Automatic emotion classification of musical segments

In this paper a novel approach is proposed to automatically classify the emotions of musical segments. The possible applications include the music recommendation system for the public or specialists like advertisers, physiotherapists, DJs, and film directors. The first step of this approach is to select musical segments from 75 musical works of different genres that may be expressive emotionally. Emotioanl labels are obtained by letting 60 subjects record their emotional appraisals (valence-arousal) by software FeelTrace when they listen to these musical works. Next, we extract 55 low-and-middle-level musical features of those 75 segments by programs MARSYAS and PsySound. The 55 musical features are then processed by feature-selection methods (MANOVA) to reduce the dimension of the feature space and the significance of features are determined according to their rank in performance. Final results by classification algorithm (Support Vector Machine) for 4-class (the four quadrants of the valence-arousal emotional space) classification are compared by using different combination of useful subsets of features. The best accuracy of classification is very satisfactory (98.67 %) by leave-one-out crossvalidation. From the results we also find some interesting relationships between the emotion class and the 55 musical features by comparing with musicians’ rules of thumb for music-emotion cues.