Datasets

The MASS dataset formed the core content of the early Signal Separation Evaluation Campaigns (SiSEC) (Vincent, Araki, and Bofill 2009), which evaluate the quality of various music separation methods. SiSEC always had a strong focus on vocals and accompaniment separation. For a long time, vocals separation methods were very demanding computationally and it was already considered extremely challenging to separate excerpts of only a few seconds.

In the following years, new datasets were proposed that improved over the MASS dataset in many directions. They are summarized in the table below:

DatasetYearReference(s)TracksTrack duration (s)Full/stereo?
MASS2008(Vinyes 2008)9(16 ±\pm 7)❌ / ✔️
MIR-1K2010(Hsu and Jang 2010)1,000(8 ±\pm 8)❌ / ❌
QUASI2011(Liutkus, Badeau, and Richard 2011; E. Vincent, Araki, et al. 2012)5(206 ±\pm 21)✔️ / ✔️
ccMixter2014(Liutkus, FitzGerald, et al. 2014)50(231 ±\pm 77)✔️ / ✔️
MedleyDB2014(Bittner et al. 2014)63(206 ±\pm 121)✔️ / ✔️
iKala2015(Chan et al. 2015)20630❌ / ❌
DSD100sigsep2015(Ono et al. 2015)100(251 ±\pm 60)✔️ / ✔️
MUSDB182017(Rafii et al. 2017)150(236 ±\pm 95)✔️ / ✔️
Last Updated: 9/25/2018, 9:00:22 AM