Polyphonic Music Genre Transfer

Abstract

This work presents the rst research (to my knowledge) to perform multi-track, polyphonic music genre transfer. In music genre transfer, the goal is to transfer some source musical piece to a target musical genre. Polyphony refers to the property that any instrument is able to play multiple notes per time step. The primary focus is on the data representation { most notably the representation of musical genre { that is optimal for polyphonic music genre transfer, supported by in-depth data exploration. Music genre transfer was performed using a generative adversarial network adapted from the image style transfer literature. The output of the models was then evaluated using two surveys on musical and transfer quality. Although the musical quality of the output did not match the musical quality of real data, the results showed that the models are able to generate musical and coherent pieces with satisfactory transfer quality.