A deep learning approach to source separation and remixing of HipHop music

Martel Baro, Héctor

A deep learning approach to source separation and remixing of HipHop music

License

Citation

Martel Baro, H. A deep learning approach to source separation and remixing of HipHop music. 2017. handle: http://hdl.handle.net/10230/32919

This citation was generated automatically.

Abstract

Audio source separation has been one of the major research fields in audio processing during the past years. The main purpose of this discipline is to decompose a mixture signal into simpler components called sources, which applied to professionally produced music means to recover the instrument tracks. This is achieved by undoing the mixing process, which might vary depending on the music genre. In this work, an existing framework based on Deep Neural Networks will be adapted to the particularities to HipHop music and tested with a proposed dataset. The aim is to determine whether or not it is a suitable approach to be implemented in a remixing application. For this reason, objective and subjective quality tests are performed to evaluate the resulting separations. The isolated instrument tracks can be used for many different purposes, specially focusing on the following two aspects. First, each instrument can be placed in a 3D space so that the song can be reproduced not only in stereo, but in more complex configurations for which the song was not initially produced. Second, all the instruments can be controlled separately to edit the song, thus, enabling the possibility to remix it. Consequently, further work on this subject will make it possible for musicians and producers to manipulate mixed songs in new powerful ways to create their content.