Skip to Main Content
We present a new approach to joint state and parameter estimation for a target-directed, nonlinear dynamic system model with switching states. The model, recently proposed for representing speech dynamics, is called the hidden dynamic model (HDM). The model parameters, subject to statistical estimation, consist of the target vector and the system matrix (also called "time-constants"), as well as parameters characterizing the nonlinear mapping from the hidden state to the observation. We implement these parameters as the weights of a three-layer feedforward multilayer perceptron (MLP) network. The new estimation approach is based on the extended Kalman filter (EKF), and its performance is compared with the traditional expectation-maximization (EM) based approach. Extensive simulation results are presented using both approaches and under typical HDM speech modeling conditions. The EKF-based algorithm demonstrates superior convergence performance compared with the EM algorithm, but the former suffers from excessive computational loads when adopted for training the MLP weights. In all cases, the simulated model output converges to the given observation sequence. However, only in the case where the MLP weights or the target vector are assumed known do the time-constant parameters converge to their true values. We also show that the MLP weights never converge to their true values, thus demonstrating the many-to-one mapping property of the feedforward MLP. We conclude that, for the system to be identifiable, restrictions on the parameter space are needed.