Agradecimentos: The authors would like to thank the University of Campinas and the Universidade Tecnológica Federal do Paraná for supporting this research through the DINTER agreement. Thanks to Prof. Dr. Rogerio Aparecido Gonçalves for lending GPUs to support large-scale SVM training. This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors Ver menos

Music Genre Classification is the problem of associating genre-related labels to digitised music tracks. It has applications in the organisation of commercial and personal music collections. Often, music tracks are described as a set of timbre-inspired sound textures. A subset of the sound textures is often selected to represent the entire track. In this paper, we evaluate the impact of texture selection on automatic music genre classification. Although previous work has selected textures by linear downsampling, no extensive work has been done to evaluate how texture selection benefits music genre classification. We also present a novel texture selector based on K-Means aimed to identify diverse sound textures within each track. Our results show that capturing texture diversity within tracks is important towards improving classification performance. Our results also indicate that our K-Means based texture selector is able to achieve significant improvements over the baseline with fewer textures per track than the other texture selectors evaluated. We also show that using multiple texture representations allows further opportunities for feature selection to improve classification performance Ver menos