A regression approach to categorical music emotion recognition

Music emotion recognition is a challenging task due to the subjective issue. Simply assigning an emotion class to a music clip is deterministic due to ignoring the individual difference when people perceive music emotion. Also, within a music clip, there always exist more than one kind of emotions. In this paper, we propose a regression approach for categorical music emotion recognition. For each music clip, based on the features extracted by various algorithms, the proposed approach determines the top two dominant emotions in it. Eight regressors for eight emotion classes are trained respectively to score the intensity of each emotion in a music clip. An experiment on a data set of 385 music clips annotated by 105 subjects is performed to demonstrate the effectiveness of the proposed method.