Add custom voice presets (F6, F7, M6) from JSON style embeddings

app/config.py

@@ -41,6 +41,12 @@ class Config:
     # Voice directory for custom voice samples
     VOICES_DIR = os.environ.get('SUPERTONIC3_VOICES_DIR', None)
 
+    # Voice presets directory (JSON style embeddings, tracked in git)
+    VOICE_PRESETS_DIR = os.environ.get(
+        'SUPERTONIC3_VOICE_PRESETS_DIR',
+        str(BASE_PATH / 'voice_presets'),
+    )
+
     DEFAULT_VOICE = os.environ.get('SUPERTONIC3_DEFAULT_VOICE', 'M1')
 
     # Built-in voice names (M1-M5 = male, F1-F5 = female)
@@ -49,6 +55,9 @@ class Config:
     # Supported extensions for custom voice samples
     VOICE_EXTENSIONS = ('.wav', '.mp3', '.flac')
 
+    # Supported extensions for voice preset files (JSON style embeddings)
+    VOICE_PRESET_EXTENSIONS = ('.json',)
+
     # Streaming default
     STREAM_DEFAULT = os.environ.get('SUPERTONIC3_STREAM_DEFAULT', 'false').lower() == 'true'
 

app/services/tts.py

@@ -1,22 +1,28 @@
+import json
 import os
 import re
 import time
 from pathlib import Path
 
+import numpy as np
+
 from app.config import Config
 from app.logging_config import get_logger
 
 logger = get_logger('tts')
 
 TTS = None
+STYLE_CLS = None
 
 
 def _ensure_supertonic():
-    global TTS
+    global TTS, STYLE_CLS
     if TTS is None:
         try:
             from supertonic import TTS as _TTS
+            from supertonic.core import Style as _Style
             TTS = _TTS
+            STYLE_CLS = _Style
         except ImportError as exc:
             raise ImportError('supertonic not found. Install with: pip install supertonic') from exc
 
@@ -26,6 +32,9 @@ class TTSService:
         self._tts = None
         self._model_loaded = False
         self._voices_dir: str | None = None
+        self._voice_presets_dir: str | None = None
+        self._preset_voices: dict[str, str] = {}
+        self._style_cache: dict[str, object] = {}
         # Pre-parse text into segments for simulated streaming
         self._sentence_splitter = re.compile(r'(?<=[.!?])\s+')
 
@@ -56,6 +65,8 @@ class TTSService:
             logger.error(f'Failed to load SuperTonic3 model: {e}')
             raise
 
+        self.discover_presets()
+
     def set_voices_dir(self, voices_dir: str | None) -> None:
         if voices_dir and os.path.isdir(voices_dir):
             self._voices_dir = voices_dir
@@ -66,16 +77,61 @@ class TTSService:
         else:
             self._voices_dir = None
 
+    def discover_presets(self) -> None:
+        self._preset_voices = {}
+        presets_dir = Config.VOICE_PRESETS_DIR
+        if not os.path.isdir(presets_dir):
+            logger.info(f'Voice presets directory not found: {presets_dir}')
+            return
+        self._voice_presets_dir = presets_dir
+        for f in sorted(Path(presets_dir).iterdir()):
+            if f.suffix.lower() in Config.VOICE_PRESET_EXTENSIONS and f.stem.isidentifier():
+                self._preset_voices[f.stem] = str(f)
+        if self._preset_voices:
+            logger.info(
+                f'Discovered {len(self._preset_voices)} voice presets: '
+                f'{", ".join(self._preset_voices.keys())}'
+            )
+
     def get_voice_style(self, voice_id: str):
         if not self.is_loaded:
             raise RuntimeError('Model not loaded. Call load_model() first.')
 
+        # Check if it's a cached preset voice
+        if voice_id in self._style_cache:
+            return self._style_cache[voice_id]
+
+        # Check if it's a known preset voice (load from JSON)
+        if voice_id in self._preset_voices:
+            return self._load_preset_style(voice_id)
+
         voice_name = self._resolve_voice(voice_id)
         t0 = time.time()
         style = self._tts.get_voice_style(voice_name=voice_name)
         logger.debug(f'Voice style loaded in {time.time() - t0:.2f}s: {voice_name}')
         return style
 
+    def _load_preset_style(self, voice_id: str):
+        filepath = self._preset_voices.get(voice_id)
+        if not filepath:
+            raise ValueError(f'Preset voice not found: {voice_id}')
+
+        t0 = time.time()
+        with open(filepath) as f:
+            data = json.load(f)
+
+        style_ttl = np.array(data['style_ttl']['data'], dtype=data['style_ttl']['type']).reshape(
+            data['style_ttl']['dims']
+        )
+        style_dp = np.array(data['style_dp']['data'], dtype=data['style_dp']['type']).reshape(
+            data['style_dp']['dims']
+        )
+
+        style = STYLE_CLS(style_ttl_onnx=style_ttl, style_dp_onnx=style_dp)
+        self._style_cache[voice_id] = style
+        logger.info(f'Loaded preset voice {voice_id} in {time.time() - t0:.2f}s')
+        return style
+
     def _resolve_voice(self, voice_id: str) -> str:
         voice_lower = voice_id.lower()
         if voice_lower in [v.lower() for v in Config.BUILTIN_VOICES]:
@@ -104,9 +160,13 @@ class TTSService:
                 False,
                 'HTTP/HTTPS URLs are not allowed for security reasons. Use hf:// for HuggingFace models.',
             )
+        # Check built-in voices
         voice_lower = voice_id.lower()
         if voice_lower in [v.lower() for v in Config.BUILTIN_VOICES]:
             return True, f'Built-in voice: {voice_id}'
+        # Check preset voices (JSON style embeddings)
+        if voice_id in self._preset_voices:
+            return True, f'Preset voice: {voice_id}'
         if self._voices_dir:
             p = Path(self._voices_dir) / voice_id
             if p.exists():
@@ -156,6 +216,8 @@ class TTSService:
         voices: list[dict] = []
         for name in Config.BUILTIN_VOICES:
             voices.append({'id': name, 'name': name, 'type': 'builtin'})
+        for name in sorted(self._preset_voices):
+            voices.append({'id': name, 'name': name, 'type': 'preset'})
         if self._voices_dir:
             for f in sorted(Path(self._voices_dir).iterdir()):
                 if f.suffix.lower() in Config.VOICE_EXTENSIONS: