$G=$;NO@J8!&2]Bj8&5f(B $BEE;R%U%!%$%k$HF1$8>l=j$KCV$$$F2<$5$$!#$h$m$7$/$*4j$$$7$^$9!#(B ->
The classic acoustic echo canceller uses the adaptive system to remove the undesired echo signal. Since this is a one-channel simple problem, adaptive system can remove the echo successfully. If the system have a more complex multi-channel structure, the adaptive system will not work as successfully as expected. Non-uniqueness problem caused by mutual coupling can not been solved by a classic stereo acoustic echo canceller.
Input sliding method will be introduced to ensure the unique identification of the echo path impulse response. We will introduce two specific models using input sliding method to add time variation into input signals. The SAEC model with pre-processing in both channels makes the cross-correlation between input signals change faster, and accelerates the convergence rate of the weight error vector.
In the analysis part, we needed to estimate the theoretical convergence rate of 2PP model. We will starte from the geometric analysis of the converging process of weight error vector, which can give us a direct-vision understanding of the converging mechanism of input-sliding method. Based on the geometrical analysis, we can finally estimate the approximation value of the theoretical convergence rate of weight error vector. It is shown clearly that only the parameter Q and N will affect the converging speed.
In order to indicate how fast it will improve by using two pre-processing units instead of one, numerical simulation experiments has been done. Both experimental data and theoretical data of convergence rate have been taken, and a comparison between the two models using the same parameter and source signal has been done. Experiment results show that the 2PP model can accelerate the convergence speed 2 times faster, which agrees with our convergence analysis.