Multi-Scale Approaches to the MediaEval 2015 "Emotion in Music" Task

The goal of the “Emotion in Music” task in MediaEval 2015 is to automatically estimate the emotions expressed by music (in terms of Arousal and Valence) in a time-continuous fashion. In this paper, considering the high context correlation among the music feature sequence, we study several multiscale approaches at different levels, including acoustic feature learning with Deep Brief Networks (DBNs) followed a modified Autoencoder (AE), bi-directional Long-Short Term Memory Recurrent Neural Networks (BLSTM-RNNs) based multi-scale regression fusion with Extreme Learning Machine (ELM), and hierarchical prediction with Support Vector Regression (SVR). The evaluation performances of all runs submitted are significantly better than the baseline provided by the organizers, illustrating the effectiveness of the proposed approaches.