Evaluation of Audio Feature Groups for the Prediction of Arousal and Valence in Music

Computer-aided prediction of arousal and valence ratings helps to automatically associate emotions with music pieces, providing new music categorisation and recommendation approaches, and also theoretical analysis of listening habits. The impact of several groups of music properties like timbre, harmony, melody or rhythm on perceived emotions has often been studied in literature. However, only little work has been done to extensively measure the potential of specific feature groups, when they supplement combinations of other possible features already integrated into the regression model. In our experiment, we measure the performance of multiple linear regression applied to combinations of energy, harmony, rhythm and timbre audio features to predict arousal and valence ratings. Each group is represented by a smaller number of dimensions estimated with the help of Minimum Redundancy–Maximum Relevance (MRMR) feature selection. The results show that cepstral timbre features are particularly useful to predict arousal, and rhythm features are the most relevant to predict valence.