Version 3.0 dev log

[yqzhishen]

Branches

• main branch will be frozen and will not update until v3 release. For latest updates (if there are any), please turn to v2-backport.
• All refactoring and new features, including bug fixes to the last v2 release, goes to v3. This branch will be merged into main once v3 is ready for release.
• v2-backport will contain all bug fixes to the last v2 release and backported features from v3, and will be discarded once v3 is ready for public testing.
• v3-dev is for daily development. Commits shall be squashed and merged into v3.

TODO list

Framework

• New configuration system based on OmegaConf and Pydantic
• Optimized binarizer workflows
• Refactor NN modules
• New training framework
• Acoustic model training
• Variance model training
• Simple inference and composed inference
• ONNX exporting with latest PyTorch versions

Features

• Muon optimizer + LYNXNet 2
• EMA (framework support)
• LoRA (framework support)
• Remake tension
• Falsetto parameter
• Mouth opening parameter
• Inpainting (retaking) in acoustic models
• Generative duration predictor
• Latent NSF vocoder