Music Genre Classification: A Multilinear Approach

In this paper, music genre classification is addressed in a multilinear perspective. Inspired by a model of auditory cortical processing, multiscale spectro-temporal modulation features are extracted. Such spectro-temporal modulation features have been successfully used in various content-based audio classification tasks recently, but not yet in music genre classification. Each recording is represented by a third-order feature tensor generated by the auditory model. Thus, the ensemble of recordings is represented by a fourth-order data tensor created by stacking the third-order feature tensors associated to the recordings. To handle large data tensors and derive compact feature vectors suitable for classification, three multilinear subspace techniques are examined, namely the Non-Negative Tensor Factorization (NTF), the High-Order Singular Value Decomposition (HOSVD), and theMultilinear Principal Component Analysis (MPCA). Classification is performed by a Support Vector Machine. Stratified cross-validation tests on the GTZAN dataset and the ISMIR 2004 Genre one demonstrate the advantages of NTF and HOSVD versusMPCA. The best accuracies obtained by the proposed multilinear approach is comparable with those achieved by state-of-the-art music genre classification algorithms.