Music genre classification using polyphonic timbre models

The increasing number of music available for download and subscriptions motivates the need for new solutions in organizing music for consumers. In this paper, several approaches for automatic genre classification of music using polyphonic timbre models are evaluated. Specifically, we compare the performance of the Gaussian mixture model (GMM), the Support Vector Machine (SVM), and the k-nearest neighbor (k-NN). Features are extracted to model the major attributes of timbre such as spectral envelope, range between tonal and noiselike character, and spectrotemporal evolution of sound. To address the scalability problem, a modified filter-and-refine method is integrated with the k-NN classifier. Results show that the 1-NN classifier with filter-and refine method achieved the highest classification accuracy on the GTZAN and ISMIR2004 datasets.