Skip to content

MUSDB18

About

MUSDB18 is the standard benchmark dataset for music source separation. It contains 150 full-length music tracks (100 training, 50 test) with isolated stems for four sources: drums, bass, vocals, and other (accompaniment). MUSDB18HQ is a higher-quality uncompressed variant. Nearly all modern source separation models — Demucs, HTDemucs, BS-RoFormer, Open-Unmix, Spleeter — report results on this dataset.

Repository

  • https://sigsep.github.io/datasets/musdb.html — official dataset page
  • https://zenodo.org/record/1117372 — MUSDB18 on Zenodo

Relevance

MUSDB18 is the primary evaluation benchmark for source separation models. For bluegrass source separation research, it provides a standard comparison point, though its genre coverage is limited (mostly pop/rock). Bluegrass-specific evaluation would require a domain-specific dataset, but MUSDB18 remains the reference for comparing model architectures and training strategies.

Mentions