Music emotion recognition based on two-level support vector classification

Music emotion recognition (MER) detects the inherently emotional expression of people for a music clip. MER is helpful in music understanding, music retrieval, and other music-related applications. As volume of online musical contents expands rapidly in recent years, demands for retrieval by emotion have been emerging lately. Determining the emotional content of music computationally is an interdisciplinary research involving not only signal processing and machine learning, but the understanding of auditory perception, psychology, cognitive science, and musicology. One of the challenges in evaluating automatic music emotion detection is that there is currently no well-developed emotion model for music emotion description. Moreover, owing to the low transparency of the acoustic feature based music emotion recognizer, it is difficult to interpret data generated by this mechanism. In this study, a two-level classification system based on both music genre and music feature pre-described by the domain knowledge is proposed. This framework has the advantage of utilizing the most suitable acoustic information. Experiments will be conducted via the measure of the correlation between diverse emotional expressions and various musical cues. To verify the performance of overall system, the proposal model will also be evaluated based on the consistency between music features and ground-truth emotions.