Automatic musical genre classification using sparsity-eager support vector machines

Constructing robust categorical and typological classifiers, i.e., finding auditory constructs utilized for describing music categories, is an important problem in music genre classification. Supervised methods such as support vector machine (SVM) achieve state of the art performance for genre classification but suffer from over-fitting on training examples. In this paper, we introduce a supervised classifier, ℓ₁-SVM, that utilizes sparse methods to deal with over-fitting for genre classification. We compare the proposed algorithm to competing learning methods such as SVM, logistic regression, and ℓ₁-regression for genre classification. Experimental results suggest that the proposed method using short-time audio features (MFCCs) outperforms the baseline algorithms in terms of the average classification accuracy rate of musical genres.