app.py

import gradio as gr
import torch
import numpy as np
import sys
from vinorm import TTSnorm
from utils_audio import convert_to_wav

sys.path.append("vits")
import commons
import utils
from models import SynthesizerTrn
from text.symbols import symbols
from text import text_to_sequence
from scipy.io.wavfile import write
import logging

numba_logger = logging.getLogger("numba")
numba_logger.setLevel(logging.WARNING)


from resemblyzer import preprocess_wav, VoiceEncoder


device = "cpu"


def get_text(texts, hps):
    text_norm_list = []
    for text in texts.split(","):
        chunk_strings = []
        chunk_len = 30
        for i in range(0, len(text.split()), chunk_len):
            chunk = " ".join(text.split()[i : i + chunk_len])
            chunk_strings.append(chunk)
        for chunk_string in chunk_strings:
            text_norm = text_to_sequence(chunk_string, hps.data.text_cleaners)
            if hps.data.add_blank:
                text_norm = commons.intersperse(text_norm, 0)
            text_norm_list.append(torch.LongTensor(text_norm))
    return text_norm_list


def get_speaker_embedding(path):
    encoder = VoiceEncoder(device="cpu")
    path = convert_to_wav(path)
    wav = preprocess_wav(path)
    embed = encoder.embed_utterance(wav)
    return embed


class VoiceClone:
    def __init__(self, checkpoint_path):
        hps = utils.get_hparams_from_file("vivos.json")
        self.net_g = SynthesizerTrn(
            len(symbols),
            hps.data.filter_length // 2 + 1,
            hps.train.segment_size // hps.data.hop_length,
            n_speakers=hps.data.n_speakers,
            **hps.model
        ).to(device)
        _ = self.net_g.eval()

        _ = utils.load_checkpoint(checkpoint_path, self.net_g, None)

        self.hps = hps

    def infer(self, text, ref_audio):
        text_norm = TTSnorm(text)
        stn_tst_list = get_text(text_norm, self.hps)
        with torch.no_grad():
            audios = []
            for stn_tst in stn_tst_list:
                x_tst = stn_tst.to(device).unsqueeze(0)
                x_tst_lengths = torch.LongTensor([stn_tst.size(0)]).to(device)

                speaker_embedding = get_speaker_embedding(ref_audio)
                speaker_embedding = (
                    torch.FloatTensor(torch.from_numpy(speaker_embedding))
                    .unsqueeze(0)
                    .to(device)
                )

                audio = self.net_g.infer(
                    x_tst,
                    x_tst_lengths,
                    speaker_embedding=speaker_embedding,
                    noise_scale=0.667,
                    noise_scale_w=0.8,
                    length_scale=1,
                )

                audio = audio[0][0, 0].data.cpu().float().numpy()
                audios.append(audio)
                print(audio.shape)

            audios = np.concatenate(audios, axis=0)
            write(ref_audio.replace(".wav", "_clone.wav"), 22050, audios)
        return ref_audio.replace(".wav", "_clone.wav"), text_norm



object = VoiceClone("G_150000.pth")


def clonevoice(text: str, speaker_wav, file_upload, language: str):
    speaker_source = ""
    if speaker_wav is not None:
        speaker_source = speaker_wav
    elif file_upload is not None:
        speaker_source = file_upload
    else:
        speaker_source = "vsontung.wav"

    print(speaker_source)

    outfile, text_norm = object.infer(text, speaker_source)

    return [outfile, text_norm]


inputs = [
    gr.Textbox(
        label="Input",
        value="muốn ngồi ở một vị trí không ai ngồi được thì phải chịu cảm giác không ai chịu được",
        max_lines=3,
    ),
    gr.Audio(label="Speaker Wav", source="microphone", type="filepath"),
    gr.Audio(label="Speaker Wav", source="upload", type="filepath"),
    gr.Radio(label="Language", choices=["Vietnamese"], value="en"),
]
outputs = [gr.Audio(label="Output"), gr.TextArea()]

demo = gr.Interface(fn=clonevoice, inputs=inputs, outputs=outputs)

demo.launch(debug=True)