A linear predictive method using extrapolated samples for modelling of voiced speech

A new method, linear prediction with extrapolated samples (LPES), is proposed for spectral estimation of voiced speech. In LPES, the nth sample of signal x(n) is predicted using its p+1 preceding samples by forming p lines of each two consecutive samples from those p+1 preceding samples. The values of these p lines are extrapolated at time instant n and treated as original p+1 preceding samples in conventional linear prediction (LP). The square of the prediction error is minimised using the autocorrelation criterion. LPES yields an all-pole filter whose order is one larger than the number of unknowns in the normal equations. This results to more accurate spectral estimation for voiced speech, especially at higher frequencies.