self.voice_combo.addItem(profile_icon, pname, f"profile:{pname}")
+            else:
+                if entry_provider != "supertonic":
+                    self.voice_combo.addItem(profile_icon, pname, f"profile:{pname}")
         # re-add voices
-        for v in get_metadata("kokoro").voices:
-            icon = QIcon()
-            flag_path = get_resource_path("abogen.assets.flags", f"{v[0]}.png")
-            if flag_path and os.path.exists(flag_path):
-                icon = QIcon(flag_path)
-            self.voice_combo.addItem(icon, f"{v}", v)
+        if provider == "supertonic":
+            for v in DEFAULT_SUPERTONIC_VOICES:
+                icon = QIcon()
+                if v.startswith("F"):
+                    icon_path = get_resource_path("abogen.assets", "female.png")
+                else:
+                    icon_path = get_resource_path("abogen.assets", "male.png")
+                if icon_path and os.path.exists(icon_path):
+                    icon = QIcon(icon_path)
+                self.voice_combo.addItem(icon, f"{v}", v)
+        else:
+            for v in VOICES_INTERNAL:
+                icon = QIcon()
+                country_code = KOKORO_LANG_TO_COUNTRY.get(v[0], v[0])
+                flag_path = get_resource_path("abogen.assets.flags", f"{country_code}.png")
+                if flag_path and os.path.exists(flag_path):
+                    icon = QIcon(flag_path)
+                self.voice_combo.addItem(icon, f"{v}", v)
         # restore selection
         idx = -1
         if self.selected_profile_name:
@@ -2069,6 +2222,9 @@ class abogen(QWidget):
             save_base_path=save_base_path,
             save_chapters_separately=getattr(self, "save_chapters_separately", None),
             merge_chapters_at_end=getattr(self, "merge_chapters_at_end", None),
+            tts_provider=self.provider_combo.currentData(),
+            supertonic_language=self.st_lang_combo.currentText(),
+            supertonic_total_steps=self.st_steps_combo.currentData(),
         )
 
         # Prevent adding duplicate items to the queue
@@ -2212,6 +2368,15 @@ class abogen(QWidget):
                 self.config["replace_numerals"] = self.replace_numerals
                 self.config["fix_nonstandard_punctuation"] = self.fix_nonstandard_punctuation
 
+                # TTS provider settings
+                tts_provider = getattr(queued_item, "tts_provider", "kokoro")
+                self.provider_combo.setCurrentText("Supertonic" if tts_provider == "supertonic" else "Kokoro")
+                self.st_lang_combo.setCurrentText(getattr(queued_item, "supertonic_language", "en"))
+                steps_val = getattr(queued_item, "supertonic_total_steps", 8)
+                idx_steps = self.st_steps_combo.findData(steps_val)
+                if idx_steps >= 0:
+                    self.st_steps_combo.setCurrentIndex(idx_steps)
+
                 # Sync Voice/Profile in config
                 self.config["selected_voice"] = self.selected_voice
                 if "selected_profile_name" in self.config:
@@ -2234,6 +2399,8 @@ class abogen(QWidget):
             self.current_queue_index = 0  # Reset for next time
 
     def get_voice_formula(self) -> str:
+        if self.provider_combo.currentData() == "supertonic":
+            return self._get_supertonic_voice()
         if self.mixed_voice_state:
             formula_components = [
                 f"{name}*{weight}" for name, weight in self.mixed_voice_state
@@ -2243,6 +2410,8 @@ class abogen(QWidget):
             return self.selected_voice
 
     def get_selected_lang(self, voice_formula) -> str:
+        if self.provider_combo.currentData() == "supertonic":
+            return self.st_lang_combo.currentText()
         if self.selected_profile_name:
             from abogen.voice_profiles import load_profiles
 
@@ -2316,9 +2485,9 @@ class abogen(QWidget):
             file_size_str = "Unknown"
 
         # pipeline_loaded_callback remains unchanged
-        def pipeline_loaded_callback(backend, error):
+        def pipeline_loaded_callback(np_module, kpipeline_class, error):
             if error:
-                self.update_log((f"Error loading TTS backend: {error}", "red"))
+                self.update_log((f"Error loading numpy or KPipeline: {error}", "red"))
                 prevent_sleep_end()
                 return
 
@@ -2332,6 +2501,10 @@ class abogen(QWidget):
             # determine selected language: use profile setting if profile selected, else voice code
             selected_lang = self.get_selected_lang(voice_formula)
 
+            tts_provider = self.provider_combo.currentData()
+            supertonic_language = self.st_lang_combo.currentText()
+            supertonic_total_steps = self.st_steps_combo.currentData()
+
             self.conversion_thread = ConversionThread(
                 self.selected_file,
                 selected_lang,
@@ -2341,13 +2514,17 @@ class abogen(QWidget):
                 self.selected_output_folder,
                 subtitle_mode=actual_subtitle_mode,
                 output_format=self.selected_format,
-                backend=backend,
+                np_module=np_module,
+                kpipeline_class=kpipeline_class,
                 start_time=self.start_time,
                 total_char_count=self.char_count,
                 use_gpu=self.gpu_ok,
                 from_queue=from_queue,
-                save_base_path=self.displayed_file_path,  # Pass the save base path (original file for EPUB)
-            )  # Use gpu_ok status
+                save_base_path=self.displayed_file_path,
+                tts_provider=tts_provider,
+                supertonic_language=supertonic_language,
+                supertonic_total_steps=supertonic_total_steps,
+            )
             # Pass the displayed file path to the log_updated signal handler in ConversionThread
             self.conversion_thread.display_path = display_path
             # Pass the file size string
@@ -2424,22 +2601,15 @@ class abogen(QWidget):
             # Store gpu_ok status to use when creating the conversion thread
             self.gpu_ok = gpu_ok
             self.update_log((gpu_msg, gpu_ok))
-            self.update_log("Loading modules...")
-
-            # Determine device based on GPU availability
-            if gpu_ok:
-                if platform.system() == "Darwin" and platform.processor() == "arm":
-                    device = "mps"
-                else:
-                    device = "cuda"
+            tts_provider = self.provider_combo.currentData()
+            if tts_provider == "supertonic":
+                # Supertonic doesn't need KPipeline, call callback directly
+                import numpy as np
+                pipeline_loaded_callback(np, None, None)
             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()
+                self.update_log("Loading modules...")
+                load_thread = LoadPipelineThread(pipeline_loaded_callback)
+                load_thread.start()
 
         threading.Thread(target=gpu_and_load, daemon=True).start()
 
@@ -2752,9 +2922,32 @@ class abogen(QWidget):
                 "Open File Error", f"Could not open file:\n{e}"
             )
 
+    def _get_supertonic_voice(self) -> str:
+        """Resolve the effective Supertonic voice from the current combo selection."""
+        if self.selected_profile_name:
+            from abogen.voice_profiles import load_profiles
+            entry = load_profiles().get(self.selected_profile_name, {})
+            if isinstance(entry, dict):
+                return str(entry.get("voice", "M1"))
+            return "M1"
+        current_data = self.voice_combo.currentData()
+        if current_data and isinstance(current_data, str) and not current_data.startswith("profile:"):
+            return current_data
+        return "M1"
+
     def _get_preview_cache_path(self):
         """Generate the expected cache path for the current voice settings."""
         speed = self.speed_slider.value() / 100.0
+        provider = self.provider_combo.currentData()
+
+        if provider == "supertonic":
+            voice_to_cache = self._get_supertonic_voice()
+            lang_to_cache = self.st_lang_combo.currentText()
+            steps = self.st_steps_combo.currentData()
+            cache_dir = get_user_cache_path("preview_cache")
+            filename = f"st_{voice_to_cache}_{lang_to_cache}_steps{steps}_{speed:.2f}.wav"
+            return os.path.join(cache_dir, filename)
+
         voice_to_cache = ""
         lang_to_cache = ""
 
@@ -2859,6 +3052,13 @@ class abogen(QWidget):
         self.btn_voice_formula_mixer.setEnabled(False)  # Disable mixer button
         self.btn_start.setEnabled(False)  # Disable start button during preview
 
+        # For Supertonic, skip KPipeline loading and use SupertonicPipeline directly
+        if self.provider_combo.currentData() == "supertonic":
+            import numpy as np
+            self.loading_movie.start()
+            self._on_pipeline_loaded_for_preview(np, None, None)
+            return
+
         # Start loading animation - ensure signal connection is always active
         if hasattr(self, "loading_movie"):
             # Disconnect previous connections to avoid multiple connections
@@ -2875,27 +3075,18 @@ class abogen(QWidget):
             )
             self.loading_movie.start()
 
-        # 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"
+        def pipeline_loaded_callback(np_module, kpipeline_class, error):
+            self._on_pipeline_loaded_for_preview(np_module, kpipeline_class, error)
 
-        lang = self.selected_lang or "a"
-        load_thread = LoadPipelineThread(
-            self._on_pipeline_loaded_for_preview, lang_code=lang, device=device
-        )
+        load_thread = LoadPipelineThread(pipeline_loaded_callback)
         load_thread.start()
 
-    def _on_pipeline_loaded_for_preview(self, backend, error):
+    def _on_pipeline_loaded_for_preview(self, np_module, kpipeline_class, 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 TTS backend: {error}"
+                "Loading Error", f"Error loading numpy or KPipeline: {error}"
             )
             self.btn_preview.setIcon(self.play_icon)
             self.btn_preview.setEnabled(True)
@@ -2926,14 +3117,28 @@ class abogen(QWidget):
                     else None
                 )
         else:
-            lang = self.selected_voice[0]
-            voice = self.selected_voice
+            if self.provider_combo.currentData() == "supertonic":
+                voice = self._get_supertonic_voice()
+            else:
+                voice = self.selected_voice or ""
+
+        tts_provider = self.provider_combo.currentData()
+        supertonic_language = self.st_lang_combo.currentText()
+        supertonic_total_steps = self.st_steps_combo.currentData()
+
+        if tts_provider == "supertonic":
+            lang = supertonic_language
+        else:
+            lang = self.selected_voice[0] if self.selected_voice else ""
 
         # use same gpu/cpu logic as in conversion
         gpu_msg, gpu_ok = get_gpu_acceleration(self.use_gpu)
 
         self.preview_thread = VoicePreviewThread(
-            backend, lang, voice, speed, gpu_ok
+            np_module, kpipeline_class, lang, voice, speed, gpu_ok,
+            tts_provider=tts_provider,
+            supertonic_language=supertonic_language,
+            supertonic_total_steps=supertonic_total_steps,
         )
         self.preview_thread.finished.connect(self._play_preview_audio)
         self.preview_thread.error.connect(self._preview_error)

abogen/pyqt/predownload_gui.py

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

abogen/pyqt/queued_item.py

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

abogen/pyqt/voice_formula_gui.py

@@ -28,11 +28,12 @@ from PyQt6.QtWidgets import (
 from PyQt6.QtCore import Qt, QTimer, QPoint, QRect, QSize
 from PyQt6.QtGui import QPixmap, QIcon, QAction
 from abogen.constants import (
+    VOICES_INTERNAL,
     SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION,
     LANGUAGE_DESCRIPTIONS,
+    KOKORO_LANG_TO_COUNTRY,
     COLORS,
 )
-from abogen.tts_backend_registry import get_metadata
 import re
 import platform
 from abogen.utils import get_resource_path
@@ -179,7 +180,7 @@ class VoiceMixer(QWidget):
         layout.addWidget(QLabel(name), alignment=Qt.AlignmentFlag.AlignCenter)
 
         # Voice name label with gender icon
-        is_female = self.voice_name in get_metadata("kokoro").voices and self.voice_name[1] == "f"
+        is_female = self.voice_name in VOICES_INTERNAL and self.voice_name[1] == "f"
 
         # Icons layout (flag and gender)
         icons_layout = QHBoxLayout()
@@ -189,8 +190,9 @@ class VoiceMixer(QWidget):
         )  # Center the icons horizontally
 
         # Flag icon
+        country_code = KOKORO_LANG_TO_COUNTRY.get(language_code, language_code)
         flag_icon_path = get_resource_path(
-            "abogen.assets.flags", f"{language_code}.png"
+            "abogen.assets.flags", f"{country_code}.png"
         )
         gender_icon_path = get_resource_path(
             "abogen.assets", "female.png" if is_female else "male.png"
@@ -512,7 +514,8 @@ class VoiceFormulaDialog(QDialog):
         header_row.addWidget(QLabel("Language:"))
         self.language_combo = QComboBox()
         for code, desc in LANGUAGE_OPTIONS:
-            flag = get_resource_path("abogen.assets.flags", f"{code}.png")
+            country_code = KOKORO_LANG_TO_COUNTRY.get(code, code)
+            flag = get_resource_path("abogen.assets.flags", f"{country_code}.png")
             if flag and os.path.exists(flag):
                 self.language_combo.addItem(QIcon(flag), desc, code)
             else:
@@ -772,7 +775,7 @@ class VoiceFormulaDialog(QDialog):
 
     def add_voices(self, initial_state):
         first_enabled_voice = None
-        for voice in get_metadata("kokoro").voices:
+        for voice in VOICES_INTERNAL:
             language_code = voice[0]  # First character is the language code
             matching_voice = next(
                 (item for item in initial_state if item[0] == voice), None

abogen/subtitle_utils.py

@@ -466,7 +466,7 @@ def sanitize_name_for_os(name, is_folder=True):
 
 
 def validate_voice_name(voice_name):
-    """Validate voice name against available voices (case-insensitive).
+    """Validate voice name against VOICES_INTERNAL list (case-insensitive).
     Handles both single voices and formulas like 'af_heart*0.5 + am_echo*0.5'.
 
     Args:
@@ -477,10 +477,10 @@ def validate_voice_name(voice_name):
             - is_valid: True if all voices in the name/formula are valid
             - invalid_voice_name: The first invalid voice found, or None if all valid
     """
-    from abogen.tts_backend_registry import get_metadata
+    from abogen.constants import VOICES_INTERNAL
 
     # Create case-insensitive lookup set (done once per call)
-    voice_lookup_lower = {v.lower() for v in get_metadata("kokoro").voices}
+    voice_lookup_lower = {v.lower() for v in VOICES_INTERNAL}
     voice_name = voice_name.strip()
 
     # Check if it's a formula (contains *)
@@ -505,7 +505,7 @@ def split_text_by_voice_markers(text, default_voice):
     """Split text by voice markers, returning list of (voice, text) tuples.
 
     IMPORTANT: Returns the last voice used so it can persist across chapters.
-    Voice names are normalized to lowercase to match canonical voice names.
+    Voice names are normalized to lowercase to match VOICES_INTERNAL.
 
     Args:
         text: Text potentially containing <<VOICE:name>> markers
@@ -518,7 +518,7 @@ def split_text_by_voice_markers(text, default_voice):
             - valid_count: Number of valid voice markers processed
             - invalid_count: Number of invalid voice markers skipped
     """
-    from abogen.tts_backend_registry import get_metadata
+    from abogen.constants import VOICES_INTERNAL
 
     voice_splits = list(_VOICE_MARKER_SEARCH_PATTERN.finditer(text))
 
@@ -560,7 +560,7 @@ def split_text_by_voice_markers(text, default_voice):
                         # Find the canonical (lowercase) voice name
                         voice_part_lower = voice_part.strip().lower()
                         canonical_voice = next(
-                            (v for v in get_metadata("kokoro").voices if v.lower() == voice_part_lower),
+                            (v for v in VOICES_INTERNAL if v.lower() == voice_part_lower),
                             voice_part.strip()
                         )
                         normalized_parts.append(f"{canonical_voice}*{weight.strip()}")
@@ -569,7 +569,7 @@ def split_text_by_voice_markers(text, default_voice):
                 # Find the canonical (lowercase) voice name
                 voice_name_lower = voice_name.lower()
                 current_voice = next(
-                    (v for v in get_metadata("kokoro").voices if v.lower() == voice_name_lower),
+                    (v for v in VOICES_INTERNAL if v.lower() == voice_name_lower),
                     voice_name
                 )
             valid_markers += 1

abogen/tts_backend.py

@@ -1,89 +0,0 @@
-"""
-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

@@ -1,146 +0,0 @@
-"""
-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 is_registered(self, backend_id: str) -> bool:
-        """Return True if a backend with the given id is registered."""
-        return backend_id in self._backends
-
-    def list_backends(self) -> list[TTSBackendMetadata]:
-        """Return metadata for all registered backends."""
-        return list(self._backends.values())
-
-    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)
-
-    def resolve_backend_for_voice(
-        self,
-        spec: str,
-        fallback: str = "kokoro",
-    ) -> str:
-        """Determine which backend owns the given voice specification.
-
-        Resolution rules:
-        1. Empty spec -> fallback
-        2. Kokoro formula (contains '*' or '+') -> "kokoro"
-        3. Exact voice ID match against registered backends -> backend id
-        4. Unknown voice -> fallback
-        """
-        raw = str(spec or "").strip()
-        if not raw:
-            return fallback
-
-        if "*" in raw or "+" in raw:
-            return "kokoro"
-
-        upper = raw.upper()
-        for metadata in self._backends.values():
-            if upper in metadata.voices:
-                return metadata.id
-
-        return fallback
-
-
-_registry = TTSBackendRegistry()
-
-
-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)
-
-
-def is_registered_backend(backend_id: str) -> bool:
-    """Return True if *backend_id* is a registered TTS backend."""
-    import abogen.tts_backends  # noqa: F401  — triggers backend registration
-    return _registry.is_registered(backend_id)
-
-
-def resolve_backend_for_voice(
-    spec: str,
-    fallback: str = "kokoro",
-) -> str:
-    """Determine which backend owns the given voice specification.
-
-    Ensures all backends are registered by importing the tts_backends
-    package on first access.
-
-    Resolution rules:
-    1. Empty spec -> fallback
-    2. Kokoro formula (contains '*' or '+') -> "kokoro"
-    3. Exact voice ID match against registered backends -> backend id
-    4. Unknown voice -> fallback
-    """
-    import abogen.tts_backends  # noqa: F401  — triggers backend registration
-    return _registry.resolve_backend_for_voice(spec, fallback=fallback)

abogen/tts_backends/__init__.py

@@ -1,20 +0,0 @@
-"""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

@@ -1,179 +0,0 @@
-"""
-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.tts_backend import TTSBackendMetadata
-
-# Internal voice list — source of truth for Kokoro voices.
-# The rest of the project accesses voices via get_metadata("kokoro").voices.
-_VOICES_INTERNAL = [
-    "af_alloy",
-    "af_aoede",
-    "af_bella",
-    "af_heart",
-    "af_jessica",
-    "af_kore",
-    "af_nicole",
-    "af_nova",
-    "af_river",
-    "af_sarah",
-    "af_sky",
-    "am_adam",
-    "am_echo",
-    "am_eric",
-    "am_fenrir",
-    "am_liam",
-    "am_michael",
-    "am_onyx",
-    "am_puck",
-    "am_santa",
-    "bf_alice",
-    "bf_emma",
-    "bf_isabella",
-    "bf_lily",
-    "bm_daniel",
-    "bm_fable",
-    "bm_george",
-    "bm_lewis",
-    "ef_dora",
-    "em_alex",
-    "em_santa",
-    "ff_siwis",
-    "hf_alpha",
-    "hf_beta",
-    "hm_omega",
-    "hm_psi",
-    "if_sara",
-    "im_nicola",
-    "jf_alpha",
-    "jf_gongitsune",
-    "jf_nezumi",
-    "jf_tebukuro",
-    "jm_kumo",
-    "pf_dora",
-    "pm_alex",
-    "pm_santa",
-    "zf_xiaobei",
-    "zf_xiaoni",
-    "zf_xiaoxiao",
-    "zf_xiaoyi",
-    "zm_yunjian",
-    "zm_yunxi",
-    "zm_yunxia",
-    "zm_yunyang",
-]
-
-_KOKORO_METADATA = TTSBackendMetadata(
-    id="kokoro",
-    name="Kokoro",
-    description="Kokoro TTS engine",
-    voices=tuple(_VOICES_INTERNAL),
-)
-
-
-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_supertonic.py

@@ -4,8 +4,9 @@ import ast
 from dataclasses import dataclass
 import logging
 import math
+import os
 import re
-from typing import Any, Dict, Iterable, Iterator, List, Optional
+from typing import Any, Iterable, Iterator, Optional
 
 import numpy as np
 
@@ -15,14 +16,12 @@ 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,
-)
+SUPERTONIC_AVAILABLE_LANGS = [
+    "en", "ko", "ja", "ar", "bg", "cs", "da", "de", "el",
+    "es", "et", "fi", "fr", "hi", "hr", "hu", "id", "it",
+    "lt", "lv", "nl", "pl", "pt", "ro", "ru", "sk", "sl",
+    "sv", "tr", "uk", "vi", "na",
+]
 
 
 @dataclass
@@ -98,7 +97,7 @@ _UNSUPPORTED_CHARS_RE = re.compile(
 
 
 def _parse_unsupported_characters(error: BaseException) -> list[str]:
-    """Best-effort extraction of unsupported characters from SuperTonic errors."""
+    """Best-effort extraction of unsupported characters from Supertonic errors."""
 
     message = " ".join(
         str(part) for part in getattr(error, "args", ()) if part is not None
@@ -164,6 +163,7 @@ def _configure_supertonic_gpu() -> None:
     except Exception as exc:
         logger.warning("Could not configure supertonic GPU providers: %s", exc)
 
+SUPERTONIC_MAX_CHUNK_LENGTH = 500
 
 class SupertonicPipeline:
     """Minimal adapter that mimics Kokoro's pipeline iteration interface."""
@@ -174,11 +174,14 @@ class SupertonicPipeline:
         sample_rate: int,
         auto_download: bool = True,
         total_steps: int = 5,
-        max_chunk_length: int = 300,
+        max_chunk_length: int = SUPERTONIC_MAX_CHUNK_LENGTH,
+        lang: str = "en",
+        intra_op_num_threads: Optional[int] = None,
     ) -> None:
         self.sample_rate = int(sample_rate)
         self.total_steps = int(total_steps)
         self.max_chunk_length = int(max_chunk_length)
+        self.lang = str(lang or "en")
 
         # Configure GPU providers before importing TTS
         _configure_supertonic_gpu()
@@ -190,7 +193,8 @@ class SupertonicPipeline:
                 "Supertonic is not installed. Install it with `pip install supertonic`."
             ) from exc
 
-        self._tts = TTS(auto_download=auto_download)
+        threads = intra_op_num_threads if intra_op_num_threads is not None else os.cpu_count()
+        self._tts = TTS(auto_download=auto_download, intra_op_num_threads=threads)
 
     def __call__(
         self,
@@ -200,12 +204,14 @@ class SupertonicPipeline:
         speed: float,
         split_pattern: Optional[str] = None,
         total_steps: Optional[int] = None,
+        lang: Optional[str] = None,
     ) -> Iterator[SupertonicSegment]:
         voice_name = (voice or "").strip() or "M1"
         steps = int(total_steps) if total_steps is not None else self.total_steps
         steps = max(2, min(15, steps))
         speed_value = float(speed) if speed is not None else 1.0
         speed_value = max(0.7, min(2.0, speed_value))
+        language = str(lang or self.lang or "en")
 
         style = self._tts.get_voice_style(voice_name=voice_name)
         chunks = _split_text(
@@ -216,12 +222,13 @@ class SupertonicPipeline:
             removed: set[str] = set()
             last_exc: Exception | None = None
 
-            # SuperTonic can raise ValueError for unsupported characters; strip and retry.
+            # Supertonic can raise ValueError for unsupported characters; strip and retry.
             for attempt in range(3):
                 try:
                     wav, duration = self._tts.synthesize(
                         text=chunk_to_speak,
                         voice_style=style,
+                        lang=language,
                         total_steps=steps,
                         speed=speed_value,
                         max_chunk_length=self.max_chunk_length,
@@ -247,14 +254,14 @@ class SupertonicPipeline:
                     chunk_to_speak = sanitized
                     if not chunk_to_speak:
                         logger.warning(
-                            "SuperTonic: dropped a chunk after removing unsupported characters: %s",
+                            "Supertonic: dropped a chunk after removing unsupported characters: %s",
                             sorted(removed),
                         )
                         break
 
                     if attempt == 0:
                         logger.warning(
-                            "SuperTonic: removed unsupported characters %s and retried.",
+                            "Supertonic: removed unsupported characters %s and retried.",
                             sorted(removed),
                         )
             else:
@@ -282,111 +289,3 @@ 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

@@ -529,20 +529,21 @@ 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, lang_code="a", device="cpu"):
+    def __init__(self, callback):
         super().__init__()
         self.callback = callback
-        self.lang_code = lang_code
-        self.device = device
 
     def run(self):
         try:
-            from abogen.tts_backend_registry import create_backend
-
-            backend = create_backend(
-                "kokoro", lang_code=self.lang_code, device=self.device
-            )
-            self.callback(backend, None)
+            np_module, kpipeline_class = load_numpy_kpipeline()
+            self.callback(np_module, kpipeline_class, None)
         except Exception as e:
-            self.callback(None, str(e))
+            self.callback(None, None, str(e))

abogen/voice_cache.py

@@ -17,7 +17,7 @@ if LocalEntryNotFoundError is None:  # pragma: no cover - fallback for tests
         pass
 
 
-from abogen.tts_backend_registry import get_metadata
+from abogen.constants import VOICES_INTERNAL
 
 _CACHE_LOCK = threading.Lock()
 _CACHED_VOICES: Set[str] = set()
@@ -26,9 +26,8 @@ _BOOTSTRAPPED = False
 
 
 def _normalize_targets(voices: Optional[Iterable[str]]) -> Set[str]:
-    kokoro_voices = get_metadata("kokoro").voices
     if not voices:
-        return set(kokoro_voices)
+        return set(VOICES_INTERNAL)
     normalized: Set[str] = set()
     for voice in voices:
         if not voice:
@@ -36,7 +35,7 @@ def _normalize_targets(voices: Optional[Iterable[str]]) -> Set[str]:
         voice_id = str(voice).strip()
         if not voice_id:
             continue
-        if voice_id in kokoro_voices:
+        if voice_id in VOICES_INTERNAL:
             normalized.add(voice_id)
     return normalized
 

abogen/voice_formulas.py

@@ -1,7 +1,7 @@
 import re
 from typing import List, Tuple
 
-from abogen.tts_backend_registry import get_metadata
+from abogen.constants import VOICES_INTERNAL
 
 
 # Calls parsing and loads the voice to gpu or cpu
@@ -22,7 +22,6 @@ 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:
@@ -31,7 +30,7 @@ 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 kokoro_voices:
+        if voice_name not in VOICES_INTERNAL:
             raise ValueError(f"Unknown voice: {voice_name}")
         try:
             weight = float(raw_weight.strip())

abogen/voice_metadata.py

@@ -1,33 +0,0 @@
-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

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

abogen/webui/Dockerfile

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

abogen/webui/conversion_runner.py

@@ -20,7 +20,7 @@ import numpy as np
 import soundfile as sf
 import static_ffmpeg
 
-from abogen.tts_backend_registry import get_metadata, is_registered_backend, resolve_backend_for_voice
+from abogen.constants import VOICES_INTERNAL
 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 (
@@ -39,15 +39,14 @@ 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
 
@@ -57,26 +56,25 @@ SAMPLE_RATE = 24000
 
 
 def _supertonic_voice_from_spec(spec: Any, fallback: str) -> str:
-    """Normalize a voice specification for Supertonic.
-
-    This function only performs Supertonic-specific normalization (uppercase conversion
-    and fallback handling). Backend resolution is handled by the registry.
-    """
     raw = str(spec or "").strip()
     fallback_raw = str(fallback or "").strip()
 
-    # Normalize to uppercase for Supertonic voice IDs
-    upper = raw.upper() if raw else ""
+    # Supertonic voices are discrete IDs (M1/F3/...). If we see a Kokoro mix
+    # formula (contains '*' or '+'), ignore it and fall back to a safe voice.
+    if not raw or "*" in raw or "+" in raw:
+        raw = fallback_raw
+    if not raw or "*" in raw or "+" in raw:
+        raw = "M1"
 
-    # If empty or contains formula characters, use fallback
-    if not upper or "*" in upper or "+" in upper:
-        upper = fallback_raw.upper() if fallback_raw else ""
+    upper = raw.upper()
+    if upper in DEFAULT_SUPERTONIC_VOICES:
+        return upper
 
-    # If still empty, use default Supertonic voice
-    if not upper or "*" in upper or "+" in upper:
-        upper = "M1"
+    fallback_upper = fallback_raw.upper() if fallback_raw else ""
+    if fallback_upper in DEFAULT_SUPERTONIC_VOICES:
+        return fallback_upper
 
-    return upper
+    return "M1"
 
 
 def _split_speaker_reference(value: Any) -> tuple[Optional[str], str]:
@@ -120,7 +118,15 @@ def _formula_from_kokoro_entry(entry: Mapping[str, Any]) -> str:
 
 
 def _infer_provider_from_spec(value: Any, fallback: str = "kokoro") -> str:
-    return resolve_backend_for_voice(str(value or ""), fallback=fallback)
+    raw = str(value or "").strip()
+    if not raw:
+        return fallback
+    upper = raw.upper()
+    if upper in DEFAULT_SUPERTONIC_VOICES:
+        return "supertonic"
+    if "*" in raw or "+" in raw:
+        return "kokoro"
+    return fallback
 
 
 class _JobCancelled(Exception):
@@ -569,7 +575,7 @@ def _spec_to_voice_ids(spec: Any) -> Set[str]:
             return set(extract_voice_ids(text))
         except ValueError:
             return set()
-    if text in get_metadata("kokoro").voices:
+    if text in VOICES_INTERNAL:
         return {text}
     return set()
 
@@ -633,7 +639,7 @@ 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(get_metadata("kokoro").voices)
+    voices.update(VOICES_INTERNAL)
     return voices
 
 
@@ -1567,7 +1573,7 @@ def run_conversion_job(job: Job) -> None:
         def get_pipeline(provider: str) -> Any:
             nonlocal kokoro_cache_ready
             provider_norm = str(provider or "kokoro").strip().lower() or "kokoro"
-            if not is_registered_backend(provider_norm):
+            if provider_norm not in {"kokoro", "supertonic"}:
                 provider_norm = "kokoro"
 
             existing = pipelines.get(provider_norm)
@@ -1575,11 +1581,10 @@ def run_conversion_job(job: Job) -> None:
                 return existing
 
             if provider_norm == "supertonic":
-                pipelines[provider_norm] = create_backend(
-                    "supertonic",
+                pipelines[provider_norm] = SupertonicPipeline(
                     sample_rate=SAMPLE_RATE,
                     auto_download=True,
-                    total_steps=int(getattr(job, "supertonic_total_steps", 5) or 5),
+                    total_steps=int(getattr(job, "supertonic_total_steps", 8) or 8),
                 )
                 return pipelines[provider_norm]
 
@@ -1589,12 +1594,16 @@ def run_conversion_job(job: Job) -> None:
             device = "cpu"
             if not disable_gpu:
                 device = _select_device()
-            # Create KPipeline instance directly (conforms to TTSBackend protocol)
-            pipelines[provider_norm] = create_backend(
-                "kokoro",
-                lang_code=job.language,
-                device=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
             if not kokoro_cache_ready:
                 _initialize_voice_cache(job)
                 kokoro_cache_ready = True
@@ -1609,7 +1618,7 @@ def run_conversion_job(job: Job) -> None:
                 provider = str(entry.get("provider") or "kokoro").strip().lower() or "kokoro"
                 if provider == "supertonic":
                     voice = str(entry.get("voice") or getattr(job, "voice", "M1") or "M1").strip() or "M1"
-                    steps = int(entry.get("total_steps") or getattr(job, "supertonic_total_steps", 5) or 5)
+                    steps = int(entry.get("total_steps") or getattr(job, "supertonic_total_steps", 8) or 8)
                     speed = float(entry.get("speed") or getattr(job, "speed", 1.0) or 1.0)
                     return "supertonic", _supertonic_voice_from_spec(voice, getattr(job, "voice", "M1")), speed, steps
                 formula = _formula_from_kokoro_entry(entry)
@@ -1625,7 +1634,7 @@ def run_conversion_job(job: Job) -> None:
             """Resolve a raw voice spec into (provider, resolved_spec, choice, speed, steps).
 
             For Kokoro formulas, `choice` will be a resolved voice tensor (via `voice_formulas`).
-            For SuperTonic, `choice` will be a valid SuperTonic voice id.
+            For Supertonic, `choice` will be a valid Supertonic voice id.
             """
 
             provider, resolved, speed, steps = resolve_voice_target(raw_spec)
@@ -1635,8 +1644,8 @@ def run_conversion_job(job: Job) -> None:
                 return provider, resolved, cached, speed, steps
 
             if provider == "kokoro":
-                kokoro_backend = get_pipeline("kokoro")
-                choice = _resolve_voice(kokoro_backend, resolved, job.use_gpu)
+                kokoro_pipeline = get_pipeline("kokoro")
+                choice = _resolve_voice(kokoro_pipeline, resolved, job.use_gpu)
             else:
                 choice = resolved
 
@@ -1765,8 +1774,8 @@ 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_backend = get_pipeline("kokoro")
-            voice_cache[f"kokoro:{base_voice_resolved}"] = _resolve_voice(kokoro_backend, base_voice_resolved, job.use_gpu)
+            kokoro_pipeline = get_pipeline("kokoro")
+            voice_cache[f"kokoro:{base_voice_resolved}"] = _resolve_voice(kokoro_pipeline, base_voice_resolved, job.use_gpu)
         processed_chars = 0
         subtitle_index = 1
         current_time = 0.0
@@ -1796,8 +1805,8 @@ 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:
-                    intro_voice_spec = get_default_voice("kokoro")
+                if not intro_voice_spec and VOICES_INTERNAL:
+                    intro_voice_spec = VOICES_INTERNAL[0]
 
                 if intro_voice_spec:
                     intro_provider, _, intro_voice_choice, intro_speed, intro_steps = resolve_voice_choice(
@@ -1848,11 +1857,11 @@ def run_conversion_job(job: Job) -> None:
                     voice=voice_name,
                     speed=float(speed_override if speed_override is not None else job.speed),
                     split_pattern=split_pattern,
-                    total_steps=int(supertonic_steps_override if supertonic_steps_override is not None else getattr(job, "supertonic_total_steps", 5)),
+                    total_steps=int(supertonic_steps_override if supertonic_steps_override is not None else getattr(job, "supertonic_total_steps", 8)),
                 )
             else:
-                kokoro_backend = get_pipeline("kokoro")
-                segment_iter = kokoro_backend(
+                kokoro_pipeline = get_pipeline("kokoro")
+                segment_iter = kokoro_pipeline(
                     normalized,
                     voice=voice_choice,
                     speed=float(speed_override if speed_override is not None else job.speed),
@@ -1941,8 +1950,8 @@ 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_backend = get_pipeline("kokoro")
-                    voice_choice = _resolve_voice(kokoro_backend, chapter_voice_resolved, job.use_gpu)
+                    kokoro_pipeline = get_pipeline("kokoro")
+                    voice_choice = _resolve_voice(kokoro_pipeline, chapter_voice_resolved, job.use_gpu)
                     voice_cache[chapter_cache_key] = voice_choice
             else:
                 voice_choice = chapter_voice_resolved
@@ -2086,9 +2095,9 @@ 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_backend = get_pipeline("kokoro")
+                                kokoro_pipeline = get_pipeline("kokoro")
                                 chunk_voice_choice = _resolve_voice(
-                                    kokoro_backend,
+                                    kokoro_pipeline,
                                     chunk_voice_resolved,
                                     job.use_gpu,
                                 )
@@ -2230,8 +2239,8 @@ 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:
-                outro_voice_spec = get_default_voice("kokoro")
+            if not outro_voice_spec and VOICES_INTERNAL:
+                outro_voice_spec = VOICES_INTERNAL[0]
 
             if outro_text and outro_voice_spec:
                 outro_start_time = current_time
@@ -2436,17 +2445,17 @@ 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 create_backend(
-            "supertonic",
+        return SupertonicPipeline(
             sample_rate=SAMPLE_RATE,
             auto_download=True,
-            total_steps=int(getattr(job, "supertonic_total_steps", 5) or 5),
+            total_steps=int(getattr(job, "supertonic_total_steps", 8) or 8),
         )
 
     device = "cpu"
     if not disable_gpu:
         device = _select_device()
-    return create_backend("kokoro", lang_code=job.language, device=device)
+    _np, KPipeline = load_numpy_kpipeline()
+    return KPipeline(lang_code=job.language, repo_id="hexgrad/Kokoro-82M", device=device)
 
 
 def _select_device() -> str:
@@ -2601,7 +2610,7 @@ def _build_ffmpeg_command(path: Path, fmt: str, metadata: Optional[Dict[str, str
 def _resolve_voice(pipeline, voice_spec: str, use_gpu: bool):
     if "*" in voice_spec:
         # Voice formulas are a Kokoro-only feature (they require a pipeline that can
-        # load individual Kokoro voices). When running with SuperTonic (or when the
+        # load individual Kokoro voices). When running with Supertonic (or when the
         # pipeline is otherwise unavailable), treat the spec as a plain string and
         # allow downstream provider-specific resolution to choose a safe fallback.
         if pipeline is None or not hasattr(pipeline, "load_single_voice"):

abogen/webui/debug_tts_runner.py

@@ -15,7 +15,7 @@ from abogen.normalization_settings import build_apostrophe_config
 from abogen.text_extractor import extract_from_path
 from abogen.voice_cache import ensure_voice_assets
 from abogen.webui.conversion_runner import SAMPLE_RATE, SPLIT_PATTERN, _select_device, _to_float32, _resolve_voice, _spec_to_voice_ids
-from abogen.tts_backend_registry import create_backend
+from abogen.utils import load_numpy_kpipeline
 
 
 _MARKER_RE = re.compile(re.escape(MARKER_PREFIX) + r"(?P<code>[A-Z0-9_]+)" + re.escape(MARKER_SUFFIX))
@@ -45,7 +45,8 @@ def _load_pipeline(language: str, use_gpu: bool) -> Any:
     device = "cpu"
     if use_gpu:
         device = _select_device()
-    return create_backend("kokoro", lang_code=language, device=device)
+    _np, KPipeline = load_numpy_kpipeline()
+    return KPipeline(lang_code=language, repo_id="hexgrad/Kokoro-82M", device=device)
 
 
 def _extract_cases_from_text(text: str) -> List[Tuple[str, str]]:

abogen/webui/routes/api.py

@@ -34,7 +34,6 @@ from abogen.normalization_settings import (
 )
 from abogen.llm_client import list_models, LLMClientError
 from abogen.kokoro_text_normalization import normalize_for_pipeline
-from abogen.tts_backend_registry import is_registered_backend
 from abogen.integrations.audiobookshelf import AudiobookshelfClient, AudiobookshelfConfig
 from abogen.integrations.calibre_opds import (
     CalibreOPDSClient,
@@ -64,7 +63,7 @@ def api_save_voice_profile() -> ResponseReturnValue:
     if profile is None:
         # Speaker Studio payload format
         provider = str(payload.get("provider") or "kokoro").strip().lower()
-        if not is_registered_backend(provider):
+        if provider not in {"kokoro", "supertonic"}:
             provider = "kokoro"
         if provider == "supertonic":
             profile = {
@@ -163,7 +162,7 @@ def api_voice_profiles_preview() -> ResponseReturnValue:
     formula = str(payload.get("formula") or "").strip()
     profile_name = str(payload.get("profile") or "").strip()
     provider = str(payload.get("tts_provider") or payload.get("provider") or "").strip().lower() or None
-    supertonic_total_steps = int(payload.get("supertonic_total_steps") or payload.get("total_steps") or settings.get("supertonic_total_steps") or 5)
+    supertonic_total_steps = int(payload.get("supertonic_total_steps") or payload.get("total_steps") or settings.get("supertonic_total_steps") or 8)
 
     voice_spec = ""
     resolved_provider = provider or "kokoro"
@@ -225,13 +224,13 @@ def api_speaker_preview() -> ResponseReturnValue:
     speed_value = payload.get("speed")
     speed = coerce_float(speed_value, 1.0)
     tts_provider = str(payload.get("tts_provider") or "").strip().lower()
-    supertonic_total_steps = int(payload.get("supertonic_total_steps") or 5)
+    supertonic_total_steps = int(payload.get("supertonic_total_steps") or 8)
     
     settings = load_settings()
     use_gpu = settings.get("use_gpu", False)
 
     base_spec, speaker_name = split_profile_spec(voice)
-    resolved_provider = tts_provider if is_registered_backend(tts_provider) else ""
+    resolved_provider = tts_provider if tts_provider in {"kokoro", "supertonic"} else ""
 
     if speaker_name:
         entry = normalize_profile_entry(load_profiles().get(speaker_name))
@@ -270,7 +269,7 @@ def api_speaker_preview() -> ResponseReturnValue:
             use_gpu=use_gpu
             ,
             tts_provider=resolved_provider,
-            supertonic_total_steps=supertonic_total_steps or int(settings.get("supertonic_total_steps") or 5),
+            supertonic_total_steps=supertonic_total_steps or int(settings.get("supertonic_total_steps") or 8),
             pronunciation_overrides=pronunciation_overrides,
             manual_overrides=manual_overrides,
             speakers=speakers,

abogen/webui/routes/settings.py

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

abogen/webui/routes/utils/form.py

@@ -7,7 +7,6 @@ from flask.typing import ResponseReturnValue
 
 from abogen.webui.service import PendingJob, JobStatus
 from abogen.webui.routes.utils.service import get_service
-from abogen.tts_backend_registry import is_registered_backend
 from abogen.webui.routes.utils.settings import (
     load_settings,
     coerce_bool,
@@ -33,7 +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.tts_backend_registry import get_default_voice
+from abogen.constants import VOICES_INTERNAL
 from abogen.speaker_configs import get_config
 from abogen.kokoro_text_normalization import normalize_roman_numeral_titles
 from dataclasses import dataclass
@@ -578,9 +577,9 @@ def apply_book_step_form(
     # NOTE: Do not auto-set a global TTS provider at the book level based on the
     # narrator defaults. Provider is resolved per-speaker/per-chunk from the voice
     # spec (e.g. "speaker:Name" for saved speakers, or a Kokoro mix formula).
-    # This enables mixed-provider conversions (e.g. narrator=SuperTonic, characters=Kokoro).
+    # This enables mixed-provider conversions (e.g. narrator=Supertonic, characters=Kokoro).
     provider_value = str(form.get("tts_provider") or "").strip().lower()
-    if is_registered_backend(provider_value):
+    if provider_value in {"kokoro", "supertonic"}:
         pending.tts_provider = provider_value
 
     # Determine the base speaker selection (saved speaker ref or raw voice).
@@ -617,8 +616,8 @@ def apply_book_step_form(
         custom_formula = ""
 
     base_voice_spec = resolved_default_voice or narrator_voice_raw
-    if not base_voice_spec:
-        base_voice_spec = get_default_voice("kokoro")
+    if not base_voice_spec and VOICES_INTERNAL:
+        base_voice_spec = VOICES_INTERNAL[0]
 
     voice_choice, resolved_language, selected_profile = resolve_voice_choice(
         pending.language,
@@ -797,8 +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:
-        base_voice = get_default_voice("kokoro")
+    if not base_voice and VOICES_INTERNAL:
+        base_voice = VOICES_INTERNAL[0]
     selected_speaker_config = (form.get("speaker_config") or "").strip()
     speaker_config_payload = get_config(selected_speaker_config) if selected_speaker_config else None
 
@@ -914,6 +913,15 @@ def build_pending_job_from_extraction(
         else:
             normalization_overrides[key] = default_val
 
+    provider_value = str(form.get("tts_provider") or "").strip().lower()
+    if provider_value not in {"kokoro", "supertonic"}:
+        provider_value = settings.get("tts_provider", "kokoro")
+    try:
+        total_steps = int(form.get("supertonic_total_steps", settings.get("supertonic_total_steps", 8)))
+        supertonic_steps = max(2, min(15, total_steps))
+    except (TypeError, ValueError):
+        supertonic_steps = int(settings.get("supertonic_total_steps", 8))
+
     pending = PendingJob(
         id=uuid.uuid4().hex,
         original_filename=original_name,
@@ -929,6 +937,8 @@ def build_pending_job_from_extraction(
         replace_single_newlines=replace_single_newlines,
         subtitle_format=subtitle_format,
         total_characters=total_chars,
+        tts_provider=provider_value,
+        supertonic_total_steps=supertonic_steps,
         save_chapters_separately=save_chapters_separately,
         merge_chapters_at_end=merge_chapters_at_end,
         separate_chapters_format=separate_chapters_format,

abogen/webui/routes/utils/preview.py

@@ -78,9 +78,10 @@ def get_preview_pipeline(language: str, device: str) -> Any:
         pipeline = _preview_pipelines.get(key)
         if pipeline is not None:
             return pipeline
-        from abogen.tts_backend_registry import create_backend
+        from abogen.utils import load_numpy_kpipeline
 
-        pipeline = create_backend("kokoro", lang_code=language, device=device)
+        _, KPipeline = load_numpy_kpipeline()
+        pipeline = KPipeline(lang_code=language, repo_id="hexgrad/Kokoro-82M", device=device)
         _preview_pipelines[key] = pipeline
         return pipeline
 
@@ -91,7 +92,7 @@ def generate_preview_audio(
     speed: float,
     use_gpu: bool,
     tts_provider: str = "kokoro",
-    supertonic_total_steps: int = 5,
+    supertonic_total_steps: int = 8,
     max_seconds: float = 8.0,
     pronunciation_overrides: Optional[Iterable[Mapping[str, Any]]] = None,
     manual_overrides: Optional[Iterable[Mapping[str, Any]]] = None,
@@ -136,9 +137,9 @@ def generate_preview_audio(
             normalized_text = source_text
 
     if provider == "supertonic":
-        from abogen.tts_backend_registry import create_backend
+        from abogen.tts_supertonic import SupertonicPipeline
 
-        pipeline = create_backend("supertonic", sample_rate=SAMPLE_RATE, auto_download=True, total_steps=supertonic_total_steps)
+        pipeline = SupertonicPipeline(sample_rate=SAMPLE_RATE, auto_download=True, total_steps=supertonic_total_steps)
         segments = pipeline(
             normalized_text,
             voice=voice_spec,
@@ -200,7 +201,7 @@ def synthesize_preview(
     speed: float,
     use_gpu: bool,
     tts_provider: str = "kokoro",
-    supertonic_total_steps: int = 5,
+    supertonic_total_steps: int = 8,
     max_seconds: float = 8.0,
     pronunciation_overrides: Optional[Iterable[Mapping[str, Any]]] = None,
     manual_overrides: Optional[Iterable[Mapping[str, Any]]] = None,

abogen/webui/routes/utils/service.py

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

abogen/webui/routes/utils/settings.py

@@ -6,8 +6,8 @@ from abogen.constants import (
     LANGUAGE_DESCRIPTIONS,
     SUBTITLE_FORMATS,
     SUPPORTED_SOUND_FORMATS,
+    VOICES_INTERNAL,
 )
-from abogen.tts_backend_registry import get_default_voice
 from abogen.normalization_settings import (
     DEFAULT_LLM_PROMPT,
     environment_llm_defaults,
@@ -174,8 +174,9 @@ 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": get_default_voice("kokoro"),
-        "supertonic_total_steps": 5,
+        "default_voice": VOICES_INTERNAL[0] if VOICES_INTERNAL else "",
+        "tts_provider": "kokoro",
+        "supertonic_total_steps": 8,
         "supertonic_speed": 1.0,
         "replace_single_newlines": False,
         "use_gpu": True,

abogen/webui/routes/utils/voice.py

@@ -17,10 +17,10 @@ from abogen.constants import (
     SUPPORTED_SOUND_FORMATS,
     SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION,
     SAMPLE_VOICE_TEXTS,
+    VOICES_INTERNAL,
 )
-from abogen.tts_backend_registry import get_metadata
 from abogen.speaker_configs import list_configs
-from abogen.tts_backend_registry import create_backend
+from abogen.utils import load_numpy_kpipeline
 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 get_metadata("kokoro").voices:
+    for voice_id in VOICES_INTERNAL:
         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": get_metadata("kokoro").voices,
+        "voices": VOICES_INTERNAL,
         "subtitle_formats": SUBTITLE_FORMATS,
         "supported_langs_for_subs": SUPPORTED_LANGUAGES_FOR_SUBTITLE_GENERATION,
         "output_formats": SUPPORTED_SOUND_FORMATS,
@@ -666,7 +666,7 @@ def resolve_voice_choice(
 
         # Provider-aware behavior:
         # - Kokoro profiles typically represent mixes (formula strings).
-        # - SuperTonic profiles represent a discrete voice id + settings.
+        # - Supertonic profiles represent a discrete voice id + settings.
         #   In that case, we return a speaker reference so downstream can
         #   resolve provider per-speaker and allow mixed-provider casting.
         if provider == "supertonic":
@@ -741,7 +741,8 @@ def get_preview_pipeline(language: str, device: str):
         pipeline = _preview_pipelines.get(key)
         if pipeline is not None:
             return pipeline
-        pipeline = create_backend("kokoro", lang_code=language, device=device)
+        _, KPipeline = load_numpy_kpipeline()
+        pipeline = KPipeline(lang_code=language, repo_id="hexgrad/Kokoro-82M", device=device)
         _preview_pipelines[key] = pipeline
         return pipeline
 

abogen/webui/routes/voices.py

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

abogen/webui/service.py

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

abogen/webui/templates/partials/new_job_step_book.html

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

abogen/webui/templates/settings.html

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

pyproject.toml

@@ -30,7 +30,7 @@ dependencies = [
   "pip",
   "kokoro>=0.9.4",
   "misaki[zh]>=0.9.4",
-  "supertonic>=0.1.0",
+  "supertonic>=1.3.1",
   "ebooklib>=0.19",
   "beautifulsoup4>=4.13.4",
   "spacy>=3.8.7,<4.0",
@@ -111,11 +111,11 @@ filterwarnings = [
 
 [project.optional-dependencies]
 # NVIDIA GPU (Windows) (CUDA 12.6)       # uv tool install abogen[cuda126]
-cuda126 = ["torch"]
+cuda126 = ["torch", "onnxruntime-gpu>=1.26.0"]
 # NVIDIA GPU (Windows) (CUDA 12.8)       # uv tool install abogen[cuda]
-cuda = ["torch"]
+cuda = ["torch", "onnxruntime-gpu>=1.26.0"]
 # NVIDIA GPU (Windows) (CUDA 13.0)       # uv tool install abogen[cuda130]
-cuda130 = ["torch"]
+cuda130 = ["torch", "onnxruntime-gpu>=1.26.0"]
 # AMD GPU (Linux) (ROCm 6.4)             # uv tool install abogen[rocm]
 rocm = ["torch", "pytorch-triton-rocm"]
 # Development dependencies               # uv tool install abogen[dev]

tests/test_conversion_voice_resolution.py

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

tests/test_epub_exporter.py

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

tests/test_kokoro_backend.py

@@ -1,216 +0,0 @@
-"""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 create_backend to capture input.
+    # with provider=supertonic and stub SupertonicPipeline to capture input.
     captured = {}
 
     class DummyPipeline:
@@ -30,16 +30,11 @@ def test_preview_applies_manual_override_before_normalization(monkeypatch):
             captured["text"] = text
             return iter(())
 
-    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)
+    monkeypatch.setitem(
+        __import__("sys").modules,
+        "abogen.tts_supertonic",
+        type("M", (), {"SupertonicPipeline": DummyPipeline}),
+    )
 
     try:
         preview.generate_preview_audio(

tests/test_tts_backend.py

@@ -1,314 +0,0 @@
-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
-
-
-class TestResolveBackendForVoice:
-    """Tests for the resolve_backend_for_voice method."""
-
-    def test_empty_spec_returns_fallback(self):
-        registry = TTSBackendRegistry()
-        assert registry.resolve_backend_for_voice("", fallback="kokoro") == "kokoro"
-        assert registry.resolve_backend_for_voice("", fallback="supertonic") == "supertonic"
-
-    def test_none_spec_returns_fallback(self):
-        registry = TTSBackendRegistry()
-        assert registry.resolve_backend_for_voice(None, fallback="kokoro") == "kokoro"
-
-    def test_kokoro_formula_with_star_returns_kokoro(self):
-        registry = TTSBackendRegistry()
-        assert registry.resolve_backend_for_voice("af_nova*0.7") == "kokoro"
-
-    def test_kokoro_formula_with_plus_returns_kokoro(self):
-        registry = TTSBackendRegistry()
-        assert registry.resolve_backend_for_voice("af_nova*0.7+am_liam*0.3") == "kokoro"
-
-    def test_kokoro_voice_id_resolves_to_kokoro(self):
-        registry = TTSBackendRegistry()
-        meta = TTSBackendMetadata(
-            id="kokoro",
-            name="Kokoro",
-            description="Kokoro TTS",
-            voices=("af_nova", "am_liam"),
-        )
-        registry.register(metadata=meta, factory=lambda: None)
-
-        assert registry.resolve_backend_for_voice("af_nova") == "kokoro"
-        assert registry.resolve_backend_for_voice("am_liam") == "kokoro"
-
-    def test_supertonic_voice_id_resolves_to_supertonic(self):
-        registry = TTSBackendRegistry()
-        meta = TTSBackendMetadata(
-            id="supertonic",
-            name="SuperTonic",
-            description="SuperTonic TTS",
-            voices=("M1", "M2", "F1", "F2"),
-        )
-        registry.register(metadata=meta, factory=lambda: None)
-
-        assert registry.resolve_backend_for_voice("M1") == "supertonic"
-        assert registry.resolve_backend_for_voice("F2") == "supertonic"
-
-    def test_unknown_voice_returns_fallback(self):
-        registry = TTSBackendRegistry()
-        meta = TTSBackendMetadata(
-            id="kokoro",
-            name="Kokoro",
-            description="Kokoro TTS",
-            voices=("af_nova",),
-        )
-        registry.register(metadata=meta, factory=lambda: None)
-
-        assert registry.resolve_backend_for_voice("unknown_voice") == "kokoro"
-        assert registry.resolve_backend_for_voice("unknown_voice", fallback="supertonic") == "supertonic"
-
-    def test_case_insensitive_matching(self):
-        registry = TTSBackendRegistry()
-        meta = TTSBackendMetadata(
-            id="supertonic",
-            name="SuperTonic",
-            description="SuperTonic TTS",
-            voices=("M1", "F1"),
-        )
-        registry.register(metadata=meta, factory=lambda: None)
-
-        assert registry.resolve_backend_for_voice("m1") == "supertonic"
-        assert registry.resolve_backend_for_voice("f1") == "supertonic"
-
-    def test_default_fallback_is_kokoro(self):
-        registry = TTSBackendRegistry()
-        assert registry.resolve_backend_for_voice("unknown") == "kokoro"
-
-    def test_multiple_backends_resolution(self):
-        registry = TTSBackendRegistry()
-        kokoro_meta = TTSBackendMetadata(
-            id="kokoro",
-            name="Kokoro",
-            description="Kokoro TTS",
-            voices=("af_nova",),
-        )
-        supertonic_meta = TTSBackendMetadata(
-            id="supertonic",
-            name="SuperTonic",
-            description="SuperTonic TTS",
-            voices=("M1",),
-        )
-        registry.register(metadata=kokoro_meta, factory=lambda: None)
-        registry.register(metadata=supertonic_meta, factory=lambda: None)
-
-        assert registry.resolve_backend_for_voice("af_nova") == "kokoro"
-        assert registry.resolve_backend_for_voice("M1") == "supertonic"
-
-    def test_global_wrapper_resolve_backend_for_voice(self):
-        from abogen.tts_backend_registry import resolve_backend_for_voice
-
-        # Test with empty spec
-        assert resolve_backend_for_voice("") == "kokoro"
-
-        # Test with formula
-        assert resolve_backend_for_voice("af_nova*0.7") == "kokoro"
-
-        # Test with a registered voice
-        assert resolve_backend_for_voice("af_nova") == "kokoro"
-        assert resolve_backend_for_voice("M1") == "supertonic"

tests/test_tts_supertonic_unsupported_chars.py

@@ -1,6 +1,6 @@
 import numpy as np
 
-from abogen.tts_backends.supertonic import SupertonicBackend, SupertonicPipeline
+from abogen.tts_supertonic import SupertonicPipeline
 
 
 class _DummyTTS:
@@ -26,23 +26,13 @@ class _DummyTTS:
         return audio, 0.05
 
 
-def _make_pipeline() -> SupertonicPipeline:
+def test_supertonic_pipeline_strips_unsupported_characters_and_retries():
+    # Avoid importing/initializing real supertonic by manually constructing the pipeline.
     pipeline = SupertonicPipeline.__new__(SupertonicPipeline)
     pipeline.sample_rate = 24000
     pipeline.total_steps = 5
     pipeline.max_chunk_length = 1000
     pipeline._tts = _DummyTTS()
-    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
@@ -53,56 +43,11 @@ def test_supertonic_pipeline_strips_unsupported_characters_and_retries():
 
 
 def test_supertonic_pipeline_drops_chunk_if_only_unsupported_characters():
-    pipeline = _make_pipeline()
+    pipeline = SupertonicPipeline.__new__(SupertonicPipeline)
+    pipeline.sample_rate = 24000
+    pipeline.total_steps = 5
+    pipeline.max_chunk_length = 1000
+    pipeline._tts = _DummyTTS()
 
     segs = list(pipeline("•", voice="M1", speed=1.0))
     assert segs == []
-
-
-# --- 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_cache.py

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

tests/test_voice_formula_resolution.py

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

tests/test_voice_metadata.py

@@ -1,233 +0,0 @@
-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"