Evaluation of Gated Recurrent Neural Networks in Music Classification Tasks

In this paper, we evaluate two popular Recurrent Neural Network (RNN) architectures employing the mechanism of gating: Long-Short Term Memory (LSTM) and Gated Recurrent Unit (GRU), in music classification tasks. We examine the performance on four datasets concerning genre, emotion and dance style recognition. Our key result is a significant improvement of classification accuracy achieved by training the recurrent network on random short subsequences of the vector sequences in the training set. We examine the effect of this training approach on both architectures and discuss the implications for the potential use of RNN in music information retrieval.