app.py

import os
import torch
import numpy as np
from fastapi import FastAPI, UploadFile, Form
from fastapi.responses import FileResponse
from TTS.api import TTS
import tempfile
import soundfile as sf

# Forzar consentimiento de licencia
os.environ["COQUI_TOS_AGREED"] = "1"

# Monkey patch temporal de torch.load
original_torch_load = torch.load

def patched_torch_load(f, *args, **kwargs):
    kwargs["weights_only"] = False
    return original_torch_load(f, *args, **kwargs)

torch.load = patched_torch_load

# Cargar modelo XTTS
tts = TTS(model_name="tts_models/multilingual/multi-dataset/xtts_v2")

app = FastAPI()

@app.post("/generate-audio/")
async def generate_audio(
    text: str = Form(...),
    language: str = Form(...),
    speaker_wav: UploadFile = Form(...)
):
    # Guardar archivo temporalmente
    with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
        contents = await speaker_wav.read()
        tmp.write(contents)
        tmp_path = tmp.name

    # Generar audio
    audio = tts.tts(
        text=text,
        speaker_wav=tmp_path,
        language=language,
        split_sentences=True,
        emotion="Angry"
    )

    # Guardar output
    out_path = tempfile.mktemp(suffix=".wav")
    sf.write(out_path, audio, 24000)

    return FileResponse(out_path, media_type="audio/wav", filename="output.wav")