BS-RoFormer

Why we included this project

Music source separation usually means training a heavy separator from scratch or settling for a generic model, and BS-RoFormer offers a more direct path. It's a PyTorch implementation of ByteDance's band-split transformer, which uses axial attention across frequency and time and swaps learned absolute positions for rotary encodings, a change the paper credits with a large improvement in quality. The repo supports stereo training and can output multiple stems in a single pass, and community members have open-sourced trained weights and training code for vocal models, including a Mel-Band RoFormer variant, so you can go from the paper to real inference without retraining everything yourself. It has even been used for actual remixes, including a Katy Perry one. If you build stem-separation or remix tools, this is a practical place to start experimenting with a model that has a track record outside benchmarks.