[speex-dev] More Speex (pre-processing) Options

Jean-Marc Valin Jean-Marc.Valin at USherbrooke.ca
Wed Nov 26 15:26:02 PST 2003


> So I have had a chance to delve deeper into  the new pre-processing
> code-  I was also able to get VAD, AGC, Denoising, etc working inside of the
> current release candidate (1.0.3).   For some reason, the 1.1.2
> release didn't properly encode/decode in our system (on windows)
> and I haven't had time to determine why.  In any case, all of the features
> jointly work very well- great job!

Can you give some details about what's not working with 1.1.2 on Windows
(compile, run-time)?

> I realize the echo cancellation element may still be in development, but I am
> having trouble getting that part of the preprocessor to work.  When
> I don't pass NULL into the speex_preprocess for the echo parameter, things
> seem to go haywire- some memory appears to be getting trampled somewhere
> (this is probably my fault).  Maybe I am misunderstanding the purpose of this
> parameter?  Except for the first few packets, the preprocessor
> appears to be removing all audio- The algorithm I am using passes in a recently
> decoded frame...  Any ideas?  Do I need to set some settings in somewhere
> in order for this to work properly?

First thing, the echo-canceller is quite experimental. The "echo"
parameter to the preprocess function is meant to be the result obtained
from the last parameter of the echo-cancellation function. *However*,
this currently does not work properly, so do not use it. Currently, the
echo canceller works in some cases, but is not as robust as it should
be. There is some tuning to do for adaptation rate and the current
crosstalk detector needs to be improved (read rewritten).

Some other issues about the preprocessor... Normally, the denoiser
should be able to remove all (>30 dB) of the *stationary* noise, but
with non-stationary noise, your mileage may vary. As for the AGC, it's
very hard to tune for all possible applications. The current tuning
seems to work in most cases, but sometimes it will tend to amplify noise
(like echo residual) when there is no foreground speech.

        Jean-Marc


-- 
Jean-Marc Valin, M.Sc.A., ing. jr.
LABORIUS (http://www.gel.usherb.ca/laborius)
Université de Sherbrooke, Québec, Canada


-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 190 bytes
Desc: Ceci est une partie de message numériquement signée
Url : http://lists.xiph.org/pipermail/speex-dev/attachments/20031126/4e104d32/signature.pgp


More information about the Speex-dev mailing list