finetune_xtts_hf.py

import os
import subprocess
import json
from huggingface_hub import HfApi, HfFolder
from datasets import load_dataset

# === Configuración ===
HF_TOKEN = os.environ.get("HF_TOKEN")  # define en los Secrets del Space
HF_REPO_ID = "sob111/xttsv2-es-finetuned"  # tu repo de destino
OUTPUT_PATH = "/tmp/output_model"
CONFIG_PATH = "./config.json"

# === Guardar token de Hugging Face ===
print("=== Guardando token de Hugging Face ===")
HfFolder.save_token(HF_TOKEN)

# === Descargar dataset desde Hugging Face ===
# 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/wav_data", exist_ok=True)
# meta_file = "/tmp/voxpopuli_es_500/metadata.json"

# with open(meta_file, "w", encoding="utf-8") as f:
#     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": 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")

# print("✅ Metadata guardada en {meta_file}")

# === Entrenamiento XTTSv2 ===
print("=== Iniciando entrenamiento XTTSv2 ===")
# Importante: usa el binario oficial de train del paquete TTS
# y pasa sólo el config. No uses rutas a 'recipes/...'
try:
    subprocess.run(
        [
            "python",
            "-m", "TTS.bin.train",
            "--config_path", CONFIG_PATH,
            "--output_path", OUTPUT_PATH
        ],
        check=True
    )
except subprocess.CalledProcessError as e:
    raise RuntimeError("❌ El entrenamiento XTTSv2 falló. Revisa los logs anteriores.") from e

print("=== Entrenamiento finalizado ===")

# === Subir modelo fine-tune a Hugging Face ===
print("=== Subiendo modelo fine-tune a Hugging Face ===")
api = HfApi()
api.create_repo(repo_id=HF_REPO_ID, exist_ok=True, token=HF_TOKEN)
api.upload_folder(
    folder_path=OUTPUT_PATH,
    repo_id=HF_REPO_ID,
    repo_type="model",
    token=HF_TOKEN
)

print("✅ Fine-tuning completado y subido a {HF_REPO_ID}")