app.py

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


class CustomFlagging(gr.FlaggingCallback):
    def setup(self, *args, **kwargs):
        pass  # Optional setup steps

    def flag(self, flag_data, flag_option=None, username=None):
        print(f"Аудио: {flag_data}, Сообщение: {flag_option}")


# Custom options
flagging_callback = CustomFlagging()
flagging_options = ["Хорошая озвучка", "Слышен механический треск", "Не совпадает произношение букв", 'Проглочены буквы']  # Customize options


# Load your fine-tuned model
model_name = "leks-forever/vits_lez_tts"  # Replace with your Hugging Face model name
tokenizer = VitsTokenizer.from_pretrained(model_name)
model = VitsModel.from_pretrained(model_name)

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


def tts_function(input_text):
    inputs = tokenizer(text=input_text, return_tensors="pt")
    speech = tts_pipeline(input_text)
    set_seed(900)  

    # make speech faster and more noisy
    model.speaking_rate = 0.9
    model.noise_scale = 0

    sampling_rate = speech["sampling_rate"]

    outputs = model(**inputs)
    waveform = outputs.waveform[0]
    waveform = waveform.detach().cpu().float().numpy()

    with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmpfile:
        write(tmpfile.name, rate=sampling_rate, data=waveform)
        return tmpfile.name  # Return the filepath


interface = gr.Interface(
    fn=tts_function,
    inputs=gr.Textbox(label="Введите текст на лезгинском"),
    outputs=gr.Audio(label="Аудио"),
    title="Text-to-speech Лезги ЧIалал",
    flagging_mode="auto",  # Enable the flagging button
)

# Launch the app
interface.launch()