Spectral analysis of speech by linear prediction

The autocorrelation method of linear prediction is formulated in the time, autocorrelation, and spectral domains. The analysis is shown to be that of approximating the short-time signal power spectrum by an all-pole spectrum. The method is compared with other methods of spectral analysis such as analysis-by-synthesis and cepstral smoothing. It is shown that this method can be regarded as another method of analysis-by-synthesis where a number of poles is specified, with the advantages of noniterative computation and an error measure which leads to a better spectral envelope fit for an all-pole spectrum. Compared to spectral analysis by cepstral smoothing in conjunction with the chirp z transform (CZT), this method is expected to give a better spectral envelope fit (for an all-pole spectrum) and to be less sensitive to the effects of high pitch on the spectrum. The normalized minimum error is defined and its possible usefulness as a voicing detector is discussed.