Automatic low-dimensional analysis of audio databases

In this paper we present an approach designed to map variable size audio sequences into fixed-length vectors, useful to discover contents of audio databases. First, we model standard audio parameters with Gaussian mixture models (GMM). Then, symmetric Kullback-Leiber divergences between models are approximated with a Monte-Carlo method. We use these statistical dissimilarities to find a low-dimensional representation of each audio sequence through Multidimensional scaling (MDS) algorithm. Vectors in low-dimensional spaces are then easily explored with kernel and clustering methods. Experiments carried out in different kind of audio databases (music, speakers and languages) show good potential of the proposed approach and provide a framework for more challenging applications.