A MIXTURE OF SUPPORT VECTOR MACHINES FOR AUDIO CLASSIFICATION
Nicolas Scaringella, D. Mlynek · Infoscience (Ecole Polytechnique Fédérale de Lausanne) · 2005
This paper describes the algorithm submitted by the authors to the Audio Genre Classification Contest organised in the context of the 2005 Music Information Retrieval Evaluation eXchange (MIREX 2005). The proposed algorithm parameterizes audio content by extracting 3 sets of features describing 3 different dimensions of music: timbre, energy and rhythm. Once features extracted, a mixture of Support Vector Machines (SVMs) is used for classification into musical genres. The underlying idea is to use separate models to approximate different parts of the problem and to combine the outputs from the experts with probabilistic methods. Using the proposed algorithm, classification of 73.11 % is achieved on the 2 databases used for the MIREX 2005 contest containing a total of 2929 songs.