Eight Music Emotions Recognition System using Neural Network with Cascaded model

Abstract

Music selection is difficult without an efficient organization based on metadata or tags, and one effective tag scheme is based on the emotion expressed by the music. The main drawback of such a system is that manually tagging music files because tagging a large number of files is a tedious work and emotional perception of each person is different. Therefore, this thesis presents a music emotion classification system for eight emotional classes with cascaded model. Russell's emotion model was adopted as a common ground for emotional annotation. The system implements on MATLAB using MIR toolbox to extract acoustic features from audio files and employed a supervised machine learning technique to recognize acoustic features to create predictive models. Four predictive models were proposed and compared. The models were composed by crossmatching two types of neural networks, i.e., Levenberg-Marquardt (LM) and resilient backpropagation (Rprop) with two types of structures: a traditional multiclass unit and multiple units of binary-class with a cascaded structure. The performance of each model was evaluated via the DEAM benchmark. The best result was achieved by the model trained with a cascaded Rprop neural network (accuracy of 89.5%). In addition, correlation coefficient analysis showed that timbre features were the most impactful for prediction. Our work offers an opportunity for a competitive advantage because only a few music providers currently tag music with emotional terms.