eloimoliner / unconditional-diff-STFT

Unconditional music synthesis using a diffusion model in the STFT domain

Unconditional synthesis of music using a diffusion model in the STFT domain.

This repository is the result of a set of experiments training a diffusion model to unconditionally synthesize piano music. The model, based on the complex-spectrogram domain shows superior performance to previous works using time-domain models as the backbone of a diffusion model.

Implementation details

References

This implementation is based on / inspired by:

• Diffusion model for gramophone noise synthesis PyTorch repo.
• DiffWave unoffifial PyTorch repo.
• CRASH PyTorch repo.