Spiking Generative Adversarial Network for Controllable Affective Music Creation

Yuan Zhao, Xianghong Lin, Zequn Zhang, Chengyang Xie, Ruidong Ma

Abstract: Several Generative Adversarial Network (GANs) models have been successful in generating music using piano rolls. However, these models typically produce continuous-valued piano rolls, which require preprocessing or specialized neural network layers for discretization before final output. Additionally, current approaches often overlook the controllability aspect of music generation. In this paper, we introduce SpikingMuseGAN, which leverages spiking neurons to generate high-quality piano rolls. Furthermore, we design a controllable emotional music generation architecture, endowing SpikingMuseGAN with the capability to generate music with specific emotional attributes. Experimental results demonstrate that utilizing spiking neurons effectively enhances the quality of generated music. In summary, SpikingMuseGAN empowers users to generate high-quality emotional music in a controllable manner, offering extensive potential applications across various domains.