Unsupervised Music Segmentation with the Genetic Algorithm

Constructing reasonable structures of sequences is one of the most important problems. Especially, artificial sequences such as music and languages have some potential structures as human beings tend to build their constructions in creating. Although a recurrent neural network (RNN) has shown remarkable performance in processing sequences, it apparently cannot handle particularly long structures. This paper proposes a novel approach for building structures of sequences based on a genetic algorithm (GA) and a hierarchical RNN. The evaluation of the GA focuses on the negative log-likelihoods when the hierarchical RNN predicts the next elements. Thus, our proposal can be used for different types of sequences. In the experiments, musical notes are applied to this method, so the GA exhibits great structures minimizing the negative log-likelihoods and the hierarchical RNN shows better performance than stacked long short-term memory as a baseline.