Music Emotion Recognition with Machine Learning Based on Audio Features

Abstract

Understanding the emotional impact of music on its audience is a common field of study in many disciplines such as science, psychology, musicology and art. In this study, a method based on acoustic features is proposed to predict the emotion of different samples from Turkish Music. The proposed method consists of 3 steps: preprocessing, feature extraction and classification on selected music pieces. As a first step, the noise in the signals is removed in the pre-process and all the signals in the data set are brought to the equal sampling frequency. In the second step, a 1x34 size feature vector is extracted from each signal, reflecting the emotional content of the music. The features are normalized before the classifiers are trained. In the last step, the data are classified using Support Vector Machines (SVM), K-Nearest Neighbor (K-NN) and Artificial Neural Network (ANN). Accuracy, precision, sensitivity and F-score are used as classification metrics. The model was tested on a new 4-class data set consisting of Turkish music data. 79.30% Accuracy, 78.77% sensitivity, 78.94% specificity and 79.03% F-score are obtained from the proposed model.