A New Method Based on Deep Belief Networks for Learning Features from Symbolic Music

As the rapid increase of music data, Music Information Retrieval (MIR) have been receiving increasing attention in both the academic and commercial spheres. Feature extraction is a crucial part of many Music Information Retrieval (MIR) tasks. In recent years, deep learning approaches have gained significant interest as a way of learning a higher abstract representation from unlabeled data. In this paper, we present a system that can automatically extract relevant from symbolic music data. Firstly, The lower level features are extracted by using toolbox Music21, the higher level feature are then learned by a Deep Belief Network (DBN), finally the activations of the trained network as inputs for a non-linear Support Vector Machine (SVM) classifier. The experiment results demonstrate that the learned features obtain a better classification accuracy than other classical methods.