Acoustic echo cancellation in non-linear and noisy environments

This paper presents a new assessment of four approaches to acoustic echo cancellation in the presence of non-linear echo. The comparison is performed with algorithms that are configured to give equivalent performance under linear-only echo conditions, thereby giving a more meaningful assessment. We also compare the effect of non-linear echo to that of noise and show that, whilst performance differs in non-linear environments, there are negligible differences in noisy environments. The computationally efficient FBLMS algorithm is shown to perform as well as other algorithms in noisy environments but performs poorly under non-linear echo conditions. We also show how the correlation between non-linearities and the speech signal can corrupt the echo path estimate and leads to small attenuation of non-linear components at the expense of reduced attenuation for linear echo. Finally we discuss the merits of modelling non-linearities as a linear environment with a time varying echo path.