All-pole spectral modelling of voiced speech with a highly compressed set of parameters

Quantisation of low-order all-pole models for the speech spectrum is analysed in the present study by comparing a new linear predictive method, Regressive Linear Prediction with Triplets (RLPT), to conventional linear prediction (LP). Effects of scalar quantisation with LARs (log-area-ratios) were analysed using both the all-pole spectra and the residuals of the prediction. It appeared that with only four bits describing the predictors, RLPT is able to model two resonances of the speech spectrum while conventional LP typically models only the overall structure of the spectrum.