Toward computer generated folk music using Recurrent Neural Networks

Abstract

In this paper, we compare the effectiveness of two different types of Recurrent Neural Networks, fully connected and Long Short Term Memory, for modeling music compositions. We compare both the categorical accuracies of these models as well as the quality of generated compositions, and find that the model based on Long Short Term Memory is more effective in both cases. We find that the fully connected model is not capable of generating non repeating note sequences longer than a few measures, and that the Long Short Term Memory model can do significantly better in some cases.