Music Mood Classification Using Audio Features

Abstract

Music emotions and moods are subjective and vary between people. Moods are categorized in to different groups where mainly divided into categorical and dimensional models. No widely accepted framework has been emerged due to ambiguity among the different moods. This work involved predicting four basic and major moods: anger, happy, neutral and sad which categorized under categorical model. Preparing ground truth dataset for analysis was part of the work due to unavailability of such dataset which are confirmed by expert judgment. Training dataset consists 800 music samples which includes 200 samples for each mood where testing dataset consists 100 samples by including 25 samples for each mood. 930 audio features are extracted through MIRToolbox using MATLAB and jAudio for feature selection. WEKA has been used as main tool for pre-processing, feature selection, experiment and classification tasks in addition to the Python sklearn. Supervised discretize, unsupervised normalize and unsupervised replace missing value filters have been used to pre-processes the original dataset. Three distinct original, discretized and normalized datasets are used to feature selection with WEKA attribute and wrapper subset evaluators. 59 datasets obtained through feature selection is prepared for classification with eight classification algorithms namely Naïve Bayes, LibSVM, SMO, IBK, AdaBoostM1, Bagging, 48 and Random Forest. WEKA experimenter is used to upload, configure algorithms, perform classification and compare the output results. According to the analysis, support vector machine algorithm LibSVM exhibited highest prediction accuracy while Bagging exhibits the highest average performance and the minimum standard deviation among the accuracy. LibSVM reported highest accuracy deviation as well as the lowest prediction results. Among three distinct datasets, only discretized datasets provided the better accuracy. During comparison, IBK, AdaBoostM1 and IBK not exhibit the prediction accuracy over 75%. Algorithms and discretized datasets predicated over 75% has been considered to perform optimization with different parameter setting obtain the better output results. Based on the results, SMO results over 77% accuracy under NormalizedPolyKernel and random seed seven