Music Emotion Recognition with the Extraction of Audio Features Using Machine Learning Approaches

Music is the covered up arithmetical exercise of a mind oblivious that it is figuring. Music not being just an extravagant language for human emotion is also a key itself for identifying the human emotion. Researches indicate that music causes stimulation through specific brain circuits to produce emotions. Listening to a piece of music can manipulate a person to feel joyous or brooding according to the emotion included in the music. But the perennial challenge is to examine the correlation between music and the subsequent effect on emotion. This music emotion recognition (MER) system can be used for simplistic music information retrieval. In this paper using (Music Information Retrieval) MIR Toolbox in Matlab, eight distinct features were extracted from 100 songs of various genres and similar emotions were clustered into four categories using the Russell's Two Dimensional Emotion Model. Mapping the extracted features into the four emotion classes, several machine-learning classifiers were trained. A set of unknown songs were used to validate the recognition accuracy. Along with the common features like pitch, timbre, rhythm etc. roll-off and brightness were also used. Roll-off showed a great priority in Random Forest feature ranking. With all these features combined, a highest prediction accuracy of 75% was found from artificial neural network (ANN) among the others classifiers like Support Vector Machine (SVM), linear discriminant, and Ensemble learner.