诸公可知目前最牛逼的TTS免费开源项目是哪一个?没错,是Bert-vits2,没有之一。它是在本来已经极其强大的Vits项目中融入了Bert大模型,基本上解决了VITS的语气韵律问题,在效果非常出色的情况下训练的成本开销普通人也完全可以接受。

BERT的核心思想是通过在大规模文本语料上进行无监督预训练,学习到通用的语言表示,然后将这些表示用于下游任务的微调。相比传统的基于词嵌入的模型,BERT引入了双向上下文信息的建模,使得模型能够更好地理解句子中的语义和关系。

BERT的模型结构基于Transformer,它由多个编码器层组成。每个编码器层都有多头自注意力机制和前馈神经网络,用于对输入序列进行多层次的特征提取和表示学习。在预训练阶段,BERT使用了两种任务来学习语言表示:掩码语言模型(Masked Language Model,MLM)和下一句预测(Next Sentence Prediction,NSP)。通过这两种任务,BERT能够学习到上下文感知的词嵌入和句子级别的语义表示。

在实际应用中,BERT的预训练模型可以用于各种下游任务,如文本分类、命名实体识别、问答系统等。通过微调预训练模型,可以在特定任务上取得更好的性能,而无需从头开始训练模型。

BERT的出现对自然语言处理领域带来了重大影响,成为了许多最新研究和应用的基础。它在多个任务上取得了领先的性能,并促进了自然语言理解的发展。

本次让我们基于Bert-vits2项目来克隆渣渣辉和刘青云的声音,打造一款时下热搜榜一的“青岛啤酒”鬼畜视频。

语音素材和模型

首先我们需要渣渣辉和刘青云的原版音频素材,原版《扫毒》素材可以参考:https://www.bilibili.com/video/BV1R64y1F7SQ/。

将两个主角的声音单独提取出来,随后依次进行背景音和前景音的分离,声音降噪以及声音切片等操作,这些步骤之前已经做过详细介绍,请参见:民谣女神唱流行,基于AI人工智能so-vits库训练自己的音色模型(叶蓓/Python3.10)。 囿于篇幅,这里不再赘述。

做好素材的简单处理后,我们来克隆项目:

git clone https://github.com/Stardust-minus/Bert-VITS2

随后安装项目的依赖:

cd Bert-VITS2  
  
pip3 install -r requirements.txt

接着下载bert模型放入到项目的bert目录。

bert模型下载地址:

中:https://huggingface.co/hfl/chinese-roberta-wwm-ext-large  
日:https://huggingface.co/cl-tohoku/bert-base-japanese-v3/tree/main

语音标注

接着我们需要对已经切好分片的语音进行标注,这里我们使用开源库whisper,关于whisper请移步:闻其声而知雅意,M1 Mac基于PyTorch(mps/cpu/cuda)的人工智能AI本地语音识别库Whisper(Python3.10)

编写标注代码:

import whisper  
import os  
import json  
import torchaudio  
import argparse  
import torch  
  
lang2token = {  
            'zh': "ZH|",  
            'ja': "JP|",  
            "en": "EN|",  
        }  
def transcribe_one(audio_path):  
    # load audio and pad/trim it to fit 30 seconds  
    audio = whisper.load_audio(audio_path)  
    audio = whisper.pad_or_trim(audio)  
  
    # make log-Mel spectrogram and move to the same device as the model  
    mel = whisper.log_mel_spectrogram(audio).to(model.device)  
  
    # detect the spoken language  
    _, probs = model.detect_language(mel)  
    print(f"Detected language: {max(probs, key=probs.get)}")  
    lang = max(probs, key=probs.get)  
    # decode the audio  
    options = whisper.DecodingOptions(beam_size=5)  
    result = whisper.decode(model, mel, options)  
  
    # print the recognized text  
    print(result.text)  
    return lang, result.text  
if __name__ == "__main__":  
    parser = argparse.ArgumentParser()  
    parser.add_argument("--languages", default="CJ")  
    parser.add_argument("--whisper_size", default="medium")  
    args = parser.parse_args()  
    if args.languages == "CJE":  
        lang2token = {  
            'zh': "ZH|",  
            'ja': "JP|",  
            "en": "EN|",  
        }  
    elif args.languages == "CJ":  
        lang2token = {  
            'zh': "ZH|",  
            'ja': "JP|",  
        }  
    elif args.languages == "C":  
        lang2token = {  
            'zh': "ZH|",  
        }  
    assert (torch.cuda.is_available()), "Please enable GPU in order to run Whisper!"  
    model = whisper.load_model(args.whisper_size)  
    parent_dir = "./custom_character_voice/"  
    speaker_names = list(os.walk(parent_dir))[0][1]  
    speaker_annos = []  
    total_files = sum([len(files) for r, d, files in os.walk(parent_dir)])  
    # resample audios  
    # 2023/4/21: Get the target sampling rate  
    with open("./configs/config.json", 'r', encoding='utf-8') as f:  
        hps = json.load(f)  
    target_sr = hps['data']['sampling_rate']  
    processed_files = 0  
    for speaker in speaker_names:  
        for i, wavfile in enumerate(list(os.walk(parent_dir + speaker))[0][2]):  
            # try to load file as audio  
            if wavfile.startswith("processed_"):  
                continue  
            try:  
                wav, sr = torchaudio.load(parent_dir + speaker + "/" + wavfile, frame_offset=0, num_frames=-1, normalize=True,  
                                          channels_first=True)  
                wav = wav.mean(dim=0).unsqueeze(0)  
                if sr != target_sr:  
                    wav = torchaudio.transforms.Resample(orig_freq=sr, new_freq=target_sr)(wav)  
                if wav.shape[1] / sr > 20:  
                    print(f"{wavfile} too long, ignoring\n")  
                save_path = parent_dir + speaker + "/" + f"processed_{i}.wav"  
                torchaudio.save(save_path, wav, target_sr, channels_first=True)  
                # transcribe text  
                lang, text = transcribe_one(save_path)  
                if lang not in list(lang2token.keys()):  
                    print(f"{lang} not supported, ignoring\n")  
                    continue  
                #text = "ZH|" + text + "\n"  
                text = lang2token[lang] + text + "\n"  
                speaker_annos.append(save_path + "|" + speaker + "|" + text)  
                  
                processed_files += 1  
                print(f"Processed: {processed_files}/{total_files}")  
            except:  
                continue

标注后,会生成切片语音对应文件:

./genshin_dataset/ying/vo_dialog_DPEQ003_raidenEi_01.wav|ying|ZH|神子…臣民对我的畏惧…  
./genshin_dataset/ying/vo_dialog_DPEQ003_raidenEi_02.wav|ying|ZH|我不会那么做…  
./genshin_dataset/ying/vo_dialog_SGLQ002_raidenEi_01.wav|ying|ZH|不用着急,好好挑选吧,我就在这里等着。  
./genshin_dataset/ying/vo_dialog_SGLQ003_raidenEi_01.wav|ying|ZH|现在在做的事就是「留影」…  
./genshin_dataset/ying/vo_dialog_SGLQ003_raidenEi_02.wav|ying|ZH|嗯,不错,又学到新东西了。快开始吧。

说白了,就是通过whisper把人物说的话先转成文字,并且生成对应的音标:

./genshin_dataset/ying/vo_dialog_DPEQ003_raidenEi_01.wav|ying|ZH|神子…臣民对我的畏惧…|_ sh en z i0 … ch en m in d ui w o d e w ei j v … _|0 2 2 5 5 0 2 2 2 2 4 4 3 3 5 5 4 4 4 4 0 0|1 2 2 1 2 2 2 2 2 2 2 1 1  
./genshin_dataset/ying/vo_dialog_DPEQ003_raidenEi_02.wav|ying|ZH|我不会那么做…|_ w o b u h ui n a m e z uo … _|0 3 3 2 2 4 4 4 4 5 5 4 4 0 0|1 2 2 2 2 2 2 1 1  
./genshin_dataset/ying/vo_dialog_SGLQ002_raidenEi_01.wav|ying|ZH|不用着急,好好挑选吧,我就在这里等着.|_ b u y ong zh ao j i , h ao h ao t iao x van b a , w o j iu z ai zh e l i d eng zh e . _|0 2 2 4 4 2 2 2 2 0 2 2 3 3 1 1 3 3 5 5 0 3 3 4 4 4 4 4 4 3 3 3 3 5 5 0 0|1 2 2 2 2 1 2 2 2 2 2 1 2 2 2 2 2 2 2 1 1  
./genshin_dataset/ying/vo_dialog_SGLQ003_raidenEi_01.wav|ying|ZH|现在在做的事就是'留影'…|_ x ian z ai z ai z uo d e sh ir j iu sh ir ' l iu y ing ' … _|0 4 4 4 4 4 4 4 4 5 5 4 4 4 4 4 4 0 2 2 3 3 0 0 0|1 2 2 2 2 2 2 2 2 1 2 2 1 1 1  
./genshin_dataset/ying/vo_dialog_SGLQ003_raidenEi_02.wav|ying|ZH|恩,不错,又学到新东西了.快开始吧.|_ EE en , b u c uo , y ou x ve d ao x in d ong x i l e . k uai k ai sh ir b a

最后,将标注好的文件转换为bert模型可读文件:

import torch  
from multiprocessing import Pool  
import commons  
import utils  
from tqdm import tqdm  
from text import cleaned_text_to_sequence, get_bert  
import argparse  
import torch.multiprocessing as mp  
  
  
def process_line(line):  
    rank = mp.current_process()._identity  
    rank = rank[0] if len(rank) > 0 else 0  
    if torch.cuda.is_available():  
        gpu_id = rank % torch.cuda.device_count()  
        device = torch.device(f"cuda:{gpu_id}")  
    wav_path, _, language_str, text, phones, tone, word2ph = line.strip().split("|")  
    phone = phones.split(" ")  
    tone = [int(i) for i in tone.split(" ")]  
    word2ph = [int(i) for i in word2ph.split(" ")]  
    word2ph = [i for i in word2ph]  
    phone, tone, language = cleaned_text_to_sequence(phone, tone, language_str)  
  
    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_path = wav_path.replace(".wav", ".bert.pt")  
  
    try:  
        bert = torch.load(bert_path)  
        assert bert.shape[-1] == len(phone)  
    except Exception:  
        bert = get_bert(text, word2ph, language_str, device)  
        assert bert.shape[-1] == len(phone)  
        torch.save(bert, bert_path)

模型训练

此时,打开项目目录中的config.json文件:

{  
  "train": {  
    "log_interval": 100,  
    "eval_interval": 100,  
    "seed": 52,  
    "epochs": 200,  
    "learning_rate": 0.0001,  
    "betas": [  
      0.8,  
      0.99  
    ],  
    "eps": 1e-09,  
    "batch_size": 4,  
    "fp16_run": false,  
    "lr_decay": 0.999875,  
    "segment_size": 16384,  
    "init_lr_ratio": 1,  
    "warmup_epochs": 0,  
    "c_mel": 45,  
    "c_kl": 1.0,  
    "skip_optimizer": true  
  },  
  "data": {  
    "training_files": "filelists/train.list",  
    "validation_files": "filelists/val.list",  
    "max_wav_value": 32768.0,  
    "sampling_rate": 44100,  
    "filter_length": 2048,  
    "hop_length": 512,  
    "win_length": 2048,  
    "n_mel_channels": 128,  
    "mel_fmin": 0.0,  
    "mel_fmax": null,  
    "add_blank": true,  
    "n_speakers": 1,  
    "cleaned_text": true,  
    "spk2id": {  
      "ying": 0  
    }  
  },  
  "model": {  
    "use_spk_conditioned_encoder": true,  
    "use_noise_scaled_mas": true,  
    "use_mel_posterior_encoder": false,  
    "use_duration_discriminator": true,  
    "inter_channels": 192,  
    "hidden_channels": 192,  
    "filter_channels": 768,  
    "n_heads": 2,  
    "n_layers": 6,  
    "kernel_size": 3,  
    "p_dropout": 0.1,  
    "resblock": "1",  
    "resblock_kernel_sizes": [  
      3,  
      7,  
      11  
    ],  
    "resblock_dilation_sizes": [  
      [  
        1,  
        3,  
        5  
      ],  
      [  
        1,  
        3,  
        5  
      ],  
      [  
        1,  
        3,  
        5  
      ]  
    ],  
    "upsample_rates": [  
      8,  
      8,  
      2,  
      2,  
      2  
    ],  
    "upsample_initial_channel": 512,  
    "upsample_kernel_sizes": [  
      16,  
      16,  
      8,  
      2,  
      2  
    ],  
    "n_layers_q": 3,  
    "use_spectral_norm": false,  
    "gin_channels": 256  
  }  
}

这里需要修改的参数是batch_size,通常情况下,数值和本地显存应该是一致的,但是最好还是改小一点,比如说一块4060的8G卡,最好batch_size是4,如果写8的话,还是有几率爆显存。

随后开始训练:

python3 train_ms.py

程序返回:

[W C:\actions-runner\_work\pytorch\pytorch\builder\windows\pytorch\torch\csrc\distributed\c10d\socket.cpp:601] [c10d] The client socket has failed to connect to [v3u.net]:65280 (system error: 10049 - 在其上下文中,该请求的地址无效。).  
[W C:\actions-runner\_work\pytorch\pytorch\builder\windows\pytorch\torch\csrc\distributed\c10d\socket.cpp:601] [c10d] The client socket has failed to connect to [v3u.net]:65280 (system error: 10049 - 在其上下文中,该请求的地址无效。).  
2023-10-23 15:36:08.293 | INFO     | data_utils:_filter:61 - Init dataset...  
100%|█████████████████████████████████████████████████████████████████████████████| 562/562 [00:00<00:00, 14706.57it/s]  
2023-10-23 15:36:08.332 | INFO     | data_utils:_filter:76 - skipped: 0, total: 562  
2023-10-23 15:36:08.333 | INFO     | data_utils:_filter:61 - Init dataset...  
100%|████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:00<?, ?it/s]  
2023-10-23 15:36:08.334 | INFO     | data_utils:_filter:76 - skipped: 0, total: 4  
Using noise scaled MAS for VITS2  
Using duration discriminator for VITS2  
INFO:OUTPUT_MODEL:Loaded checkpoint './logs\OUTPUT_MODEL\DUR_4600.pth' (iteration 33)  
INFO:OUTPUT_MODEL:Loaded checkpoint './logs\OUTPUT_MODEL\G_4600.pth' (iteration 33)  
INFO:OUTPUT_MODEL:Loaded checkpoint './logs\OUTPUT_MODEL\D_4600.pth' (iteration 33)

说明没有问题,训练日志存放在项目的logs目录下。

随后可以通过tensorboard来监控训练过程:

python3 -m tensorboard.main --logdir=logs\OUTPUT_MODEL

当loss趋于稳定说明模型已经收敛:

栩栩如生,音色克隆,Bert-vits2文字转语音打造鬼畜视频实践(Python3.10)-LMLPHP

模型推理

最后,我们就可以使用模型来生成我们想要听到的语音了:

python3 webui.py -m ./logs\OUTPUT_MODEL\G_47700.pth

注意参数为训练好的迭代模型,如果觉得当前迭代的模型可用,那么直接把pth和config.json拷贝出来即可,随后可以接着训练下一个模型。

结语

基于Bert-vits2打造的渣渣辉和刘青云音色的鬼畜视频已经上线到Youtube(B站),请检索:刘悦的技术博客,欢迎诸君品鉴和臻赏。

10-24 01:52