Task¶
The task of the Music Demixing (MDX) Challenge is music demixing, which is a special case of source separation.
Source Separation¶
Source separation is the process of isolating individual sounds in an auditory mixture of multiple sounds [CFL+18, RLStoter+18, VVG18] , as defined in the previous tutorial [MSS20].
Mathematically, we assume that a mixture signal \(y(t)\) consists of \(N\) sources, \(x_n(t)\), for \(n=1...N\), such that
Source separation is the inverse operation of the mixing process. We aim to find the original signals, \(x_n(t)\) for \(n=1...N\), for a given mixture signal \(y(t)\).
Music Demixing¶
Music Demixing (MDX) is a special case of Source Separation. A given mixture signal is a song, and source signals are sounds from musical instruments.
Recently, many data-driven approaches have been proposed for music demixing. Especially, deep learning methods [CKCJ21, CKC+20, DefossezUBB19a, DefossezUBB19b, HKVM20, KCL+21, LKJX21, LXWY20, LluisPS19, TM20] have become mainstream because they have shown outstanding performance.
To train a deep neural network for music demixing, we need a training dataset. We explore datasets for music demixing in the following section.