Autoregressive spectral estimation in noise with application to speech analysis

An improved "weighted information" method of spectral estimation has recently been introduced [1,2]. The method treats the problem of estimating autoregressive (AR) process parameters from discrete time observations corrupted by additive independent noise, with known power spectral density. The estimation procedure is cast as the problem of minimizing a weighted information measure and has a theoretical foundation relating it to asymptotic maximum likelihood and minimum information divergence procedures. This paper develops several topological properties of the weighted information measure. Computational procedures appropriate to the estimation problem are discussed elsewhere [1,2]; procedures appropriate to the detection ("vector quantization") problem are presented here. Simulation results and a real speech example demonstrate the advantages of the method.