RAVE

Why we included this project

RAVE lets artists and developers train a variational autoencoder on their own recordings and then run the result live, which is what makes it a reference point for real-time neural audio synthesis. A trained model can transform timbre, apply learned sound mappings, and render audio at interactive latency, so it has become a common tool in electronic music and sound design. The practical pull is the whole pipeline: training scripts for custom data, exported models, and clear paths into a DAW or a live performance setup, helped along by the VST and Max 8 integrations. The codebase is clean PyTorch with a published paper behind it, so it works as a study reference for VAE-based generative audio as much as a working instrument. Non-commercial use is straightforward, but teams planning commercial products should read the CC BY-NC terms before building on it.