Add train_narrator_tts.py

train_narrator_tts.py

@@ -138,8 +138,8 @@ def prepare_data() -> None:
     raw_dir = token_dir / "raw"
     raw_dir.mkdir(parents=True, exist_ok=True)
 
-    # No Audio feature, no cast_column — just plain string paths and text
-    ds = hf_datasets.Dataset.from_dict({"audio_path": wav_paths, "text": texts})
+    # CustomDataset.__getitem__ also reads row["duration"] per sample
+    ds = hf_datasets.Dataset.from_dict({"audio_path": wav_paths, "text": texts, "duration": durations})
     ds.save_to_disk(str(raw_dir))
     print(f"Saved dataset -> {raw_dir}")