Skip to content

v2.3 Final Release

Compare
Choose a tag to compare
@Stardust-minus Stardust-minus released this 19 Dec 11:17
· 102 commits to master since this release
76653b5

本次更新内容如下:
1.完善了 webui_preprocess.py(By @OedoSoldier
2.修正了存在于 Duration Predictor Discriminator 的 bug。并添加了 LSTM。对 SDP 有较大提升。
3.支持 BF16 训练。
4.对英语的 G2P 进行修复。
5.添加基于 WavLM 的 Discriminator(来源于 StyleTTS2)
6.因情感控制效果不佳,去除了 CLAP。更换了更简单的 BERT 融合语义方式。(By @litagin02

这可能是 Bert-VITS2 近期的 final release,接下来我们将会加入对 Fish Speech 的开发。

Bert-VITS2项目组
敬上

This might be our final release for the time being as we shift our focus toward our new project Fish Speech.

We continue to appreciate your constructive advice for further improvements. Please feel free to share your thoughts in pull requests.


Base model: https://huggingface.co/OedoSoldier/Bert-VITS2-2.3


Update: The configuration for the gin_channels parameter in the pre-trained 2.3 weights has been adjusted to 512. To align with this change, we've updated the corresponding entry in the config.json file.