Uncertainty modeling in acoustic echo control

Acoustic echo control (AEC) is a crucial component of hands-free voice interfaces. While the adaptive echo canceler provides a seemingly ideal solution, the modeling errors and the time-variance of real-world acoustic echo paths prevent a perfect cancellation. To achieve sufficient echo suppression the acoustic echo canceler therefore needs to be complemented by an adaptive echo suppression postfilter. Based on the stochastic echo path model proposed by Enzner, we derive in this contribution an MMSE solution for echo canceler and postfilter jointly. While the resulting postfilter bears some similarity with the traditional Wiener filter, it assumes a deterministic far-end signal and employs the undermodeling error and the uncertainty of the acoustic echo path in its gain computation. It thus compensates the typical deficiencies of acoustic echo cancelers in real-world applications. Another implication of the presented theory lies in the deep justification of a recent Kalman filtering trend in AEC, which delivers an echo-path estimate in conjuction with the echo-path uncertainty in the form of a posterior distribution.