A nonlinear acoustic echo canceller for hands-free telephony

In this paper, we propose an Acoustic Echo Canceller (AEC) composed by a cascade of a Pipelined Recurrent Neural Network (PRNN) and a linear filter. The PRNN is so called because of its recurrent and its modular processing. The Recurrent processing is used in order to more accurately model the room/speakerphone transfer function at high volumes. The linear subsection is a linear filter using a Fast Affine Projection Algorithm (FAP). FAP's key features include LMS like complexity and memory requirement (low), and RLS like convergence (fast) for the important case where the excitation signal is speech. Simulations results show, at high volume, an improvement in the echo return loss enhancement (ERLE) of 9 dB over a conventional linear adaptive filter with little additional computation.