Add train_narrator_tts.py

train_narrator_tts.py

@@ -164,7 +164,33 @@ def prepare_data() -> None:
         raise FileNotFoundError(f"Cannot find F5-TTS built-in vocab.txt — searched {builtin_vocab}")
 
 
+def _patch_torchaudio() -> None:
+    """Write sitecustomize.py into the venv so torchaudio.load uses soundfile.
+
+    torchaudio 2.6+ delegates load() to torchcodec which needs FFmpeg system libs.
+    sitecustomize.py runs at Python startup before any module import, so our patch
+    replaces torchaudio.load before F5-TTS's dataset.py ever calls it.
+    """
+    sc = _lib_root() / "site-packages" / "sitecustomize.py"
+    sc.write_text(
+        "try:\n"
+        "    import torchaudio, soundfile as _sf, torch as _torch, numpy as _np\n"
+        "    def _sf_load(uri, frame_offset=0, num_frames=-1, normalize=True,\n"
+        "                 channels_first=True, format=None, backend=None):\n"
+        "        arr, sr = _sf.read(str(uri), dtype='float32', always_2d=False)\n"
+        "        if arr.ndim == 1: arr = arr[_np.newaxis, :]\n"
+        "        elif channels_first: arr = arr.T\n"
+        "        return _torch.from_numpy(arr.copy()), sr\n"
+        "    torchaudio.load = _sf_load\n"
+        "    print('sitecustomize: torchaudio.load patched to use soundfile')\n"
+        "except Exception as _e: print(f'sitecustomize patch failed: {_e}')\n",
+        encoding="utf-8",
+    )
+    print(f"Installed torchaudio soundfile patch -> {sc}")
+
+
 def fine_tune() -> Path:
+    _patch_torchaudio()
     env = {**os.environ, "HF_TOKEN": HF_TOKEN}
 
     cmd = [