Machine Learning Approaches for Emotion Classification of Music: A Systematic Literature Review

Music is ubiquitous in every corner of the world. Fundamental mechanisms for how a listener perceives emotions in music as well as the way in which the emotional goal of a composer is expressed through music have been widely explored. Not only humans but also machines can recognize emotions in music once they are trained. As a result, various machine learning approaches are tested by previous researchers for their capability in music emotion recognition. To identify the current status of research and the research gaps in the above domain, we conducted a systematic literature review, through which we investigated several key aspects including types of music, acoustic features, feature extraction mechanisms, classification algorithms, and their performance. Six electronic databases were searched for studies published until 2018. Initial set of studies comprised of 198 studies from which 14 were selected for detailed analysis. Findings revealed that a considerable number of studies have used western music whereas other cultural-specific music still remains to be explored. Acoustic features pertaining to pitch, intensity, timbre, tempo, rhythm, melody, and harmony have been commonly used. While Support Vector Machine, Naive Bayes, and k-Nearest Neighbor are among the frequently used standard classifiers, Fuzzy classifiers and ensemble learning also have been attempted. The discrepancies among the classifier performances reported in previous studies could be partially attributed to the differences in the key aspects we have considered. The study makes a significant contribution by providing a comprehensive and up-to-date review of the previous attempts made in the selected domain.