Speech parameter estimation by time-weighted-error Kalman filtering

In this paper, we discuss a method of improving the parameter-tracking performance of the Kalman filter for modeling time-varying signals. The Kalman filter is an effective means of recursively estimating the coefficients of an AR (or ARMA) model; however, its effectiveness is diminished by the weight which the filter gives to the history of the signal. With a view toward improved modeling of speech signals, we examine the use of a time-weighted error criterion to remedy this situation.