A Deep Neural Network based approach for vocal extraction from songs

Songs and media files on the internet have grown at an unprecedented rate. Music has been one of the primary sources of entertainment for mankind for a long time now. With such large amounts of media available to us, it has become possible to use this to our advantage to solve problems which have been considered difficult to solve traditionally. One such problem is the separation of vocals and instrumental part from a song. This problem has largely remain unsolved despite a lot of work having been done on it, largely due to the difficulty in separating these two components of a song due to the high correlation and coherence between the two. In this paper we present a Deep Neural Network based approach to approach the problem and demonstrate how it shows a lot of promise for several type of songs and outperforms the existing techniques for most songs. Several Neural network architectures are experimented with and a detailed comparison between the results obtained from the various architectures are discussed in this paper.