Music genre classification using multiscale scattering and sparse representations

An effective music genre classication approach is proposed that combines the translation-invariance and deformation-robustness property of scattering coefficients and the discriminative power of sparse representation-based classifiers. We argue that these two approaches to feature selection and classification complement each other in reducing the in-class variability of data, and this should lead to enhanced performance. Our results show clear improvement over a variety of previous approaches. A music genre classication accuracy of approximately 91.2% on the GTZAN database is reported.