app.py

import spaces
import gradio as gr
import torch
from TTS.api import TTS
import os
import json
import scipy.io.wavfile as wavfile
import numpy as np

os.environ["COQUI_TOS_AGREED"] = "1"

device = "cuda"

tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to(device)

@spaces.GPU(enable_queue=True)
def clone(text, audio):
    # Generowanie mowy
    wav, alignment, text_info, _ = tts.tts(text=text, speaker_wav=audio, language="pl", return_dict=True)
    
    # Konwersja do numpy array i zapisanie jako plik WAV
    wav_np = np.array(wav)
    wavfile.write("./output.wav", 24000, (wav_np * 32767).astype(np.int16))
    
    # Przygotowanie informacji o fonemach
    phonemes_data = []
    cumulative_duration = 0
    for phoneme, duration in zip(text_info['phonemes'], alignment['durations']):
        start_time = cumulative_duration
        end_time = start_time + duration
        phonemes_data.append({
            "phoneme": phoneme,
            "start": float(start_time),
            "end": float(end_time),
            "duration": float(duration)
        })
        cumulative_duration = end_time
    
    # Zapisywanie informacji o fonemach do pliku JSON
    with open("./phonemes_info.json", "w", encoding="utf-8") as f:
        json.dump(phonemes_data, f, ensure_ascii=False, indent=2)
    
    return "./output.wav", "./phonemes_info.json"

# Interfejs Gradio
iface = gr.Interface(
    fn=clone, 
    inputs=[
        gr.Textbox(label='Tekst do syntezy'),
        gr.Audio(type='filepath', label='Plik audio z głosem referencyjnym')
    ], 
    outputs=[
        gr.Audio(type='filepath', label='Zsyntezowana mowa'),
        gr.File(label='Informacje o fonemach (JSON)')
    ],
    title='Klonowanie Głosu z Informacjami o Fonemach',
    theme=gr.themes.Base(primary_hue="teal", secondary_hue="teal", neutral_hue="slate")
)

iface.launch()