Kim_Vocal_2
The community-trained Mel-Band RoFormer vocal checkpoint that became the default 'best vocals' button across the UVR ecosystem.
About
Kim_Vocal_2 is the rare community checkpoint that became infrastructure. Trained by Kimberley Jensen on the Mel-Band RoFormer architecture (the ByteDance ISMIR 2023 design), it is a single-stem vocals model — roughly 228M parameters at 44.1 kHz stereo — and it earned its reputation the hard way: producers kept picking it in blind comparisons inside Ultimate Vocal Remover, python-audio-separator, and MVSEP, until it became the recommended default vocal model in several tools.\n\nThe instrumental is derived as mixture-minus-vocals, the standard practice for single-stem checkpoints. The weights are MIT-licensed on Hugging Face (KimberleyJSN/melbandroformer), with MLX and GGUF ports for Apple Silicon and llama.cpp-style runtimes.\n\nThe honest framing: there is no paper and no published SDR benchmark for this checkpoint — its quality case is entirely community consensus and leaderboard-adjacent adoption. For pure vocal isolation it is arguably the most battle-tested free model available; for full 4-stem splits, pair it with a dedicated multi-stem model.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- No published benchmarks — the author has not released SDR numbers; quality is established by community consensus (recommended default in UVR-family tools)\n- Architecture basis (Mel-Band RoFormer) is author-reported to beat BS-RoFormer on vocals on MUSDB18-HQ, but that number is for the paper's model, not this checkpoint
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 metrics, metrics-v1). Evidence: MIT-licensed weights on Hugging Face from a known community trainer, deep adoption as the default vocal model across the UVR/python-audio-separator ecosystem, and MLX/GGUF ports with parity testing. Strengths: battle-tested on real music, open license, wide tooling support. Docked for: (1) no paper and no published benchmarks — quality rests on community consensus; (2) vocals-only, single stem. License: MIT (relicensed from GPL-3.0 per the HF card).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · Community-trained Mel-Band RoFormer checkpoint (~228M params, 44.1 kHz stereo, vocals-only) by Kimberley Jensen (@KimberleyJSN); instrumental derived as mixture-minus-vocals.
- · No paper and no published SDR benchmarks — quality rests on community consensus: the recommended default vocal model across the UVR / python-audio-separator / MVSEP ecosystem, with MLX and GGUF ports (parity-tested).
- · Single stem only; pair with a multi-stem model for full splits. Exact release date unpublished; community references date it to 2024.