BS-RoFormer

About

Current state-of-the-art music source separation (9.80 dB SDR on MUSDB18HQ). ByteDance/SAMI model using band-split processing with RoPE Transformers in the frequency domain. Won the SDX23 MSS track. No official pip package, but community reimplementations available (lucidrains/BS-RoFormer is the most popular, 811 stars).

Relevance

Highest quality general separation available. Same fixed-stem limitation applies — separates into vocals/bass/drums/other. For bluegrass, it would produce the cleanest possible "other" stem, but banjo/mandolin/fiddle would still be mixed together. Useful as a preprocessing step before specialized separation.

Mentions

• ../sources/2023-09-05-bs-roformer — paper
• ../entities/demucs — previous SOTA
• ../entities/asteroid — PyTorch toolkit, could incorporate BS-RoFormer

Links

• Paper: https://arxiv.org/abs/2309.02612
• Community impl: https://github.com/lucidrains/BS-RoFormer (811 stars)
• MVSEP wrapper: https://github.com/ZFTurbo/MVSEP-CDX23