BS-RoFormer
ByteDance's band-split Transformer that won the Sound Demixing Challenge 2023 — the architecture behind today's best stem splitters.
About
BS-RoFormer is the model that moved music separation from RNNs to Transformers. It splits the complex spectrogram into subbands, projects each band with MLPs instead of CNN front-ends, then runs hierarchical Transformers across bands and across time to estimate separation masks — with rotary position embeddings (RoPE) as the key training trick. The ByteDance team's system won the music separation track of the Sound Demixing Challenge 2023.\n\nThe paper reports a compact variant reaching 9.80 dB average SDR on MUSDB18-HQ without any extra training data, and the full competition system hit 9.97 average global SDR. The design is now the template everyone copies: Mel-Band RoFormer, the Kim_Vocal_2 UVR model, and several 2024 papers are direct descendants.\n\nThe honest caveat: ByteDance never released official code or weights. What the community uses is lucidrains' open reimplementation (MIT) plus weights trained by community members like ZFTurbo. Treat it as a verified research design with community-built artifacts, not a vendor-supported model.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Author-reported (arXiv 2309.02612): SDX'23 music separation track, 1st place (SAMI-ByteDance), 9.97 avg global SDR with extra data\n- Author-reported: compact variant 9.80 dB median SDR on MUSDB18-HQ without extra training data\n- No independent replication published yet — all numbers are author-reported; official weights never released
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER verified, metrics-v1). Evidence: SDX'23-winning system with a public arXiv paper (2309.02612) from ByteDance's SAMI team, and a widely used MIT community reimplementation with openly trained weights. Strengths: landmark architecture, the template for the current generation of separators, strong author-reported results. Docked for: (1) no official code or weights from the authors — artifacts are community-built; (2) all benchmark numbers are author-reported. License: MIT (community implementation); author weights were never released.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · arXiv paper 2309.02612 (Sep 2023) from ByteDance SAMI (Lu, Wang, Kong, Hung): band-split RoPE Transformer; SDX'23 music separation track 1st place.
- · Author-reported: compact variant 9.80 dB median SDR on MUSDB18-HQ without extra data; competition system 9.97 avg global SDR with extra data. No independent replication.
- · Template architecture for the current generation (Mel-Band RoFormer, Kim_Vocal_2); usable via the MIT lucidrains reimplementation plus community-trained weights (e.g. ZFTurbo).