The Musdb18 Corpus For Music Separation
Authors
- Zafar Rafii
- Antoine Liutkus
- Fabian-Robert Stöter
- Stylianos Ioannis Mimilakis
- Rachel Bittner
Abstract
The sigsep musdb18 data set consists of a total of 150 full-track songs of different styles and includes both the stereo mixtures and the original sources, divided between a training subset and a test subset. Its purpose is to serve as a reference database for the design and the evaluation of source separation algorithms. The objective of such signal processing methods is to estimate one or more sources from a set of mixtures, e.g. for karaoke applications. It has been used as the official dataset in the professionally-produced music recordings task for SiSEC 2018, which is the international campaign for the evaluation of source separation algorithms. musdb18 contains two folders, a folder with a training set: “train”, composed of 100 songs, and a folder with a test set: “test”, composed of 50 songs. Supervised approaches should be trained on the training set and tested on both sets. All files from the musdb18 dataset are encoded in the Native Instruments stems format (.mp4). It is a multitrack format composed of 5 stereo streams, each one encoded in AAC @256kbps. These signals correspond to: 0 - The mixture, 1 - The drums, 2 - The bass, 3 - The rest of the accompaniment, 4 - The vocals. For each file, the mixture correspond to the sum of all the signals. All signals are stereophonic and encoded at 44.1kHz. As the MUSDB18 is encoded as STEMS, it relies on ffmpeg to read the multi-stream files. We provide a python wrapper called stempeg that allows to easily parse the dataset and decode the stem tracks on-the-fly.