Hammerstein model-based nonlinear echo cancelation using a cascade of neural network and adaptive linear filter

We present a novel nonlinear echo cancellation method that assumes the Hammerstein nonlinear system model. Model parameters are identified using a neural network followed by an adaptive linear filter. The parameters of both subsystems are estimated separately, which allows the utilization of computationally efficient conventional methods. The proposed method is verified on simulated as well as on real-world signals. In comparison to power-filter echo cancelers, the method achieves significantly higher echo suppression provided that the excitation signal is white noise, and it yields comparable performance when the signal is speech.