Multi-Modal Music Mood Classification Using Co-Training

In this paper, we present a new approach to content-based music mood classification. Music, especially song, is born with multi-modality natures. But current studies are mainly focus on its audio modality, and the classification capability is not good enough. In this paper we use three modalities which are audio, lyric and MIDI. After extracting features from these three modalities respectively, we get three feature sets. We devise and compare three variants of standard co-training algorithm. The results show that these methods can effectively improve the classification accuracy.