Nonlinear echo cancellation using decoupled A-B net structure

Abstract

This paper proposes a general nonlinear digital filter structure for echo cancellation applications. Although echo cancelers employing linear digital filter structures are more widely used, there are many applications where nonlinear filters must be used. We propose using the DABNet (Decoupled A-B Net) filter, which is composed of a decoupled linear dynamic system followed by a nonlinear static map, for echo cancellation. The linear dynamic system is initially spanned by a set of discrete Laguerre systems, and then cascaded with a single hidden layer perceptron. A model reduction technique can be performed not only to identify the main time constants, but also to reduce the dimensionality of the perceptron input. The DABNet structure is able to approximate any nonlinear, causal, discrete time invariant, multiple-input single output system with fading memory. Comparisons between echo cancelers implemented with the DABnet and nonlinear FIR filters are presented.