A spectral distance measure for speech detection in noise and speech segmentation

Abstract

A new spectral distance measure is introduced and its properties explained. This measure is especially designed to evaluate distances between spectral densities, and presents important properties, such as invariance to scaling factors or shifts in amplitude. The measure may be used as a test for whiteness, to determine the similarity between independent processes, or to check the quasi-stationarity condition in a single process. Its special ability to detect spectral similarities may be exploited for speech segmentation and in the detection of speech under strong noise levels, and may be used in end-point detection applications. The fundamentals of the measure are given, some case studies are described and the results discussed.