Add train_narrator_tts.py

train_narrator_tts.py

@@ -56,6 +56,7 @@ MAX_SECS = 12.0
 
 DATASET_NAME = "narrator"
 EXP_ARCH = "F5TTS_v1_Base"
+WAVS_DIR = Path("/tmp/f5_work/wavs")
 
 
 def _lib_root() -> Path:
@@ -87,12 +88,12 @@ def prepare_data() -> None:
             tar.extractall("/tmp/", filter="data")
         print(f"Extracted -> {lj_root}")
 
-    # Store audio as pre-decoded float arrays — NO hf_datasets.Audio feature.
-    # Using Audio feature triggers torchcodec at read time, which requires FFmpeg
-    # system libs that aren't available in HF Jobs. Storing plain dicts sidesteps this:
-    # row["audio"]["array"] is a Python list[float], row["audio"]["sampling_rate"] is int.
-    # F5-TTS CustomDataset calls torch.FloatTensor(row["audio"]["array"]) which accepts lists.
-    audio_data: list[dict] = []
+    WAVS_DIR.mkdir(parents=True, exist_ok=True)
+
+    # F5-TTS CustomDataset.__getitem__ accesses row["audio_path"] (string path) and
+    # row["text"]. It loads audio itself (soundfile/torchaudio) — no HF Audio feature,
+    # no torchcodec, no FFmpeg dependency. Just store plain string paths + text.
+    wav_paths: list[str] = []
     texts: list[str] = []
     durations: list[float] = []
     count = 0
@@ -122,7 +123,9 @@ def prepare_data() -> None:
                 import librosa
                 arr = librosa.resample(arr, orig_sr=orig_sr, target_sr=TARGET_SR)
 
-            audio_data.append({"array": arr.tolist(), "sampling_rate": TARGET_SR})
+            dest = WAVS_DIR / f"{clip_id}.wav"
+            sf.write(str(dest), arr, TARGET_SR)
+            wav_paths.append(str(dest))
             texts.append(text)
             durations.append(len(arr) / TARGET_SR)
             count += 1
@@ -135,7 +138,8 @@ def prepare_data() -> None:
     raw_dir = token_dir / "raw"
     raw_dir.mkdir(parents=True, exist_ok=True)
 
-    ds = hf_datasets.Dataset.from_dict({"audio": audio_data, "text": texts})
+    # No Audio feature, no cast_column — just plain string paths and text
+    ds = hf_datasets.Dataset.from_dict({"audio_path": wav_paths, "text": texts})
     ds.save_to_disk(str(raw_dir))
     print(f"Saved dataset -> {raw_dir}")
 
@@ -210,17 +214,6 @@ def push_model(ckpt_dir: Path) -> None:
 
 
 if __name__ == "__main__":
-    # torchcodec (datasets audio decoder) needs FFmpeg system libs — install before training
-    print("Installing FFmpeg for torchcodec audio decoding...")
-    r = subprocess.run(
-        ["apt-get", "install", "-y", "--no-install-recommends", "ffmpeg"],
-        capture_output=True, text=True,
-    )
-    if r.returncode == 0:
-        print("FFmpeg installed OK")
-    else:
-        print(f"apt-get failed (rc={r.returncode}): {r.stderr[-300:]}")
-
     prepare_data()
     ckpt_dir = fine_tune()
     push_model(ckpt_dir)