Estimation of short-time spectral envelope of voiced speech using maximum entropy method of spectral analysis

Abstract

The Burg maximum entropy (ME) method of spectral analysis is used to estimate the short-time spectral envelope of the voiced speech signal. A number of vowel segments from continuous speech are analyzed using this method. The Burg ME method is compared with the autocorrelation and covariance methods of linear prediction using normalized linear prediction error and accuracy in estimating the speech spectrum as criteria. It is shown that the Burg ME method consistently gives better results than the autocorrelation method of linear prediction for both pitch-synchronous and pitch-asynchronous analyses. For pitch-synchronous analysis, this method performs slightly better than the covariance method, while for pitch-asynchronous analysis, its performance is marginally inferior to that of the covariance method.