musegan

Why we included this project

Polyphonic multitrack music is hard to generate because instruments overlap in time, and MuseGAN is the project that showed how GANs could handle it. It can write a full pop-style arrangement from scratch or take a track you supply and build accompaniment around it, using bass, drums, guitar, piano, and strings. The code and training pipeline are built around the Lakh Pianoroll Dataset, so you can retrain on your own multitrack pianoroll data instead of relying on a pretrained model. It is primarily a research-oriented reference implementation, though the release includes scripts for downloading and processing training data and for training or sampling from the generator. That makes it a useful starting point for teams studying music generation with adversarial networks, or anyone who wants to experiment with multitrack composition.