Searching for dominant high-level features for music information retrieval

Music Information Retrieval systems are often based on the analysis of a large number of low-level audio features. When dealing with problems of musical genre description and visualization, however, it would be desirable to work with a very limited number of highly informative and discriminant macro-descriptors. In this paper we focus on a specific class of training-based descriptors, which are obtained as the loglikelihood of a Gaussian Mixture Model trained with short musical excerpts that selectively exhibit a certain semantic homogeneity. As these descriptors are critically dependent on the training sets, we approach the problem of how to automatically generate suitable training sets and optimize the associated macro-features in terms of discriminant power and informative impact. We then show the application of a set of three identified macro-features to genre visualization, tracking and classification.