fishaudio /Bert-VITS2

vits2 backbone with multilingual-bert

★ 8,803⑂ 1,307👁 49⑃ 352● 0▤ 12
PythonAGPL-3.0Created 2023-07-29
下载:Extra:Japanese Specialized VersionJP-Exta ↓JP-Exta2024-02-01
+0
日增长
8,803
stars
1,307
forks
49
watchers
352
prs
0
issues
12
releases

README

Bert-VITS2

VITS2 Backbone with multilingual bert

For quick guide, please refer to webui_preprocess.py.

简易教程请参见 webui_preprocess.py。

 【项目推介】
 FishAudio下的全新自回归TTS Fish-Speech现已可用,效果为目前开源SOTA水准,且在持续维护,推荐使用该项目作为BV2/GSV的替代。本项目短期内不再进行维护。
 Demo Video: https://www.bilibili.com/video/BV18E421371Q
 Tech slides Video: https://www.bilibili.com/video/BV1zJ4m1K7cj
 请注意,本项目核心思路来源于anyvoiceai/MassTTS 一个非常好的tts项目
 MassTTS的演示demo为ai版峰哥锐评峰哥本人,并找回了在金三角失落的腰子

[//]:  ( 本项目与[PlayVoice/vits_chinese]&40;https://github.com/PlayVoice/vits_chinese&41; 没有任何关系)

[//]:  ()
[//]:  (本仓库来源于之前朋友分享了ai峰哥的视频,本人被其中的效果惊艳,在自己尝试MassTTS以后发现fs在音质方面与vits有一定差距,并且training的pipeline比vits更复杂,因此按照其思路将bert)

 成熟的旅行者/开拓者/舰长/博士/sensei/猎魔人/喵喵露/V应当参阅代码自己学习如何训练。

 严禁将此项目用于一切违反《中华人民共和国宪法》,《中华人民共和国刑法》,《中华人民共和国治安管理处罚法》和《中华人民共和国民法典》之用途。
 严禁用于任何政治相关用途。
 Video:https://www.bilibili.com/video/BV1hp4y1K78E
 Demo:https://www.bilibili.com/video/BV1TF411k78w
 References
+ anyvoiceai/MassTTS
+ jaywalnut310/vits
+ p0p4k/vits2_pytorch
+ svc-develop-team/so-vits-svc
+ PaddlePaddle/PaddleSpeech
+ emotional-vits
+ fish-speech
+ Bert-VITS2-UI
 感谢所有贡献者作出的努力

  


[//]:  ( 本项目所有代码引用均已写明,bert部分代码思路来源于[AI峰哥]&40;https://www.bilibili.com/video/BV1w24y1c7z9&41;,与[vits_chinese]&40;https://github.com/PlayVoice/vits_chinese&41;无任何关系。欢迎各位查阅代码。同时,我们也对该开发者的[碰瓷,乃至开盒开发者的行为]&40;https://www.bilibili.com/read/cv27101514/&41;表示强烈谴责。)

Discussions

Latest release

Extra:Japanese Specialized Version JP-Exta 2024-02-01
After seeing the success of the Chinese special version, we also trained the Japanese special version.
Update log:
1.Delete Chinese and English support.
2.Merge Pull Request(312 ), which fixed an error existed in previous Japanese g2p. (Thanks @litagin02 )
3.We increased Japanese training data several times. Now up to ~800h single language.
4.Thrown away DurationDiscriminator, which often lead to training unstable.
5.Changed the emotion VectorQuantize, try to have a larger codebook_size but smaller codebook_dim to help increase the usage of codebook. By the hand, We also fix the loss_com

Snapshot: 2026-11-16T09:23:14Z · Data license Apache-2.0 · JSON