Nonlinear backward ADPCM speech coding using kernel methods and online sparsification

Abstract

ABSTRACTModeling using neural networks and Volterra expansion are widely used in signal processing to account for nonlinear characteristics inherent in natural signals. In spite of their good performance in a wide range of applications, adaptive implementations of these models suffer from severe problems of a large number of parameters, slow convergence, and the risk of being trapped in local minima. Furthermore, inherent instability of the Volterra filters makes them unsuitable for applications that require synthesis or reconstruction, for example, speech coding. Nonetheless, kernel methods in the framework of reproducing kernel Hilbert spaces are emerging solutions that can tackle these problems by extending linear algorithms to spaces created by nonlinear mapping of the input signal. In this paper, we address the problems encountered in nonlinear adaptive coding by employing, for the first time, several kernel adaptive algorithms in the framework of speech coding using the adaptive differential pulse code modulation technique. Particular attention is focused on the effects of various sparsification techniques on the performance of the resultant algorithms. Simulation results show that utilizing kernel methods results in an average improvement of up to 3.4 dB in the signal‐to‐noise ratio of the reconstructed speech. This is while a lower improvement of up to 1.4 dB is achieved by exploiting Volterra filters at the cost of a demanding increase in the complexity, a slight increase in the bit‐rate and introducing a relatively long delay. Copyright © 2012 John Wiley & Sons, Ltd.