What’s Out There
← Back to swiping
Audio & musicFreeAcademic

MUSDB18: a corpus for music separation

Rafii, Liutkus, Stöter, et al. (SigSep)

150 full songs delivered as a mix plus separated vocals, drums, bass, and other stems.

A reference data set for music source separation: 150 full-length tracks of various styles, each with a stereo mix and its original sources, split into training and test subsets. It has been the official data set for professionally produced music tasks in separation evaluation campaigns. A lossless MUSDB18-HQ version exists.

Access

Cost
FreeFree download from Zenodo; no account needed.
License
Open, non-commercial onlyCustom non-commercial terms (research use only)
Sign-up
No sign-up
How you get it
Bulk downloadMP4 stems (ZIP)

What's in it

One row is
One song with mixture and separate stems
Covers
Not geographic
Years
Release 2017
Updated
One-time release
Size
150 tracks
stereo mixturevocalsdrumsbassother stemtrain or test split

Quality

4 of 5

Standard benchmark for separation research; small by modern standards and not updated.

  • Non-commercial license and mixed track licenses
  • Only 150 tracks, Western pop and rock heavy

Connects to

  • MUSDB18-HQ: Uncompressed audio version of the same tracks
  • MedleyDB: Multitrack set with instrument-level stems
  • MoisesDB: Larger later multitrack separation set

Questions it could answer

  • Which stems are hardest for models to separate cleanly?
  • How does model quality vary by genre?

Similar data sets

  • RAVDESS: Ryerson Audio-Visual Database of Emotional Speech and Song

    7,356 recordings of 24 actors performing emotions, in audio, video, and song.

    Audio & musicLivingstone & Russo (Ryerson University)Free

  • Million Song Dataset

    Audio features for a million songs, with no audio, frozen in December 2010.

    Audio & musicLabROSA (Columbia University) and The Echo NestFree

  • MusicNet

    330 classical recordings with over a million labeled notes aligned to the second.

    Audio & musicThickstun, Harchaoui, Kakade (University of Washington)Free

  • GuitarSet

    360 guitar excerpts with hexaphonic pickups, giving note-level labels for each string.

    Audio & musicXi, Bittner, Pauwels, Ye, Bello (NYU MARL)Free

  • MusicBrainz Database

    An open, crowd-edited encyclopedia of recorded music, with the core data in the public domain.

    Audio & musicMetaBrainz FoundationFree

  • FSD50K

    51,197 Freesound clips labeled with 200 sound-event classes drawn from the AudioSet Ontology.

    Audio & musicFonseca, Favory, Pons, et al. (Universitat Pompeu Fabra)Free