Background on Speech Analysis

Abstract

After presenting the basic principles of speech analysis, we focus on the mathematical techniques which constitute the foundations of most of the methods currently in use in speech processing, such as the Fourier transforms and the linear prediction analysis. Then, we review typical parameter sets generally proposed to encode the speech signal prior recognition. While these methods give a reasonable representation of speech spectra, they do not provide a very accurate temporal localization of a signal’s spectral components. Two classes of techniques having the potential to deal with this problem, such as time-frequency analyses and wavelets, are presented. Finally, we address the problem of robust speech analysis and give a brief overview of the fields of higher-order spectral analysis and auditory modeling, illustrating our presentation with recent applications of these techniques to speech processing. We conclude this chapter by mentioning the limits of standard analysis methods in the presence of noise.