Automatic System for Music Genre Classification

Abstract-The huge amount of multimedia dataset, especially digital music, available on-line nowadays is still causing crucial challenges for both developers and users in order to effectively organize and categorize such information. In this paper, one develops a new automatic classification tool for genre classification based on features involving timbral texture features, Mel frequency cepstral coefficients and wavelet features. The classification scheme uses a combination of three classifies using majority voting rule. It consists of combining the outputs of K nearest Neighbour (KNN), Gaussian Mixture Model (GMM) and Neural Network (NNET) classifiers. Each of these classifiers involves training, testing and validation stages, which are appropriately chosen to ensure sufficient coverage of parameter values. The outcome produces an average overall accuracy rate of 83.5 % for ten music genres. By implementing additional pre-processing of data involving data normalization and principal component analysis (PCA), the overall accuracy over the ten music genres has peaked at an average of 86%, which corresponds to a significant improvement over similar works in the literature. I