Stepsize control in NLMS acoustic echo cancellation using a neural network

Correctly estimating the best filter adaptation rate in acoustic echo cancellation (AEC) is known to be a complex problem justifying advanced algorithms. In this paper we illustrate a new neural network estimator based on a set of classic statistical estimators and a proper, mostly generalized, offline training. The output step size estimated is shown to give good convergence speed and to correctly behave in case of local noise, "double talk" and changes in the room impulse response. Complexity is evaluated and proves acceptable for actual implementations, and results are compared with both a simple NLMS (normalized least mean square) implementation and an advanced fuzzy methodology which are also known in the literature.