Speech processing using higher order statistics

The potential use of third-order statistics (TOS) in speech processing applications is examined. A source-system speech representation and analysis algorithm is considered. It computes the autoregressive (AR) parameters of speech from its third-order cumulants. The TOS-based AR estimates are compared with those obtained using second order statistics. Results are given voiced and unvoiced speech segments which are corrupted by additive Gaussian noise. The TOS-based AR estimation algorithm is also used to estimate the parameters of the short-term prediction filter of a code excited linear prediction (CELP) algorithm.