Spectral estimation of voiced speech with regressive linear prediction

A new predictive algorithm for the spectral estimation of voiced speech, Regressive Linear Prediction (RLP), is presented in this paper. Unlike conventional linear prediction (LP), RLP first combines the values that preceed the sample x(n) into groups of three consecutive samples (i.e., x(n-i), x(n-i-1), x(n-i-2)). Each of these p groups is then used to determine a regression line the value of which at time n are used as a data sample in the prediction. The prediction is optimised by minimising the square of the error between the original sample and the predicted value according to the autocorrelation criterion. With the same number p of unknowns in the normal equations, RLP yields an all-pole filter of order p+2, whilst the all-pole filter given by LP is of order p. The additional poles of the RLP filter improve the spectral modelling of voiced speech. This is shown in the present study by comparing the performance of conventional LP and the RLP method.