Popular music representation: chorus detection & emotion recognition

This paper proposes a popular music representation strategy based on the song’s emotion. First, a piece of popular music is decomposed into chorus and verse segments through the proposed chorus detection algorithm. Three descriptive features: intensity, frequency band and rhythm regularity are extracted from the structured segments for emotion detection. A hierarchical Adaboost classifier is employed to recognize the emotion of a piece of popular music. The general emotion of the music is classified according to Thayer’s model into four emotions: happy, angry, depressed and relaxed. Experiments conducted on a 350-popular-music database show the average recall and precision of our proposed chorus detection are approximately 95 % and 84 %, respectively; and the average precision rate of emotion detection is 92 %. Additional tests are performed on songs with cover versions in different lyrics and languages, and the resultant precision rate is 90 %. The proposes approaches have been tested and proven by the professional online music company, KKBOX Inc. and show promising performance for effectively and efficiently identifying the emotions of a variety of popular music.