Regressive linear prediction with triplets-an effective all-pole modelling technique for speech processing

This paper presents a new linear predictive method for speech processing, Regressive Linear Prediction with Triplets (RLPT). The RLPT-algorithm yields from p normal equations an all-pole filter of order 2p+1 (i.e., an all-pole filter of order 2p+1 is defined from p numerical values). In comparison to conventional linear prediction of order p RLPT takes into account 2p+1 preceding samples of x(n) in the computation of the linear prediction. Consequently, the obtained all-pole filter is able to model the spectrum of voiced speech more accurately than conventional linear prediction, especially with very small p-values. The experiments show that the RLPT-method is effective in finding one or two lowest resonances in voiced speech spectra when p equals 1 or 2