app.py

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

model = VitsModel.from_pretrained("facebook/mms-tts-chv")
tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-chv")

def tts_chuvash(text):
    inputs = tokenizer(text, return_tensors="pt")
    with torch.no_grad():
        speech = model(**inputs).waveform

    audio_np = speech.squeeze().cpu().numpy()
    sampling_rate = model.config.sampling_rate
    return (sampling_rate, audio_np)

demo = gr.Interface(
    fn=tts_chuvash,
    inputs=gr.Textbox(label="Текст чăвашла чĕлхи (Chuvash Cyrillic)"),
    outputs=gr.Audio(label="Аудио"),
    title="Чăваш TTS (Facebook VITS)",
    description="Модель TTS на чăваш чĕлхи с помощью MMS-TTS от Facebook (поддерживает только кириллицу)"
)

demo.launch()