On the use of radial basis function networks for nonlinear speech processing

In this paper, nonlinear prediction of speech is performed by using a radial basis function (RBF) network. The RBF network has the parameters: (1) dimension of the input vector; (2) number of hidden nodes; (3) basis function; (4) centers; (5) widths and (6) weights to be selected. The complexity and the performance heavily depend on the input dimension and the number of centers. Reduction of complexity while maintaining the performance by decreasing both is still an open problem. The hierarchically self-organizing algorithm, which automatically adjusts the number of centers, with a dynamical systems approach is suggested as a solution. This algorithm is compared to standard approaches vis extensive computer simulations