MUSDB18 - a corpus for music separation

Contents & fields

The dataset contains 150 full-track songs with stereo mixtures and original sources, encoded in Native Instruments stems format (.mp4). Each file contains 5 stereo streams: mixture, drums, bass, other accompaniment, vocals. All signals are stereo, 44.1kHz, AAC @256kbps. Training set: 100 songs; test set: 50 songs.

• 0 - The mixture——Mixture signal
• 1 - The drums——Drums signal
• 2 - The bass——Bass signal
• 3 - The rest of the accompaniment——Other accompaniment signal
• 4 - The vocals——Vocals signal

Research uses

Suitable for designing and evaluating music source separation algorithms, e.g., karaoke applications, signal processing research, and as the official dataset for the professionally-produced music recordings task of SiSEC 2018.

This card was drafted from the source page; institution, coverage, time span, scale, fields and license are subject to the official page (pending human review).