Pipelined nonlinear spline filter for speech prediction

Abstract In this paper, a new pipelined nonlinear spline adaptive filter (PNSF) is presented for the speech prediction application. The proposed architecture is essentially an improved pipelined cascade model, where each module consists of a FIR filter followed by a spline activation function. Based on minimum mean square error cost and stochastic gradient method, the on-line learning adaptive algorithms for updating the nonlinear and linear weights are derived. We analyze the selection range of the learning rate involved in the learning algorithms to ensure the convergence of the algorithms. Simulations are carried out to evaluate the performance of the PNSF on nonlinear system identification and speech prediction. Experimental results show that the PNSF provides better performance compared to the spline adaptive filter (SAF), joint process filter using pipelined second-order Volterra filter (JPPSOV) and pipelined neural IIR (PNIIR) models.