Proceeding of the past Conferences /

Below you can read the papers of the past Conferences


New Music Concepts 2019

MGTGAN:Cycle-Consistent Adversarial Networks for Symbolic Multi-track Music Genre Transfer

Part of the New Music Concepts book series (NMC volume 6)
ISBN: 978-88-944350-0-9

Author(s): YenLun Peng, Haitao Zheng

Abstract: Transferring music from one genre to another has a few differences from transferring image from one domain to other domain. Color digital images are made of pixels, and pixels are made of combinations of primary colors represented by a series of code. However, Music is consist of multiple sound tracks, in symbolic music representation like MIDI, we tend to use different instruments with their temporal dynamics to represent to each track. In order to discover the difference between these two types of representation for channels, we proposed MGTGAN to analyze how different channel affect to the result of music genre transfer. Following the recent research from MuseGAN and CycleGAN, we inspired by MuseGAN's three type of model, and chose two of them, which are jamming model, composer model. By combining these two models with CycleGAN. We successfully transfer multi-track music from one genre to another. We also investigate applying different loss function like WGAN in CycleGAN and adding two distinct discriminators to get better convergence. Due to the complexity of the multi-track music is hard to train and evaluate, we also propose a Desert Camel MIDI dataset to simplify the experiment. Each song is well defined in multiple genre, which means it is a paired data and have a correct answer. The evaluation can more focus on judging the certain chord for invariant content and the rhythm style for genre transfer. Our result show that jamming model can transfer each track better than composer model. However, subjective evaluation from human gave composer model higher harmonic score. The code and dataset are available at https://github.com/AllenPeng0209/MGTGAN

Keywords: Music Genre Transfer, Deep Learning, Generative Adversarial Networks