Non Negative Matrix Factorization for Music Emotion Classification

Abstract

Classification of emotion is a fundamental problem in music information retrieval where it addresses the query and retrieval of desirable types of music from large music data set. Until recently, there are only few works on music emotion classification that are carried out by incorporating instrumental and vocal timbre. Generally, vocal timbre alone can be used in distinguishing emotion in music but it became less effective when mixed with the instrumental part. Thus, a new research interest has led to identifying instrumental and vocal timbre as features capable of influencing human affect and analysis of sounds in regards to their emotional content. In this research, non-negative matrix factorization (NMF) is applied to separate music into both instrumental and vocal components. Extracted timbre features from audio using signal processing technique will be used to train and test artificial neural network (ANN) classifier. The ANN learn from supervised and unsupervised training to classify the emotional contents in music data as sad, happy anger or calm. The efficiency of the ANN classifier is verified by a subjective test including inputs from annotators by manual categorization of the audio data. The efficiency of this method reached up to 90 %.