MoonInTheRiver/DiffSinger

Based on the README, here's a technical summary: Implements shallow diffusion for mel-spectrogram generation with dual pipelines: DiffSpeech handles text-to-speech with explicit F0 prediction, while DiffSinger performs singing voice synthesis from lyrics and MIDI/ground-truth pitch inputs. Supports multiple inference modes including end-to-end MIDI-to-mel synthesis and cascade architectures with explicit pitch estimation, accelerated via PNDM sampling, and integrates NSF-HiFiGAN or HiFiGAN vocoders for waveform reconstruction.

4,746 stars. No commits in the last 6 months.