A new predictive method for all-pole modelling of speech spectra with a compressed set of parameters

This paper presents a new linear predictive method for speech processing, Linear Prediction with Linear Extrapolation (LPLE). The proposed method reformulates the computation of linear prediction by combining preceding samples of x(n) into pairs of consecutive samples. Each sample pair determines a line, which is then extrapolated at time instant n to be used as a data sample in the prediction. Consequently, from p normal equations the LPLE-technique yields an all-pole filter of order 2p, while the corresponding all-pole filter given by conventional linear prediction (LP) is of order p. It appears that LPLE is able to model the spectrum of voiced speech more accurately than conventional LP with small p-values. In this paper the effect of the prediction order is studied by varying the p-value from 1 to 12 for both conventional LP and the LPLE-method. Finally, the resulting all-pole spectra as well as the residual energies are compared