infer.py

import argparse

import numpy
import numpy as np
import pydub
import torch

import commons
import utils
from models import SynthesizerTrn
from text import cleaned_text_to_sequence, get_bert
from text.cleaner import clean_text
from text.symbols import symbols

# 当前版本信息
latest_version = "2.0"


def get_net_g(model_path: str, device: str, hps):
    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)
    _ = net_g.eval()
    _ = utils.load_checkpoint(model_path, net_g, None, skip_optimizer=True)
    return net_g


def get_text(text, language_str, hps, device):
    # 在此处实现当前版本的get_text
    norm_text, phone, tone, word2ph = clean_text(text, language_str)
    phone, tone, language = cleaned_text_to_sequence(phone, tone, language_str)

    if hps.data.add_blank:
        phone = commons.intersperse(phone, 0)
        tone = commons.intersperse(tone, 0)
        language = commons.intersperse(language, 0)
        for i in range(len(word2ph)):
            word2ph[i] = word2ph[i] * 2
        word2ph[0] += 1
    bert = get_bert(norm_text, word2ph, language_str, device)
    del word2ph
    assert bert.shape[-1] == len(phone), phone

    if language_str == "ZH":
        bert = bert
        sh_bert = torch.zeros(1024, len(phone))
        en_bert = torch.zeros(1024, len(phone))
    elif language_str == "SH":
        bert = torch.zeros(1024, len(phone))
        sh_bert = bert
        en_bert = torch.zeros(1024, len(phone))
    elif language_str == "EN":
        bert = torch.zeros(1024, len(phone))
        sh_bert = torch.zeros(1024, len(phone))
        en_bert = bert
    else:
        raise ValueError("language_str should be ZH, SH or EN")

    assert bert.shape[-1] == len(phone), f"Bert seq len {bert.shape[-1]} != {len(phone)}"

    phone = torch.LongTensor(phone)
    tone = torch.LongTensor(tone)
    language = torch.LongTensor(language)
    return bert, sh_bert, en_bert, phone, tone, language


def infer(
        text,
        sdp_ratio,
        noise_scale,
        noise_scale_w,
        length_scale,
        sid,
        language,
        hps,
        net_g,
        device,
):
    bert, sh_bert, en_bert, phones, tones, lang_ids = get_text(text, language, hps, device)
    with torch.no_grad():
        x_tst = phones.to(device).unsqueeze(0)
        tones = tones.to(device).unsqueeze(0)
        lang_ids = lang_ids.to(device).unsqueeze(0)
        bert = bert.to(device).unsqueeze(0)
        sh_bert = sh_bert.to(device).unsqueeze(0)
        en_bert = en_bert.to(device).unsqueeze(0)
        x_tst_lengths = torch.LongTensor([phones.size(0)]).to(device)
        del phones
        speakers = torch.LongTensor([hps.data.spk2id[sid]]).to(device)
        audio = (
            net_g.infer(
                x_tst,
                x_tst_lengths,
                speakers,
                tones,
                lang_ids,
                bert,
                sh_bert,
                en_bert,
                sdp_ratio=sdp_ratio,
                noise_scale=noise_scale,
                noise_scale_w=noise_scale_w,
                length_scale=length_scale,
            )[0][0, 0]
            .data.cpu()
            .float()
            .numpy()
        )
        del x_tst, tones, lang_ids, bert, x_tst_lengths, speakers
        torch.cuda.empty_cache()
        return audio


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('--config', type=str, default='configs/config.json')
    parser.add_argument('--device', type=str, default='cpu')
    parser.add_argument('--model_path', type=str, default='models/G_73000.pth')
    parser.add_argument('--output', type=str, default='sample')
    args = parser.parse_args()

    hps = utils.get_hparams_from_file(args.config)
    net_g = get_net_g(args.model_path, device=args.device, hps=hps)

    # noise_scale = 0.667
    # noise_scale_w = 0.8
    # length_scale = 0.9

    sdp_ratio = 0
    noise_scale = 0.667
    noise_scale_w = 0.8
    length_scale = 0.9

    def do_sample(texts, sid, export_tag):
        audio_data = numpy.array([], dtype=numpy.float32)

        for (sub_text, language) in texts:
            sub_audio_data = infer(sub_text, sdp_ratio, noise_scale, noise_scale_w, length_scale, sid, language, hps, net_g, args.device)
            audio_data = np.concatenate((audio_data, sub_audio_data))

        audio_data = audio_data / numpy.abs(audio_data).max()
        audio_data = audio_data * 32767
        audio_data = audio_data.astype(numpy.int16)
        sound = pydub.AudioSegment(audio_data, frame_rate=hps.data.sampling_rate, sample_width=audio_data.dtype.itemsize, channels=1)
        export_filename = args.output + export_tag + sid + '.mp3'
        sound.export(export_filename, format='mp3')
        print(export_filename)

    text = [('我觉得有点贵。', 'ZH'), ('so expensive, can they?', 'EN'), ('哈巨，吃不消它。', 'SH')]

    do_sample(text, '小庄', '_1_')
    do_sample(text, '小嘟', '_1_')
    do_sample(text, 'Jane', '_1_')
    do_sample(text, '小贝', '_1_')
    do_sample(text, '老克勒', '_1_')
    do_sample(text, '美琳', '_1_')

    pass


if __name__ == "__main__":
    main()