JazzGAN: Improvising with Generative Adversarial Networks

Abstract

For the purpose of creating a jazz teaching tool in the open-source Impro-Visor (Improvisation Advisor) application, we trained JazzGAN, a generative adversarial network (GAN) using recurrent neural networks (RNN) to improvise monophonic jazz melodies over chord progressions. Improvising jazz melodies creates several challenges not addressed by previous generative adversarial neural networks for music generation, including (1) frequent and diverse key changes; (2) unconventional and off-beat rhythms; (3) flexibility with off-chord notes. To address these issues, we compare the performance of several data representations with JazzGAN and propose the use of harmonic bricks for phrase segmentation. We define metrics to quantify the aforementioned issues, compare several data encodings of rhythm, and show that JazzGAN compares favorably against Magenta's ImprovRNN.