Vocal Separation by Constrained Non-Negative Matrix Factorization

The vocal separation is to separate vocal part and remove the accompaniment part from the mixed music data. Vocal part include many information, singer, lyric and emotion of the song. If we can extraction only the vocal part from the original sound from CD source, it can be applied to various applications. In this paper, we propose a new method to take out the natural vocal parts from mixed music by using non-negative matrix factorization (NMF). This NMF-based framework separates the harmonic, percussive, and vocal structures from the input signal. We impose the constraint into each component to enforce its feature such as harmonic or temporal continuity. In addition, we propose a framework utilizing the prior information in order to achieve the valid vocal separation over this mathematical procedure. The experiments over some vocal databases show the proposed framework has the superior separation performance compared to the conventional methods. Also by considering the characteristics of music, we intend to obtain high accuracy result.