MUSEGAN : DEMONSTRATION OF A CONVOLUTIONAL GAN BASED MODEL FOR GENERATING MULTI-TRACK PIANO-ROLLS

Generating realistic and aesthetic pieces is one of the most exciting tasks in the field. We present in this demo paper a new neural music generation model we recently proposed, called MuseGAN. We exploit the potential of applying generative adversarial networks (GANs) to generate multi-track pop/rock music of four bars, using convolutions in both the generators and the discriminators. Moreover, we propose an efficient approach for pre-processing symbolic data and share the data with the community. Our model can generate music either from scratch, or by following (accompanying) a track given by user.