Mel-Band RoFormer
The mel-scale successor to BS-RoFormer — overlapping psychoacoustic bands instead of heuristic splits, and the basis of the best UVR vocal models.
About
Mel-Band RoFormer is the follow-up to BS-RoFormer from the same ByteDance team, and its one big idea is principled band design. Where BS-RoFormer split the spectrum into non-overlapping bands by heuristic rules, Mel-Band RoFormer maps frequency bins into overlapping subbands on the mel scale — the psychoacoustic scale that mirrors how humans hear — so low frequencies get finer resolution where it matters. On MUSDB18-HQ the authors report it beating BS-RoFormer on vocals, drums, and other stems.\n\nThis is the architecture that escaped the lab: community trainer Kimberley Jensen built the Kim_Vocal_2 checkpoint on it, and it became the recommended default vocal model in Ultimate Vocal Remover, python-audio-separator, and half the open-source vocal-splitting tools. The ISMIR 2024 follow-up extended the same backbone to vocal melody transcription.\n\nLike its predecessor, there is no official ByteDance release — the ecosystem runs on lucidrains' MIT reimplementation and community-trained weights. The paper itself was peer-reviewed at ISMIR 2023.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Author-reported (arXiv 2310.01809, ISMIR 2023): outperforms BS-RoFormer on vocals, drums, and other stems on MUSDB18-HQ\n- Community adoption: basis of the Kim_Vocal_2 UVR vocal model and ZFTurbo variants; widely used as the default vocal separator in open-source tools\n- No independent replication of the paper's numbers published yet — all headline figures are author-reported
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: peer-reviewed paper at ISMIR 2023 (arXiv 2310.01809) from ByteDance SAMI, plus massive community adoption — the Kim_Vocal_2 checkpoint built on this architecture is the default vocal model across the UVR ecosystem. Strengths: principled mel-band design, real-world validation through community use, MIT reimplementation available. Docked for: (1) no official weights from the authors; (2) paper numbers are author-reported with no independent replication. License: MIT (community implementation).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Peer-reviewed paper at ISMIR 2023 (arXiv 2310.01809) from ByteDance SAMI (Wang, Lu, Won): overlapping mel-scale subbands instead of heuristic splits; author-reported to beat BS-RoFormer on vocals, drums, and other on MUSDB18-HQ.
- · The architecture that escaped the lab: Kimberley Jensen's Kim_Vocal_2 checkpoint and ZFTurbo variants are built on it; it became the default vocal model across the UVR ecosystem. ISMIR 2024 follow-up extended it to vocal melody transcription.
- · No official weights from the authors; the ecosystem runs on the MIT lucidrains reimplementation plus community checkpoints.