Music Emotion Regression based on Multi-modal Features1

Music emotion regression is considered more appropriate than classification for music emotion retrieval, since it resolves some of the ambiguities of emotion classes. In this paper, we propose an AdaBoost-based approach for music emotion regression, in which emotion is represented in PAD model and multi-modal features are employed, including audio, MIDI and lyric features. We first demonstrate the effectiveness of our approach, and then focus on exploring the contribution of individual modalities to the regression of each emotion dimension. A series of experiments show that lyric contributes the most to the regression of emotion dimension P, while audio and MIDI contribute more to the regression of dimension A and D. Thinking that the three modalities provide complementary information from different angles, we combine them and show that the best regression performance is obtained when all modalities are used.