Transfer symbolic music style from latent representation

Abstract

Generative models has been widely applied in many computer vision scenarios. Two series of models, GenerativeAdversarial Network(GAN) and Variational Autoencoder(VAE), are getting more and more popular in represen-tation learning. Training these model on discrete sequence data generation is still challenging. We want to takeadvantage of both kind of models. In this work, we first improved a CycleGAN based model to transfer MIDI musicgenre. Then we want to find to combine the CycleGAN model together with a disentangled latent representationfrom VAE to have better understanding of music style.