Nonlinear prediction of speech signals using radial basis function networks

In this paper, we compare the capabilities of various forms of radial basis function networks as nonlinear short-term predictors for speech signals representing sustained utterances of German vowels. We use RBF and RBF-AR¹ network architectures, trained using a standard algorithm or alternatively the extended Kalman filter (EKF) algorithm, and linear least squares predictors. We also look at cascaded forms of linear/nonlinear predictors. We evaluate both prediction gain and spectral flatness measure of the residual. The results indicate: The RBF-AR structure is the most powerful, EKF training yields better results than standard training for RBF networks, and a non-cascaded RBF-AR predictor produces results superior to cascaded predictors.