Independent features for content-based music genre classifiaction

International audience In this paper, we propose a new feature approach for genre classification of musical signals based on Independent Component Analysis. This approach take account of the fact that the redundancy of information between features may decrease the music genre classification rates. We thus introduce Independent Component Analysis to achieve independency of the features components. The new independent features are then used for genre classification through Support Vector Machine (SVM) or artificial neural network (ANN) classifiers. We will show through different experiments that this approach gives better accuracy rates than classical feature sets such as wavelet based, spectral, temporal or MFCC feature sets associated with different classifiers such as Multiclass SVM, Multilabel SVM and also ANN. These results are obtained with a database of 800 songs issued from the database of the Algerian radio. We thus obtain scores of 83% to 92% for eight genres. Interesting comparative results are reported and commented.