Update finetune_xtts_hf.py

finetune_xtts_hf.py

@@ -19,15 +19,21 @@ print("=== Descargando dataset sob111/voxpopuli_es_500 ===")
 ds = load_dataset("sob111/voxpopuli_es_500", split="train", token=HF_TOKEN)
 
 # Guardar metadata.json en el formato esperado por Coqui TTS
-os.makedirs("/tmp/voxpopuli_es_500", exist_ok=True)
+os.makedirs("/tmp/voxpopuli_es_500/wav_data", exist_ok=True)
 meta_file = "/tmp/voxpopuli_es_500/metadata.json"
 
 with open(meta_file, "w", encoding="utf-8") as f:
-    for sample in ds:
+    for i, sample in enumerate(ds):
+        # Guardar cada audio en wav_data
+        audio_path = f"/tmp/voxpopuli_es_500/wav_data/sample_{i}.wav"
+        array = sample["audio"]["array"]
+        import soundfile as sf
+        sf.write(audio_path, array, sample["audio"]["sampling_rate"])
+
         entry = {
-            "audio_file": sample["audio_file"],
-            "text": sample["text"],
-            "speaker_name": sample.get("speaker_name", "speaker")
+            "audio_file": audio_path,
+            "text": sample.get("text") or sample.get("sentence", ""),
+            "speaker_name": str(sample.get("speaker_id", "speaker"))
         }
         f.write(json.dumps(entry, ensure_ascii=False) + "\n")