SEPARATED LINEAR PREDICTION - A NEW ALL-POLE MODELING TECHNIQUE FOR SPEECH ANALYSIS

This study presents a new predictive method, Separated Linear Predicti on (SLP), for spectral estimation of speech. The prediction of the sig nal value x(n) is computed from its p + 1 preceding samples by emphasi sing among the previous samples the one which is located next to sampl e x(n). All;he p samples from x(n - 2) to x(n - (p + 1)) are linearly extrapolated with x(n - 1) to obtain p new values, which are used for prediction. Optimisation of the filter coefficients is computed using the autocorrelation criterion as in conventional LP. The SLP-analysis yields an all-pole filter of the order p + 1 with p unknowns in the no rmal equations. The performance of SLP was compared to conventional LP by analysing vowels produced by two female and four male speakers. Re sults showed that the proposed method yielded in general more accurate modelling of higher formants, residuals with smaller energy and incre ased flatness. (C) 1998 Elsevier Science B.V. All rights reserved.