A new linear predictive method for spectral estimation of voiced speech

Abstract

A new linear predictive method for analysis of voiced speech is presented. The new technique, Separated Linear Prediction (SLP), is based on predicting sample x(n) from its previous samples as in conventional Linear Prediction (LP). SLP, when compared to conventional LP-analysis, separates p+1 previous samples into two groups: (a) x(n-1) and (b) x(n-1-i), 1/spl les/i/spl les/p. Sample x(n-1) is treated differently since it most likely has the highest correlation with sample x(n). By using linear extrapolation between x(n-1) and each of the samples in group (b) a new prediction model is formulated. By minimizing the square of the prediction error an optimal SLP-predictor is derived. When analyzing voiced speech it shown that SLP yields more accurate higher formants in comparison to conventional LP.