How sparsely can a signal be approximated while keeping its class identity?

This paper explores the degree of sparsity of a signal approximation that can be reached while ensuring that a sufficient amount of information is retained, so that its main characteristics remains. Here, sparse approximations are obtained by decomposing the signals on an overcomplete dictionary of multiscale time-frequency "atoms". The resulting representation is highly dependent on the choice of dictionary, decomposition algorithm and depth of the decomposition. The class identity is measured by indirect means as the speech/music discrimination power of features derived from the sparse representation compared to those of classical PCM-based features. Evaluation is performed on French Broadcast TV and Radio recordings from the QUAERO project database with two different statistical classifiers.