Kernel-based identification of Hammerstein systems for nonlinear acoustic echo-cancellation

Traditional acoustic echo cancelers use a linear model to represent the echo path. Nevertheless, many consumer devices include loudspeakers and audio power amplifiers that may generate significant nonlinear distortions, creating the need for acoustic echo cancelers to produce a nonlinear filter response. To address this issue, we propose a nonlinear acoustic echo cancellation algorithm based on the framework of kernel methods. We model the echo path as a Hammerstein system, and we propose a resource-efficient strategy to identify the nonlinear and linear parts. While the basic algorithm is presented as an iterative batch method, we show that a simple extension allows it to be used in online scenarios as well. Results for both types of scenarios show that the algorithm produces good results on a system with a clipping nonlinearity and a realistic room impulse response.