Linear Prediction and the Spectral Analysis of Speech

Abstract

The report gives a detailed treatment of the use of linear prediction in speech analysis. New concepts are developed and more familiar concepts are seen in a new way. The covariance and autocorrelation methods are derived in the time and frequency domains. Both methods are shown to be derivable from a more general concept, that of generalized analysis-by-synthesis, where a nonstationary two-dimensional spectrum is approximated by another model spectrum. Linear prediction analysis is a special case where the model spectrum is all-pole. Also, under the assumption of stationarity the general covariance method reduces to the autocorrelation method.