A perceptually reweighted mixed-norm method for sparse approximation of audio signals

In this paper, we consider the problem of finding sparse representations of audio signals for coding purposes. In doing so, it is of utmost importance that when only a subset of the present components of an audio signal are extracted, it is the perceptually most important ones. To this end, we propose a new iterative algorithm based on two principles: 1) a reweighted 1-norm based measure of sparsity; and 2) a reweighted 2-norm based measure of perceptual distortion. Using these measures, the considered problem is posed as a constrained convex optimization problem that can be solved optimally using standard software. A prominent feature of the new method is that it solves a problem that is closely related to the objective of coding, namely rate-distortion optimization. In computer simulations, we demonstrate the properties of the algorithm and its application to real audio signals.