Automatic Music Genre Classification of Audio Signals with Machine Learning Approaches

Musical genre classification is put into context byexplaining about the structures in music and how it is analyzedand perceived by humans. The increase of the music databaseson the personal collection and the Internet has brought a greatdemand for music information retrieval, and especiallyautomatic musical genre classification. In this research wefocused on combining information from the audio signal thandifferent sources. This paper presents a comprehensivemachine learning approach to the problem of automaticmusical genre classification using the audio signal. Theproposed approach uses two feature vectors, Support vectormachine classifier with polynomial kernel function andmachine learning algorithms. More specifically, two featuresets for representing frequency domain, temporal domain,cepstral domain and modulation frequency domain audiofeatures are proposed. Using our proposed features SVM act asstrong base learner in AdaBoost, so its performance of theSVM classifier cannot improve using boosting method. Thefinal genre classification is obtained from the set of individualresults according to a weighting combination late fusionmethod and it outperformed the trained fusion method. Musicgenre classification accuracy of 78% and 81% is reported onthe GTZAN dataset over the ten musical genres and theISMIR2004 genre dataset over the six musical genres,respectively. We observed higher classification accuracies withthe ensembles, than with the individual classifiers andimprovements of the performances on the GTZAN andISMIR2004 genre datasets are three percent on average. Thisensemble approach show that it is possible to improve theclassification accuracy by using different types of domainbased audio features