A Universal Music Translation Network

  • 2018-05-21 00:20:03
  • Noam Mor, Lior Wolf, Adam Polyak, Yaniv Taigman
  • 179

Abstract

We present a method for translating music across musical instruments, genres,and styles. This method is based on a multi-domain wavenet autoencoder, with ashared encoder and a disentangled latent space that is trained end-to-end onwaveforms. Employing a diverse training dataset and large net capacity, thedomain-independent encoder allows us to translate even from musical domainsthat were not seen during training. The method is unsupervised and does notrely on supervision in the form of matched samples between domains or musicaltranscriptions. We evaluate our method on NSynth, as well as on a datasetcollected from professional musicians, and achieve convincing translations,even when translating from whistling, potentially enabling the creation ofinstrumental music by untrained humans.

 

Quick Read (beta)

loading the full paper ...