Skip to content

Latest commit

 

History

History
executable file
·
314 lines (223 loc) · 12.1 KB

README.md

File metadata and controls

executable file
·
314 lines (223 loc) · 12.1 KB
Raw

Best practice TTS based on BERT and VITS with some Natural Speech Features Of Microsoft

Hugging Face Spaces GitHub Repo stars GitHub forks GitHub issues GitHub

这是一个用于TTS算法学习的项目，如果您在寻找直接用于生产的TTS，本项目可能不适合您！

vits_bert.mp4

天空呈现的透心的蓝，像极了当年。总在这样的时候，透过窗棂，心，在天空里无尽的游弋！柔柔的，浓浓的，痴痴的风，牵引起心底灵动的思潮；情愫悠悠，思情绵绵，风里默坐，红尘中的浅醉，诗词中的优柔，任那自在飞花轻似梦的情怀，裁一束霓衣，织就清浅淡薄的安寂。

风的影子翻阅过淡蓝色的信笺，柔和的文字浅浅地漫过我安静的眸，一如几朵悠闲的云儿，忽而氤氲成汽，忽而修饰成花，铅华洗尽后的透彻和靓丽，爽爽朗朗，轻轻盈盈

时光仿佛有穿越到了从前，在你诗情画意的眼波中，在你舒适浪漫的暇思里，我如风中的思绪徜徉广阔天际，仿佛一片沾染了快乐的羽毛，在云环影绕颤动里浸润着风的呼吸，风的诗韵，那清新的耳语，那婉约的甜蜜，那恬淡的温馨，将一腔情澜染得愈发的缠绵。

Features，特性

1, Hidden prosody embedding from BERT，get natural pauses in grammar

2, Infer loss from NaturalSpeech，get less sound error

3, Framework of VITS，get high audio quality

4, Module-wise Distillation, get speedup

💗Tip: It is recommended to use Infer Loss fine-tune model after base model trained, and freeze PosteriorEncoder during fine-tuning.

💗意思就是：初步训练时，不用loss_kl_r；训练好后，添加loss_kl_r继续训练，稍微训练一下就行了，如果音频质量差，可以给loss_kl_r乘以一个小于1的系数、降低loss_kl_r对模型的贡献；继续训练时，可以尝试冻结音频编码器Posterior Encoder；总之，玩法很多，需要多尝试！

naturalspeech

为什么不升级为VITS2

VITS2最重要的改进是将Flow的WaveNet模块使用Transformer替换，而在TTS流式实现中，通常需要用纯CNN替换Transformer。

Online demo，在线体验

https://huggingface.co/spaces/maxmax20160403/vits_chinese

Install，安装依赖和MAS对齐

pip install -r requirements.txt

cd monotonic_align

python setup.py build_ext --inplace

Infer with Pretrained model，用示例模型推理

Get from release page vits_chinese/releases/

put prosody_model.pt To ./bert/prosody_model.pt

put vits_bert_model.pth To ./vits_bert_model.pth

python vits_infer.py --config ./configs/bert_vits.json --model vits_bert_model.pth

./vits_infer_out have the waves inferred, listen !!!

Infer with chunk wave streaming out，分块流式推理

as key parameter, hop_frame = ∑decoder.ups.padding 💗

python vits_infer_stream.py --config ./configs/bert_vits.json --model vits_bert_model.pth

Ceil duration affect naturalness

So change w_ceil = torch.ceil(w) to w_ceil = torch.ceil(w + 0.35)

All Thanks To Our Contributors:

Train，训练

download baker data https://aistudio.baidu.com/datasetdetail/36741, more info: https://www.data-baker.com/data/index/TNtts/

change sample rate of waves to 16kHz, and put waves to ./data/waves

python vits_resample.py -w [input path]:[./data/Wave/] -o ./data/waves -s 16000

put 000001-010000.txt to ./data/000001-010000.txt

python vits_prepare.py -c ./configs/bert_vits.json

python train.py -c configs/bert_vits.json -m bert_vits

bert_lose

额外说明

原始标注为

000001	卡尔普#2陪外孙#1玩滑梯#4。
  ka2 er2 pu3 pei2 wai4 sun1 wan2 hua2 ti1
000002	假语村言#2别再#1拥抱我#4。
  jia2 yu3 cun1 yan2 bie2 zai4 yong1 bao4 wo3

标注规整后：

  • BERT需要汉字 卡尔普陪外孙玩滑梯。 (包括标点)
  • TTS需要声韵母 sil k a2 ^ er2 p u3 p ei2 ^ uai4 s uen1 ^ uan2 h ua2 t i1 sp sil
000001	卡尔普陪外孙玩滑梯。
  ka2 er2 pu3 pei2 wai4 sun1 wan2 hua2 ti1
  sil k a2 ^ er2 p u3 p ei2 ^ uai4 s uen1 ^ uan2 h ua2 t i1 sp sil
000002	假语村言别再拥抱我。
  jia2 yu3 cun1 yan2 bie2 zai4 yong1 bao4 wo3
  sil j ia2 ^ v3 c uen1 ^ ian2 b ie2 z ai4 ^ iong1 b ao4 ^ uo3 sp sil

训练标注为

./data/wavs/000001.wav|./data/temps/000001.spec.pt|./data/berts/000001.npy|sil k a2 ^ er2 p u3 p ei2 ^ uai4 s uen1 ^ uan2 h ua2 t i1 sp sil
./data/wavs/000002.wav|./data/temps/000002.spec.pt|./data/berts/000002.npy|sil j ia2 ^ v3 c uen1 ^ ian2 b ie2 z ai4 ^ iong1 b ao4 ^ uo3 sp sil

遇到这句话会出错

002365	这图#2难不成#2是#1Ｐ过的#4？
  zhe4 tu2 nan2 bu4 cheng2 shi4 P IY1 guo4 de5

拼音错误修改

将正确的词语和拼音写入文件： ./text/pinyin-local.txt

渐渐 jian4 jian4
浅浅 qian3 qian3

数字播报支持

已支持，基于WeNet开源社区WeTextProcessing；当然，不可能是完美的

不使用Bert也能推理

python vits_infer_no_bert.py --config ./configs/bert_vits.json --model vits_bert_model.pth

虽然训练使用了Bert，但推理可以完全不用Bert，牺牲自然停顿来适配低计算资源设备，比如手机

低资源设备通常会分句合成，这样牺牲的自然停顿也没那么明显

ONNX非流式

导出：会有许多警告，直接忽略

python model_onnx.py --config configs/bert_vits.json --model vits_bert_model.pth

推理

python vits_infer_onnx.py --model vits-chinese.onnx

ONNX流式

具体实现，将VITS拆解为两个模型，取名为Encoder和Decoder。

  • Encoder包括TextEncoder与DurationPredictor等；

  • Decoder包括ResidualCouplingBlock与Generator等；

  • ResidualCouplingBlock，即Flow，可以放在Encoder或Decoder，放在Decoder需要更大的hop_frame

并且将推理逻辑也进行了切分；特别的，先验分布的采样过程放在了Encoder中：

z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale

ONNX流式模型导出

python model_onnx_stream.py --config configs/bert_vits.json --model vits_bert_model.pth

ONNX流式模型推理

python vits_infer_onnx_stream.py --encoder vits-chinese-encoder.onnx --decoder vits-chinese-decoder.onnx

在流式推理中，hop_frame是一个重要参数，需要去尝试出合适的值

Model compression based on knowledge distillation，应该叫迁移学习还是知识蒸馏呢？

Student model has 53M size and 3× speed of teacher model.

To train:

python train.py -c configs/bert_vits_student.json -m bert_vits_student

To infer, get student model at the release page

python vits_infer.py --config ./configs/bert_vits_student.json --model vits_bert_student.pth

代码来源

Microsoft's NaturalSpeech: End-to-End Text to Speech Synthesis with Human-Level Quality

Nix-TTS: Lightweight and End-to-End Text-to-Speech via Module-wise Distillation

https://github.com/Executedone/Chinese-FastSpeech2 bert prosody

https://github.com/wenet-e2e/WeTextProcessing

https://github.com/TensorSpeech/TensorFlowTTS Heavily depend on

https://github.com/jaywalnut310/vits

https://github.com/wenet-e2e/wetts

https://github.com/csukuangfj onnx and android

BERT应用于TTS

2019 BERT+Tacotron2: Pre-trained text embeddings for enhanced text-tospeech synthesis.

2020 BERT+Tacotron2-MultiSpeaker: Improving prosody with linguistic and bert derived features in multi-speaker based mandarin chinese neural tts.

2021 BERT+Tacotron2: Extracting and predicting word-level style variations for speech synthesis.

2022 https://github.com/Executedone/Chinese-FastSpeech2

2023 BERT+VISINGER: Towards Improving the Expressiveness of Singing Voice Synthesis with BERT Derived Semantic Information.

AISHELL3多发音人训练，训练出的模型可用于克隆