SSCL: Music Generation in Long-term with Cluster Learning

Deep neural networks have been used in music related applications such as polyphony music generation, music style transfer and hidden emotions identification. However, it's still challenge to generate long-term music with harmonious structures. In this paper, we propose a Self-similarity Clustering Learning (SSCL) model to meet the above challenge. Our main idea is to reduce the dimensionality of the self-similarity matrix followed by clustering, so that the generated music samples in one cluster have similar long-term strucutres. We built a music dataset by collecting 198 hours music data from Internet to evaluate the quality of the generated music. Experimental results show that our model can synthesize music with better long-term properties which are closer to real music.