The amount of music information available in the Internet and portable devices applications is huge and has considerably increased over the last decade. This makes even more crucial the need of organizing, compressing, representing and indexing large volume of music data. In this context, musical styles are particular interesting descriptors, since they summarize patterns between music and are used for years to organize large databases. This project proposes the application of pattern recognition and machine learning methods in order to classify musical styles. To this end, it is aimed to verify the contribution of measures extracted from the melody and from the percussion of different music styles. The discriminative power of the extracted features will be verified with the use of supervised classification and clustering techniques. In addition, it is proposed to aggregate graphs modeling with the purpose of enhancing the classification performance. Supervised classification will be performed by the k-nn and Gaussian classifiers, while clustering will be performed by k-means and hierarchical clustering. The performance evaluation will be conducted through metrics derived from the confusion matrix, as, for example, the Kappa coefficient and the classification accuracy.
News published in Agência FAPESP Newsletter about the scholarship: