feat: add OpenAI-compatible Audio Speech API endpoint

infer.py

@@ -297,7 +297,31 @@ def maybe_print_voice_clone_text_chunks(
         print()
 
 
+def _patch_torchaudio_backend() -> None:
+    """Patch torchaudio to avoid the SoX backend, which segfaults on some systems."""
+    try:
+        import torchaudio
+        _original_load = torchaudio.load
+        _original_save = torchaudio.save
+
+        def _load_with_soundfile(uri, *args, backend=None, **kwargs):
+            if backend is None:
+                backend = "soundfile"
+            return _original_load(uri, *args, backend=backend, **kwargs)
+
+        def _save_with_soundfile(uri, src, sample_rate, *args, backend=None, **kwargs):
+            if backend is None:
+                backend = "soundfile"
+            return _original_save(uri, src, sample_rate, *args, backend=backend, **kwargs)
+
+        torchaudio.load = _load_with_soundfile
+        torchaudio.save = _save_with_soundfile
+    except ImportError:
+        pass
+
+
 def main(argv: Optional[Sequence[str]] = None) -> dict[str, object]:
+    _patch_torchaudio_backend()
     set_logging()
     args = parse_args(argv)
     if args.debug == 1:

moss_tts_nano_runtime.py

@@ -25,21 +25,15 @@
 
 _DEFAULT_VOICE_FILES: dict[str, tuple[str, str]] = {
     "Junhao": ("zh_1.wav", "Chinese male voice A"),
-    "Zhiming": ("zh_2.wav", "Chinese male voice B"),
-    "Weiguo": ("zh_5.wav", "Chinese male voice C"),
     "Xiaoyu": ("zh_3.wav", "Chinese female voice A"),
     "Yuewen": ("zh_4.wav", "Chinese female voice B"),
     "Lingyu": ("zh_6.wav", "Chinese female voice C"),
-    "Trump": ("en_1.wav", "Trump reference voice"),
     "Ava": ("en_2.wav", "English female voice A"),
     "Bella": ("en_3.wav", "English female voice B"),
     "Adam": ("en_4.wav", "English male voice A"),
-    "Nathan": ("en_5.wav", "English male voice B"),
-    "Sakura": ("jp_1.mp3", "Japanese female voice A"),
     "Yui": ("jp_2.wav", "Japanese female voice B"),
-    "Aoi": ("jp_3.wav", "Japanese female voice C"),
-    "Hina": ("jp_4.wav", "Japanese female voice D"),
-    "Mei": ("jp_5.wav", "Japanese female voice E"),
+    "男播音": ("zh_10.wav", "Chinese male broadcaster voice"),
+    "杨幂": ("zh_11.wav", "Chinese female voice YangMi style"),
 }
 
 DEFAULT_VOICE = "Junhao"

onnx_tts_runtime.py

@@ -490,9 +490,26 @@ def resolve_prompt_audio_codes(
             return self.encode_reference_audio(prompt_audio_path)
         resolved_voice = str(voice or self.list_builtin_voices()[0]["voice"])
         voice_row = next((item for item in self.list_builtin_voices() if item["voice"] == resolved_voice), None)
-        if voice_row is None:
-            raise ValueError(f"Built-in voice not found: {resolved_voice}")
-        return list(voice_row["prompt_audio_codes"])
+        if voice_row is not None:
+            return list(voice_row["prompt_audio_codes"])
+        # Fallback: try to find wav file in preset voices directory
+        from moss_tts_nano.defaults import DEFAULT_PROMPT_AUDIO_DIR
+        wav_candidates = [
+            DEFAULT_PROMPT_AUDIO_DIR / f"{resolved_voice}.wav",
+            DEFAULT_PROMPT_AUDIO_DIR / f"{resolved_voice}_reference.wav",
+        ]
+        # Also check the PyTorch preset map for wav filename
+        try:
+            from moss_tts_nano_runtime import _DEFAULT_VOICE_FILES
+            if resolved_voice in _DEFAULT_VOICE_FILES:
+                wav_candidates.insert(0, DEFAULT_PROMPT_AUDIO_DIR / _DEFAULT_VOICE_FILES[resolved_voice][0])
+        except ImportError:
+            pass
+        for wav_path in wav_candidates:
+            if wav_path.exists():
+                logging.info("ONNX voice %r not in builtin manifest, encoding from %s", resolved_voice, wav_path)
+                return self.encode_reference_audio(wav_path)
+        raise ValueError(f"Built-in voice not found: {resolved_voice}")
 
     def decode_full_audio_safe(self, generated_frames: list[list[int]]) -> np.ndarray:
         try: