Audio representations on overcomplete sets

Abstract

We present in this paper the advantages of using overcomplete sets for the representation and coding of audio signals. More specifically, we have investigated hybrid signal models of the type {tonal+transients+stochastic residual} where these features can be simultaneously present at all times. The extraction of the two "deterministic" components (tonal and transient parts) is made in such a way that each of them can be represented in a compact way in appropriate orthogonal basis (MDCT and dyadic wavelets, respectively). The use of structured representations significantly reduces the potentially high cost of encoding of significance maps. Preliminary results show that this approach is well-adapted for transform coding of very high-quality sounds.