Music Genre Classification Using Music Information Retrieval and Self Organizing Maps

Abstract

Recent times have seen a wide proliferation of use of Genre as basis of music classification and documentation. This paper proposes a novel approach of music genre classification using unsupervised Neural Network method—Kohonen Self Organizing Maps (SOM). Various music features like timbrel features (attack slope), spectral distribution of energy (brightness) and tonality (major or minor) were extracted from the audio files using MATLAB toolbox. An inventory of a number of musical pieces was developed for building clusters of different genre. A database containing the extracted musical features was then clustered using SOM. Each cluster encapsulated a different genre of music. The centroid of each cluster was taken as the representative point of that genre in the considered n dimensional musical feature space. Then a number of songs were then mapped onto the said space and classified into different genre depending on their Euclidean distance from the center of each cluster.