MVSEP-MDX23
The community-built 4-stem splitter that took 3rd place at the Sound Demixing Challenge 2023 — Demucs4 and MDX fused with UVR's best vocal weights.
About
MVSEP-MDX23 is what happens when a top community trainer assembles the best of everything: it fuses Demucs4 and MDX-Net architectures with the high-quality vocal weights from the Ultimate Vocal Remover project (credit to Kimberley Jensen), plus web-UI work and dataset augmentation help from the MVSEP community. The result took 3rd place on Leaderboard C of the Sound Demixing Challenge 2023 and became one of the most-used free 4-stem separators.\n\nOn the author's own MultiSong benchmark it self-reports 12.50 dB bass, 11.69 dB drums, 6.54 dB other, 9.51 dB vocals and 15.82 dB instrumental SDR — comfortably ahead of stock Demucs HT 4 on the same test. It ships with a PyQt GUI, a browser web UI, a Colab notebook, and a ~730 MB standalone Windows build, so non-programmers can actually use it.\n\nThe honest framing: it is a competition entry by an individual researcher, not a product — the repo declares no software license, and all quality numbers are self-reported. But with 860+ stars and real challenge placement, it is one of the most credible community models in this space.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Challenge-verified: 3rd place, Sound Demixing Challenge 2023 Leaderboard C (music demixing track)\n- Self-reported (repo README, MultiSong dataset): SDR bass 12.50, drums 11.69, other 6.54, vocals 9.51, instrumental 15.82 — above Demucs HT 4 on the same set\n- Self-reported numbers only; no independent replication published
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: verified 3rd-place finish at SDX'23, 860+ GitHub stars, an actively maintained repo with GUI/Colab/standalone builds from a known community trainer (ZFTurbo). Strengths: real challenge placement, practical packaging, strong self-reported numbers vs Demucs. Docked for: (1) no declared software license — verify before commercial use; (2) quality numbers are self-reported on the author's own dataset. License: none declared; verify before commercial use.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Community competition entry by Roman Solovyev (ZFTurbo): fuses Demucs4 and MDX-Net architectures with high-quality UVR vocal weights (Kimberley Jensen); 3rd place on SDX'23 Leaderboard C (music demixing track).
- · Self-reported MultiSong SDR: bass 12.50, drums 11.69, other 6.54, vocals 9.51, instrumental 15.82 — above stock Demucs HT 4 on the same set. No independent replication.
- · 860+ stars; ships with PyQt GUI, browser web UI, Colab notebook, and a standalone Windows build.