Testing a variety of features for music mood recognition

Abstract

Music collections are organized in a very different way depending on a target, number of songs or a distribution method, etc. One of the high-level feature, which can be useful and intuitive for listeners, is “mood.” Even if it seems to be the easiest way to describe music for people who are non-experts, it is very difficult to find the exact correlation between physical features and perceived impressions. The paper presents experiments aimed at testing a variety of low-level features dedicated to music mood recognition. Musical excerpts to be tested comprise individual (solo) tracks and mixes of these tracks. First FFT- and wavelet-based analyses, performed on musical excerpts, are shown. A set of “energy-based” parameters is then proposed. These are mainly rms coefficients normalized over the total energy derived from wavelet- based decomposed subbands, variance and some statistical moments. They are then incorporated into the feature vector describing music mood. Further part of experiments consists in testing to what extent these features are correlated to the given music mood. Results of the experiments are shown as well as the correlation analysis between two main mood dimensions—Valence and Arousal assigned to music excerpts during the subjective tests.