A Symbolic-domain Music Generation Method Based on Leak-GAN

To efficiently train deep generative neural works, generative adversarial networks have been applied in plenty of fields including music generation. We introduce a novel way of generating symbolic-domain music via adversarial training, which allows the discriminator to leak its information to the generator for better guidance. To prove good quality of the generated pieces, we conduct comparative test on LSTM model and evaluate the randomly-sampled music on five metrics in the light of statistics and music theory. The results shows that our model is more capable of generating coherent, natural and realistic music works.