Abstract: Approaches to non-linear acoustic echo cancellation for mobile devices all require a reliable, non-linear model of the loudspeaker. This paper investigates a recent approach to nonlinear system identification. The loudspeaker is characterised using specially-crafted test signals, in this case simple exponential sine-sweep inputs, which facilitate the derivation of a polynomial Hammerstein model. Model performance is then assessed using the same artificial test input signal and then using real-speech. Model performance is assessed objectively using the mean cepstral distance between real loudspeaker outputs and that estimated using the model. Results show the potential of the approach but also the challenge in estimating reliable non-linear models which accurately predict the response to complex real-speech inputs.

Download: