Multitrack Symbolic Music Composition with LSTM-GAN

Music generation is becoming a heated subject in the field of machine learning. Particularly, in the direction of multitrack, sequential, and symbolic music composition, methods based on generative adversarial networks (GAN) such as MuseGAN and BinaryMuseGAN have shown outstanding performance and have been adopted by many researchers. However, there are still issues in musicality and temporal correlation that need to be addressed. In this paper, an LSTM-based GAN is proposed in order to capture more temporal information from real-world music samples. One bar is seen as the smallest unit and fed into the LSTM network in the discriminator. Additionally, a processor is added between the generator and the discriminator to extract important information from the data. This LSTM-GAN scheme helps the discriminator find important differentiators more quickly and hence improves the ability of the generator to fabricate authentic data. The proposed method is validated via experiments, where the generated data shows similarities with the authentic ones.