Multi-class classification of song emotions using Machine learning

Recent developments in Music Information Retrieval (MIR) have importantly brought several opportunities to enhance those applications that require music emotion recognition. Combining the terms ‘music’, ‘emotion,’ and ‘recognition’ firstly triggers the meaning that it is the detection of audio emotion in a musical form. The work in this paper uses the Music4All dataset and explores the possibility of audio metadata features-based music emotion recognition. The audio features, 13 in number, were initially analyzed to determine the emotion label. The audio emotion recognition task discussed in this paper is based on Russell's Circumplex model, which examines four classes of music emotion. The features which are relevant according to this model are valence and energy (same as arousal feature). Based on the range of values of valence and energy, three classifiers were built: Support Vector Machine (SVM), Logistic regression, and artificial neural networks (ANN). Logistic regression performed significantly better than the deep learning model by generating 99.9% accuracy in multi-class audio classification.