Low-level feature selection for emotional music preferences based on subjective audience rating

Recently, many researches of modeling or measuring human feeling have been conducted to understand human emotions. However, researches on music-related human emotions have much difficulty due to the subjective perception of emotions. We have selected low-level musical features which may trigger human emotions, based on subjective audience ratings. This experiment is based on the subjective audience ratings of five hundred participants in a very popular Korean TV music program. In this program, audience is requested to rate music of the contestants and to select their preferred music based on their emotional feelings. The most relevant low-level features with respect to human emotions are selected by backward elimination method and experimental results show that selected low-level features touch human emotions quite positively.