MUSDB18 Music Separation Benchmark

A separation benchmark needs mixtures and the isolated sources that were combined to make them. MUSDB18 provides full-length music tracks with four named stems: vocals, drums, bass and other instruments or sounds. The SigSep dataset documentation lists 150 tracks, with 100 in its training split and 50 in its test split. This shared setup lets researchers train and compare systems without inventing a private reference collection. It also makes the limits of a score easier to state: the result applies to a specific corpus, split and scoring method. The “other” stem is broad. It can contain many instruments and production elements, so a separator’s error there is not one simple instrument mistake. Stems may overlap in frequency, and effects such as reverb can blur boundaries. A system can have a strong vocal score and still distort a bass note or lose a cymbal transient. Report each source rather than one average, and listen to representative failures. Objective metrics such as SDR or SI-SDR depend on their definitions and cannot replace auditory judgment. Dataset variants matter. MUSDB18-HQ is a related high-quality WAV version; the standard release has its own encoding and usage conventions. A paper should identify exactly which version and whether extra training data or postprocessing was used. Test-set songs should not be used repeatedly to choose a model. Large pretraining collections can also contain overlapping music, so teams should investigate contamination where feasible. The catalog is valuable but not the full world of recorded sound. Live concerts, unusual regional genres, heavily compressed social clips and film dialogue differ from many studio songs. Rights to training data and separated outputs also require attention; having access to a benchmark does not grant permission to publish every derivative. Use MUSDB18 to compare research and add representative local tests before claiming a separator will satisfy real musicians or listeners.