Demucs: Deep Extractor for Music Sources with extra unlabeled data remixed

3 Sep 2019Alexandre DéfossezNicolas UsunierLéon BottouFrancis Bach

We study the problem of source separation for music using deep learning with four known sources: drums, bass, vocals and other accompaniments. State-of-the-art approaches predict soft masks over mixture spectrograms while methods working on the waveform are lagging behind as measured on the standard MusDB benchmark... (read more)

PDF Abstract

Results from the Paper


 SOTA for Music Source Separation on MUSDB18 (using extra training data)

     Get a GitHub badge
TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK USES EXTRA
TRAINING DATA
RESULT LEADERBOARD
Music Source Separation MUSDB18 DEMUCS SDR (vocals) 6.29 # 7
SDR (drums) 6.08 # 4
SDR (other) 4.12 # 5
SDR (bass) 5.83 # 3
Music Source Separation MUSDB18 DEMUCS (extra) SDR (vocals) 7.05 # 1
SDR (drums) 7.08 # 2
SDR (other) 4.47 # 1
SDR (bass) 6.70 # 2