Colloquium - Details
You will receive information about presentations in time if you subscribe to the newsletter of the Colloquium Communications Technology.
Master-Vortrag: Online Learning of Loudspeaker Nonlinearities for Acoustic Echo Cancellation
Montag, 13. Juni 2022
Hands-free communication is pervasive throughout modern society, requiring a robust cancellation of the echo from the far end speech signal that is emitted from the loudspeaker to the microphone. Acoustic echo cancellation addresses this issue typically by employing linear adaptive filters. However, in reality, the echo path is nonlinear due to the non-ideal characteristics of acoustic transducers and power amplifiers operated at their physical limits.
This thesis introduces and investigates a novel approach to tackle nonlinear AEC by estimating the nonlinear reference signal using a deep neural network and a differentiable Kalman filter. The hybrid system is designed to learn loudspeaker nonlinearities directly from data, enabling end-to-end training on data composed of pairs of the far end reference and near end microphone signals. In contrast to previous neural network-based solutions that have been tailored toward one particular loudspeaker, the proposed system aims to be generalizable for different loudspeaker nonlinearities. Therefore, inspired by linear adaptive filtering, the recurrent architecture explicitly takes advantage of the information in the residual echo in order to estimate the nonlinearity adaptively.
The proposed approach was evaluated for both simulated and measured data. The results indicate that the architecture could enable faster convergence and better steady state performance than related adaptive approaches. Furthermore, some examples demonstrated that the performance of a method that makes use of oracle knowledge could be surpassed, evidently because the models adapt to the linear acoustic echo path, too.