app.py

from transformers import AutoProcessor, VitsModel
import torch
import gradio as gr
import scipy.io.wavfile

model_id = "facebook/mms-tts-bod"

processor = AutoProcessor.from_pretrained(model_id)
model = VitsModel.from_pretrained(model_id)
model.eval()

def tts_fn(text):
    inputs = processor(text=text, return_tensors="pt")
    with torch.no_grad():
        output = model(**inputs)
    audio = output.waveform.squeeze().numpy()
    sample_rate = model.config.sampling_rate
    return (sample_rate, audio)

demo = gr.Interface(fn=tts_fn, inputs=gr.Textbox(label="Nhập văn bản tiếng Tây Tạng"), outputs="audio")
demo.launch()