Probabilistic Estimation of a Novel Music Emotion Model

An approach is proposed to estimate the emotional probability distribution of a novel music emotion model based on the updated Hevner's 8 emotion groups. Possible application includes browsing and mixing music with different emotion distributions. It is based on ground truths collected for 200 30-s clips (subdivided into 1200 segments further) chosen from soundtrack and labeled by 328 subjects online. Averagely, there are 28.2 valid emotional labeling events per clip, and constructing a probability distribution. Next, 88 musical features were extracted by 4 existing programs. The most discriminative 29 features were selected out by the pair-wise F-score comparison. The resultant 1200 segments were randomly separated into 600 training and 600 testing data, and input to SVM to estimate an 8-class probability distribution. They are finally evaluated by cosine, intersection, and quadratic similarity with the ground truth, where the quadratic metric achieves the best 87.3% ± 12.3% similarity.