Hi I have a question about the AEC and preprocessor. I have seen that when near end is talking for a longer time (about 10 s) without being interrupted by the far end the residual echo (or rather leak_estimate) increases making the preprocessor suppress the near end talk when it shouldn't. Is there a way to make the leakage estimate only update when far-end is present or similar in order to avoid this behavior? If I need to implement this myself, does it make sense to put some kind of far_end threshold on the update of leak_estimate or should I rather pause the adaption or do something else? Best Regards Johan -------------- next part -------------- An HTML attachment was scrubbed... URL: http://lists.xiph.org/pipermail/speex-dev/attachments/20090706/4de200e4/attachment.htm
Hi again I have now investigated further why the leak_estimate increases when near end is talking for a longer time. What happens is that the accumulated crosscorrelation of y (Pyy) in mdf.c is decreasing towards zero. The leak_estimate is calculated as the divistion of Pee with Pyy and when Pyy is close to zero the output will dramaticily increase. At least in my application there is need of handeling this behaviour. Have anyone else seen this behaviour in your application? One idea I have is to add a constant or floor to Pyy in the divisoin. Does anyone have any other ideas? Best Regards Johan -------------- next part -------------- An HTML attachment was scrubbed... URL: http://lists.xiph.org/pipermail/speex-dev/attachments/20090707/b14ae879/attachment.htm
Normally, leak_estimate shouldn't adapt when there's no echo, so maybe that's where the problem is. Jean-Marc Johan Nilsson a ?crit :> Hi > > > > I have a question about the AEC and preprocessor. > > > > I have seen that when near end is talking for a longer time (about 10 s) > without being interrupted > > by the far end the residual echo (or rather leak_estimate) increases > making the preprocessor > > suppress the near end talk when it shouldn?t. > > > > Is there a way to make the leakage estimate only update when far-end is > present or similar > > in order to avoid this behavior? > > > > If I need to implement this myself, does it make sense to put some kind > of far_end threshold > > on the update of leak_estimate or should I rather pause the adaption or > do something else? > > > > Best Regards > > Johan > > > ------------------------------------------------------------------------ > > _______________________________________________ > Speex-dev mailing list > Speex-dev at xiph.org > http://lists.xiph.org/mailman/listinfo/speex-dev
> Normally, leak_estimate shouldn't adapt when there's no echo, so maybe > that's where the problem is.I have made a some plots to illustrates what happens. You can find the plot on the following address: https://extranet.svep.se/leak_estimate_090709.PNG Best Regards Johan