A Modified Autocorrelation Method of Linear Prediction for Pitch-synchronous Analysis of Voiced Speech 2. Method

A modified autocorrelation method of linear prediction is proposed for pitch-synchronous analysis of voiced speech. The method needs one full period of speech data for analysis and assumes periodic extension of the data. This method guarantees the stability of the estimated all-pole filter and is shown to perform better than the covariance and autocorrelation methods of linear prediction. R6sum6. Pour l'analyse synchronis6e ~ la fondamentale de la parole vois6e, on propose une m6thode d'autocorr61ation modifi6e de la pr6diction lin6aire. La m6thode n6cessite une p6riode complete des donn6es pour l'analyse et est bas6e sur l'hypoth~se d'une extention p6riodique des don6es. Cette m6thode garantie la stabilit6 du filtre tout-pSle 6stim~, et il est montr6 qu'elle est meilleure que les m6thodes de covariance et d'autocorr61ation de la pr6diction lin6aire. 1. Motivation For pitch-synchronous analysis of voiced speech (where the analysis-segment duration is less than or equal to one pitch period), the autocorrelation method as well as the covariance method of linear prediction are unacceptable because of the following reasons. The performance of the auto-correlation method is not good [3], though it guarantees the stability of the estimated all-pole filter. The covariance method performs well, but it does not always lead to a stable all-pole filter [2]. So there is a need for a method for pitch-synchronous analysis of voiced speech which can perform as well as or better than the covariance method and can guarantee the stability of the estimated all-pole filter. In the present paper, we have proposed one such method. In the autocorrelation method of linear prediction , it is assumed that the signal is defined for all time such that it is identically zero outside a portion of the signal N samples long, where N is some positive integer [5, 6]. This is accomplished by weighting the speech signal by a finite window