Double linear prediction

A new method to achieve a reduction in the number of poles of the linear predictive filter is described. This method is called double linear prediction since the method consists of solving consecutively two sets of autocorrelative normal equations, where the second set is derived from the first. A first "pass" of the linear predictive analysis is done on \´pre-empasized" speech and yields a p-pole filter from which the impulse response is derived. A "suitable" portion of this impulse response is selected and windowed. Then autocorrelative, linear predictive analysis is again applied in a second "pass", where now the number of coefficients is much less than in the first "pass," i.e., ususally p/2 or (p/2)+2 depending on the number of formants present. The method is discussed and examples are presented illustrating the procedure.