Recursive bayesian controllers for non-linear acoustic echo cancellation and suppression systems

Abstract

Both a cascade and a multichannel joint Bayesian estimator are provided for suppressing acoustic echo. An expansion basis (Power/Fourier series) is selected to convert a sample-based input signal xinto a DFT-domain multichannel signal [X, . . . X]. The posterior of unknown states (e.g., mean Ŵand covariance Pof the echo path Wand the mean âand covariance Qof the nonlinear coefficients a; or channel-wise mean Ŵand multichannel covarianceof a compound quantity formed by merging together the echo path Wand the inonlinear coefficient a) and model parameters θare estimated; and Kalman gain factor(s) Kare computed for optimal adaptation of the posterior of unknown states. An echo signal Ŷis estimated using the multichannel input signal [X, . . . X] and the adapted posterior; and an error signal Eis generated. Residual echo is suppressed by post-filtering the error signal Ewith a weighting function ψwhich depends on the adapted posterior, and the filtered error signal ŝ'is then transmitted to a far-end.