app.py

import gradio as gr
import torch
import numpy as np
from transformers import VitsModel, AutoTokenizer

# Configuración del modelo
model_id = "facebook/mms-tts-yua"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

print(f"Cargando modelo en: {device}")
model = VitsModel.from_pretrained(model_id).to(device)
tokenizer = AutoTokenizer.from_pretrained(model_id)

def generate_tts(text):
    if not text.strip():
        return None
    
    inputs = tokenizer(text, return_tensors="pt").to(device)
    
    with torch.no_grad():
        output = model(**inputs)
    
    waveform = output.waveform[0].cpu().numpy()
    sample_rate = model.config.sampling_rate
    
    return (sample_rate, waveform)

# Interfaz de Gradio
demo = gr.Interface(
    fn=generate_tts,
    inputs=gr.Textbox(label="Escribe en Maya Yucateco", placeholder="Bix yanilech..."),
    outputs=gr.Audio(label="Audio Generado"),
    title="Maya Yucateco TTS (Facebook MMS)",
    description="Servidor de síntesis de voz para el idioma Maya Yucateco."
)

if __name__ == "__main__":
    demo.launch()