app.py

import spaces
import gradio as gr
import torch
from TTS.api import TTS
import os
import argparse
import os
import sys
import tempfile
import librosa.display
import numpy as np

import torchaudio
import traceback
from TTS.demos.xtts_ft_demo.utils.formatter import format_audio_list
from TTS.demos.xtts_ft_demo.utils.gpt_train import train_gpt

from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import Xtts

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

device = "cuda"

tts = TTS("tts_models/multilingual/multi-dataset/xtts_bill_spa").to(device)
model_path = '/home/user/.local/share/tts/tts_models--multilingual--multi-dataset--xtts_bill_spa/model.pth'
config_path = '/home/user/.local/share/tts/tts_models--multilingual--multi-dataset--xtts_bill_spa/config.json'
vocab_path = '/home/user/.local/share/tts/tts_models--multilingual--multi-dataset--xtts_bill_spa/vocab.json'


def clear_gpu_cache():
    # clear the GPU cache
    if torch.cuda.is_available():
        torch.cuda.empty_cache()

XTTS_MODEL = None
def load_model(xtts_checkpoint, xtts_config, xtts_vocab):
    global XTTS_MODEL
    clear_gpu_cache()
    if not xtts_checkpoint or not xtts_config or not xtts_vocab:
        return "You need to run the previous steps or manually set the `XTTS checkpoint path`, `XTTS config path`, and `XTTS vocab path` fields !!"
    config = XttsConfig()
    config.load_json(xtts_config)
    XTTS_MODEL = Xtts.init_from_config(config)
    print("Loading XTTS model! ")
    XTTS_MODEL.load_checkpoint(config, checkpoint_path=xtts_checkpoint, vocab_path=xtts_vocab, use_deepspeed=False)
    if torch.cuda.is_available():
        XTTS_MODEL.cuda()

    print("Model Loaded!")

def run_tts(lang, tts_text, speaker_audio_file):
    if XTTS_MODEL is None or not speaker_audio_file:
        return "You need to run the previous step to load the model !!", None, None

    gpt_cond_latent, speaker_embedding = XTTS_MODEL.get_conditioning_latents(audio_path=speaker_audio_file, gpt_cond_len=XTTS_MODEL.config.gpt_cond_len, max_ref_length=XTTS_MODEL.config.max_ref_len, sound_norm_refs=XTTS_MODEL.config.sound_norm_refs)
    out = XTTS_MODEL.inference(
        text=tts_text,
        language=lang,
        gpt_cond_latent=gpt_cond_latent,
        speaker_embedding=speaker_embedding,
        temperature=XTTS_MODEL.config.temperature, # Add custom parameters here
        length_penalty=XTTS_MODEL.config.length_penalty,
        repetition_penalty=XTTS_MODEL.config.repetition_penalty,
        top_k=XTTS_MODEL.config.top_k,
        top_p=XTTS_MODEL.config.top_p,
    )

    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as fp:
        out["wav"] = torch.tensor(out["wav"]).unsqueeze(0)
        out_path = fp.name
        torchaudio.save(out_path, out["wav"], 24000)
    print("Speech generated !")

    return out_path, speaker_audio_file


@spaces.GPU(enable_queue=True)
def generate(text, audio):
    load_model(model_path, config_path, vocab_path)
    out_path, speaker_audio_file = run_tts(lang='es', tts_text=text, speaker_audio_file=audio)
    return out_path

demo = gr.Interface(
    fn=generate, 
    inputs=[gr.Textbox(label='Frase a generar'), gr.Audio(type='filepath', label='Voz de referencia')], 
    outputs=gr.Audio(type='filepath')
)

demo.launch()