Description and generation of spherically invariant speech-model signals

Abstract

In this paper, spherically invariant random processes (SIRPs) are used as stationary models for bandlimited speech signals in telephone channels. These processes are completely described by the univariate probability density function (PDF) together with the covariances of their amplitudes. A comprehensive mathematical treatment is achieved by means of Meijer's G-functions. An algorithm that has been implemented on a 16-bit signal processor (TMS320) is given, which generates SIRP-like signals with independently selectable PDF and correlation. The statistical features of these signals coincide extremely well with those derived for SIRPs. Furthermore, it is demonstrated that both the correlation as well as the univariate PDF can completely be adjusted to the stochastic parameters measured from real speech signals.