Affect prediction in music using boosted ensemble of filters

Music influences the affective states of its listeners. For this reason, music is extensively used in various media forms to enhance and induce emotional feeling. Automatic evaluation of affect from music can have impact on music design and can also aid further analysis of music. In this work, we present a novel scheme for affect prediction in music using a Boosted Ensemble of Single feature Filters (BESiF) model. Given a set of frame-wise features, the BESiF model predicts the affective rating as a weighted sum of filtered feature values. The BESiF model improves the Signal to Noise Ratio for arousal and valence prediction by a factor of 1.92 and 1.06, respectively, over the best baseline method. This performance is achieved using only 14 signal features for arousal (16 for valence). We further analyze the transformation of one of the features selected towards arousal prediction.