Master Theses - Details

Online Learning of Loudspeaker Nonlinearities for Acoustic Echo Cancellation

Supervisors: Till Hardenbicker, Stefan Kühl

Topics: Recurrent Neural Networks, Acoustic Echo Cancellation


With remote work becoming more and more common, hands-free communication via loudspeakers is indispensable. Specifically, it alleviates the need to wear headphones for extended periods of time. For the task of acoustic echo cancellation (AEC), adaptive algorithms like the normalized least mean square algorithm (NLMS) or the Kalman filter (KF) are well known and widely used in modern full-duplex communication systems.

While these algorithms are optimal for linear systems, they exhibit strong residual artifacts when confronted with nonlinearities in the echo path. Such behaviour is mostly caused by small and cheap loudspeakers. Furthermore, it is known that these nonlinearities have a temporal memory. Given that similar loudspeakers result in similar artifacts, data-driven approaches can exploit prior knowledge about the nonlinear behaviour.

In this thesis, a recurrent network structure for non linear AEC shall be investigated. It is based on a nonlinear expansion whose coefficients are updated online by a recurrent neural network (RNN). In order to generate training data for this model, measurements with a real device are required. These should contain different realistic acoustic scenarios. In the next step this data shall be used to train the nonlinear echo canceler. For the training procedure, optimal parameters and input features need to be determined. The best performing model shall be compared to related approaches.