Music-genre classification system based on spectro-temporal features and feature selection

Abstract

An automatic classification system of the music genres is proposed. Based on the timbre features such as mel-frequency cepstral coefficients, the spectro-temporal features are obtained to capture the temporal evolution and variation of the spectral characteristics of the music signal. Mean, variance, minimum, and maximum values of the timbre features are calculated. Modulation spectral flatness, crest, contrast, and valley are estimated for both original spectra and timbre-feature vectors. A support vector machine (SVM) is used as a classifier where an elaborated kernel function is defined. To reduce the computational complexity, an SVM ranker is applied for feature selection. Compared with the best algorithms submitted to the music information retrieval evaluation exchange (MIREX) contests, the proposed method provides higher accuracy at a lower feature dimension for the GTZAN and ISMIR2004 databases.