Relative impulse response estimation during doubletalk with an artificial neural network-based step size control

Abstract

The Normalized Least-Mean Squares (NLMS) algorithm is a widely used method for linear system identification (e.g., for Acoustic Echo Cancellation (AEC), where the acoustic path between loudspeaker and microphone needs to be estimated). As soon as interferers or background noise are active, step size control becomes a crucial task in order to ensure a fast but stable adaptation. Conventional step size control methods address the case of additive noise contaminating the system output, i.e., microphone signal. When the NLMS algorithm is used for Relative Impulse Response (RIR) estimation, however, both the input signal and the output signal of the unknown system are noisy (since both of them are microphone signals) and a different step size control is required. In this paper, the derivation of a new step size is presented. Since the resulting step size cannot be determined directly, an Artificial Neural Network (ANN)-based estimation of the step size is proposed and its effectiveness is demonstrated.