Nonlinear echo cancellation using generalized power filters

This paper deals with the problem of nonlinear echo cancellation. One of the established approaches to this task is that of adaptive filtering using power filters. These filters model the unknown nonlinear system as a set of monomials, followed by a set of linear systems. The monomials are monotonously increasing functions useful for modeling nonlinearities such as soft-clipping, but only to a limited extent. We propose to replacing the monomials by a set of basis functions which resemble known real-world nonlinear distortions. Next, we propose performing the identification using the Multichannel Recursive Least Squares algorithm instead of the single-channel Normalized Least Mean Squares. In this way, the method can avoid the orthogonalization procedure used in the conventional power filter, thereby improving the reliability of the algorithm. We demonstrate the performance of the proposed algorithm in experiments with artificial and real-world data and compare it to several stateof- the-art methods.