Dimensional music emotion recognition by valence-arousal regression

Abstract

As hot topics in current research, music emotion recognition (MER) have been addressed by different disciplines such as physiology, psychology, musicology, cognitive science, etc. In this paper, music emotions was modeled as continuous variables composed of valence and arousal values (VA values) based on Valence-Arousal model, and MER is formulated as a regression problem. 548 dimensions of music features were extracted and selected. The support vector regression, random forest regression and regression neural networks were adopted to recognize music emotion. Experimental results show that these regression algorithms achieved good regression effect. The optimal R2 statistics of values of VA values are 29.3% and 62.5%, which are achieved respectively by RFR and SVR in Relief feature space.