app.py

import tempfile
from scipy.io.wavfile import write
import gradio as gr
from transformers import VitsTokenizer, VitsModel, set_seed, pipeline
import torch
from datetime import datetime


model_name = "leks-forever/vits_lez_tts"  
tokenizer = VitsTokenizer.from_pretrained(model_name)
model = VitsModel.from_pretrained(model_name)

tts_pipeline = pipeline("text-to-speech", model=model_name)

new_sentence = '!.?'
in_sentence = ',-.:;'


def canonize_lez(text):
    for abruptive_letter in ['к', 'К', 'п', 'П', 'т', 'Т', 'ц', 'Ц', 'ч', 'Ч']:
        for abruptive_symbol in ['1', 'l', 'i', 'I', '|', 'ӏ', 'Ӏ', 'ӏ']:
            text = text.replace(abruptive_letter+abruptive_symbol, abruptive_letter+'Ӏ')
    return text


def tts_function(input_text, speaking_rate, noise_scale, add_pauses):
    device = 'cuda' if torch.cuda.is_available() else 'cpu'

    fixed_text = canonize_lez(input_text)

    if add_pauses:
            
        for symb in new_sentence:
            fixed_text = fixed_text.replace(symb, '   ')

        for symb in in_sentence:
            fixed_text = fixed_text.replace(symb, '  ')

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

    inputs = {key: value.to(device) for key, value in inputs.items()}

    model.to(device)
    model.eval()

    set_seed(900)

    model.speaking_rate = speaking_rate
    model.noise_scale = noise_scale

    with torch.no_grad():
        outputs = model(**inputs)
        waveform = outputs.waveform[0]

    waveform = waveform.detach().cpu().float().numpy()

    sampling_rate = model.config.sampling_rate

    timestamp = datetime.now().strftime("H%M%S")
    filename_part = input_text[:20].replace(' ', '_')  
    filename = f"{filename_part}_{timestamp}.wav"

    with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmpfile:
        write(filename, rate=sampling_rate, data=waveform)
        return filename


with gr.Blocks() as interface:
    gr.Markdown("### Text-to-speech Лезги ЧIалал")

    with gr.Row():
        with gr.Column():
            input_text = gr.Textbox(label="Введите текст на лезгинском", lines=4)
            add_pauses = gr.Checkbox(label="Добавить больше пауз у знаков препинания", value=False)
            speaking_rate = gr.Slider(label="Скорость речи (speaking_rate)", minimum=0, maximum=2, step=0.1, value=0.9)
            noise_scale = gr.Slider(label="Шум (noise_scale)", minimum=0, maximum=5, step=0.1, value=0)
            submit_button = gr.Button("Сгенерировать")
        
        with gr.Column():
            output_audio = gr.Audio(label="Аудио")

    submit_button.click(
        fn=tts_function,
        inputs=[input_text, speaking_rate, noise_scale, add_pauses],
        outputs=output_audio,
    )
    
interface.launch()