On feature combination for music classification

Abstract

We address the problem of combining different types of audio features for music classification. Several feature-level and decision-level combination methods have been studied, including kernel methods based on multiple kernel learning, decision level fusion rules and stacked generalization. Eight widely used audio features were examined in the experiments on multi-feature based music classification. Results on benchmark data set have demonstrated the effectiveness of using multiple types of features for music classification and identified the most effective combination method for improving classification performance. © 2010 Springer-Verlag Berlin Heidelberg.