This is a refactored and enhanced version of DiffSinger: Singing Voice Synthesis via Shallow Diffusion Mechanism based on the original paper and implementation, which provides:
- Cleaner code structure: useless and redundant files are removed and the others are re-organized.
- Better sound quality: the sampling rate of synthesized audio are adapted to 44.1 kHz instead of the original 24 kHz.
- Higher fidelity: improved acoustic models and diffusion sampling acceleration algorithms are integrated.
- More controllability: introduced variance models and parameters for prediction and control of pitch, energy, breathiness, etc.
- Production compatibility: functionalities are designed to match the requirements of production deployment and the SVS communities.
| Overview | Variance Model | Acoustic Model |
|---|---|---|
 |
 |
 |
- Installation & basic usages: See Getting Started
- Dataset creation pipelines & tools: See MakeDiffSinger
- Best practices & tutorials: See Best Practices
- Editing configurations: See Configuration Schemas
- Deployment & production: OpenUTAU for DiffSinger, DiffScope (under development)
- Communication groups: QQ Group (907879266), Discord server
- Progress since we forked into this repository: See Releases
- Roadmap for future releases: See Project Board
- Thoughts, proposals & ideas: See Discussions
ONNX exporting pipelines, OpenUtau support and other deployment-related components are supported ONLY FOR MAIN BRANCH. Without in-depth knowledge on how and why DiffSinger code works, you can easily make things break just like many other machine-learning projects. All experimental branches are only intended for feature development, only a limited feature set is tested, and are never meant to be deployed as ONNX models. This wasn't clearly stated before and has caused us debugging headaches multiple times in the past, in some cases users had:
- distributed ONNX model that won't work in OpenUtau (possibly worked in their own modified revision) and led to mysterious bug report in OpenUtau repo (#xxxx)
- exported ONNX model with Torch 2.2 (which is unsupported), had a success (which is broken, ONNX export has A LOT of quirks we had to iron out with our code) and distributed the model, led to mysterious bug report in OpenUtau repo (#xxxx)
Please refrain from distributing model files generated by experimental branches, for the sake of developers' maintenance burden. This is NOT SUPPORTED.
TBD
TBD
- Paper: DiffSinger: Singing Voice Synthesis via Shallow Diffusion Mechanism
- Implementation: MoonInTheRiver/DiffSinger
- Denoising Diffusion Probabilistic Models (DDPM): paper, implementation
- DDIM for diffusion sampling acceleration
- PNDM for diffusion sampling acceleration
- DPM-Solver++ for diffusion sampling acceleration
- UniPC for diffusion sampling acceleration
- Rectified Flow (RF): paper, implementation
- HiFi-GAN and NSF for waveform reconstruction
- pc-ddsp for waveform reconstruction
- RMVPE and yxlllc's fork for pitch extraction
Any organization or individual is prohibited from using any functionalities included in this repository to generate someone's speech without his/her consent, including but not limited to government leaders, political figures, and celebrities. If you do not comply with this item, you could be in violation of copyright laws.
This forked DiffSinger repository is licensed under the Apache 2.0 License.