Frequency-Domain Analysis of Digital Speech Signals

Abstract

In this chapter, we reconsider many of the frequency-domain concepts discussed in earlier chapters in terms of digital analysis. We begin by discussing some properties of digital sinusoids and the discrete Fourier transform, which is the principal operation for deriving spectra from digital time signals. Having established the procedure for obtaining digital spectra, we can consider how they can be further parameterised to highlight some of the salient spectral differences between speech sounds. One of the parameterisations, known as cepstral analysis, will provide the first technique for separating the acoustic source from the filter. In the final part of this Chapter (6.6), we will reanalyse time-domain convolution in terms of frequency-domain multiplication, thereby providing another kind of model for relating the output to the filter and the source.