A robust sequential parameter estimation for time-varying speech signal analysis

In the conventional Kalman estimator (CKE), the ability to track the time-varying formants could be improved by using a forgetting factor. However, due to such a forgetting factor, the CKE becomes very sensitive to the impulsive pitch excitations and produces poor parameter estimates. To reduce the effects of the impulsive pitch excitations, the authors propose the robust Kalman estimator (RKE) with a forgetting factor, in which the error signal is assumed to have a normal mixture distribution. From the simulation results on the synthetic speech data, it is found that this proposed robust Kalman estimator could be well utilized as a sequential parameter estimation method