Speex dev - Sep 2006 - Little question about the article "on Adjusting..."

Hello Jean-Marc Valin and all the speex-dev team,

I just read your article "On Adjsusting the learning Rate in Frequency
Domain Acho Cancellation With Double-Talk"
It is a very very interresting article, but I have a little questions
about the last section,

I understood that AEC with Gansler or cross-correlation algorithms have a
learning rate of 0.2 (in your experiment) or 0 if a double-talk is
detected.
But, in the figure 4, where you have plot the learning rate function of
the time, I can see that in the first 2 seconds, cross-correlation
algorithm's learning rate is about 0.22... How can it be possible ? Is it
just a measurement problem ?

Thank your for your answer,
Sincerly,

Jean-Christophe.

> I understood that AEC with Gansler or cross-correlation algorithms have a
> learning rate of 0.2 (in your experiment) or 0 if a double-talk is
> detected.
> But, in the figure 4, where you have plot the learning rate function of
> the time, I can see that in the first 2 seconds, cross-correlation
> algorithm's learning rate is about 0.22... How can it be possible ? Is
it
> just a measurement problem ?
I just checked with my code and the cross-correlation algorithm actually
used a learning rate of 0.25, not 0.2 (the Gansler algo was 0.2 though).
Thanks for pointing that out. I'll try to fix it before the article
shows up in IEEE Trans.

	Jean-Marc