app.py

import tempfile

import gradio as gr
import soundfile as sf

from transformers import (
    AutoTokenizer,
    VitsModel
)

import torch

MODEL_NAME = "facebook/mms-tts-nan"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = VitsModel.from_pretrained(MODEL_NAME)


def tts(text):

    inputs = tokenizer(text, return_tensors="pt")

    with torch.no_grad():
        output = model(**inputs)

    waveform = output.waveform.squeeze().cpu().numpy()

    file = tempfile.NamedTemporaryFile(
        suffix=".wav",
        delete=False
    )

    sf.write(
        file.name,
        waveform,
        model.config.sampling_rate
    )

    return file.name


demo = gr.Interface(
    fn=tts,
    inputs=gr.Textbox(label="輸入文字"),
    outputs=gr.Audio(type="filepath"),
    title="Taiwanese TTS"
)

demo.launch()