Robust sample-selective LP parameter estimation in USFS1016 coder based on quadratic classifier

Abstract

In order to decrease the LP spectral degradation in the USFS1016 4.8 kb/s CELP speech coder, the application of robust sample-selective LP parameter estimation procedures based on a quadratic classifier of nonstationary signals is proposed. Additionally, applications of the quadratic classifier with optimal and heuristically decision thresholds are considered. Comparative experimental analysis is done based on the cepstral distance as an objective spectral measure. The obtained results show that the proposed robust procedures achieve a smaller LP spectral degradation in the standard CELP coder, compared to standard LP methods, sample-selective methods, and robust method based on Huber's (1994) M-estimation theory.