In acoustic echo cancellation (AEC) an attempt is made to cancel the echo path between the loudspeaker and the microphone by a linear adaptive filter. In real world applications, the loudspeaker microphone enclosure introduces nonlinear distortions into the echo path. The majority of the nonlinear distortions are from overdriven loudspeaker signals. Adaptive filters can only cancel the linear portions of the echo path, thus the non-linear portions cannot be removed via the adaptive filter. Later, it will be discussed how an adaptive algorithm will react to these non-linear distortions, and how residual echo suppressors are used to eliminate unmodeled parts of the echo path.
There are many approaches to echo cancellation. Common approaches include time-domain normalized least means squares (NLMS), affine projection algorithm, recursive least squares and transform-domain (block processing) NLMS. As expected not all adaptive algorithms behave the same to non-linear distortions. How an adaptive algorithm reacts to these disturbances is similar to how these algorithms react to doubletalk situations. In other words, the faster an algorithm converges the more it is affected by nonlinearities. Block processing algorithms have worse performance in regards to convergence during nonlinearities because a when the system diverges a whole block of data gets degraded.
Non-linear processing is the suppression of residual echo left by the linear adaptive filter. Suppression approaches are more applicable than nonlinear adaptive filters because of the slowed convergence and high computational complexity. Center clipping and adaptive gain suppression are two approaches to non-linear processing. In center clipping, when the level of the microphone signal after the adaptive filter is below the excitation signal by a tunable threshold, it can be determined that this signal is related to echo and can be clipped. The downfall to this approach is that if the near-end speaker level is too low, it will be clipped during doubletalk. In adaptive echo suppression, the amount of attenuation applied is related to the level of the estimated residual echo. The difficulty of this approach is that an accurate estimate of the residual echo level, which is not explicitly available, is required.