audio-diffusion-pytorch

Why we included this project

This PyTorch library gives you the building blocks for training diffusion models directly on raw audio waveforms, skipping spectrograms and latent codes. It includes configurable U-Net backbones, several diffusion methods and samplers, and covers unconditional generation, text conditioning, autoencoding, upsampling, and vocoding. The core diffusion and sampling logic is dimension-agnostic, so you can reuse it on other data formats with little modification. Note that there are no pretrained weights, and the example configs in the README are illustrative; the Moûsai paper describes the actual configurations used in experiments. That makes it a solid research foundation and learning tool rather than a plug-and-play audio generator.