Speex dev - Jan 2005 - Problem encoding sine wave in 1.1.6 and somewhat in 1.0.4

On Thu, 2005-01-13 at 12:42 -0500, Jean-Marc Valin
wrote:
> Le jeudi 13 janvier 2005 ? 10:59 -0500, Jared Whitby a ?crit :
> > Interestingly enough.. I started playing around with preprocessing
> > options in 1.1.6 and happened upon the denoise filter
> > (SPEEX_PREPROCESS_SET_DENOISE). When i run the test tone using that
> > option it is completely filtered out and I just get (complete)
> > silence. When the test tone is intermixed with regular voice I only
> > get the voice. So while i still don't quite understand why the
test
> > tone dorks up the encoder, it looks like I have a way around it.
> 
> Well, the denoiser is designed to remove any kind of stationary
> background noise. You sinusoid fits that definition pretty well :-)
> 
> 	Jean-Marc
The "Is is a bug? No, it's a feature" has just make another
victim!

Claude

Well I think a sinusoid shouldn't totally trash the encoder state, so
I still think theres a bug lurking around in there. The denoiser just
prevents it from ever making it to the encoder. Just from browsing
through the speex code and from what I've learned reading through the
mailing list, it looks like there are a few places where checks are in
place to prevent sinusoids from destabilizing the stream. Somehow this
particular sine wave is getting through the cracks I think. I actually
tried a couple of lower frequency samples and It didn't seem to cause
the problem I was having.

But, I seem to be in good shape with the filter in place... so just
figured I would share what i have learned.

Jared


On Thu, 13 Jan 2005 19:40:11 +0100, Claude Brisson <claude@renegat.net>
wrote:
> On Thu, 2005-01-13 at 12:42 -0500, Jean-Marc Valin wrote:
> > Le jeudi 13 janvier 2005 ? 10:59 -0500, Jared Whitby a ?crit :
> > > Interestingly enough.. I started playing around with
preprocessing
> > > options in 1.1.6 and happened upon the denoise filter
> > > (SPEEX_PREPROCESS_SET_DENOISE). When i run the test tone using
that
> > > option it is completely filtered out and I just get (complete)
> > > silence. When the test tone is intermixed with regular voice I
only
> > > get the voice. So while i still don't quite understand why
the test
> > > tone dorks up the encoder, it looks like I have a way around it.
> >
> > Well, the denoiser is designed to remove any kind of stationary
> > background noise. You sinusoid fits that definition pretty well :-)
> >
> >       Jean-Marc
> 
> The "Is is a bug? No, it's a feature" has just make another
victim!
> 
> Claude
> 
> 

-- 
Jared
jwhitby@gmail.com

Le jeudi 13 janvier 2005 ? 14:19 -0500, Jared Whitby a ?crit
:
> Well I think a sinusoid shouldn't totally trash the encoder state, so
> I still think theres a bug lurking around in there. The denoiser just
> prevents it from ever making it to the encoder. Just from browsing
> through the speex code and from what I've learned reading through the
> mailing list, it looks like there are a few places where checks are in
> place to prevent sinusoids from destabilizing the stream. Somehow this
> particular sine wave is getting through the cracks I think. I actually
> tried a couple of lower frequency samples and It didn't seem to cause
> the problem I was having.
I think the "problem" comes from the pitch predictor and the fact that
at certain frequencies, it will be able to perfectly predict the signal.
By doing so, its poles will be right on the unit circle and the impulse
response will not decay in time. Preventing that from happening without
hurting the general performance (where you *do* want the pitch predictor
to make the error as small as possible) is not trivial.

	Jean-Marc

-- 
Jean-Marc Valin <Jean-Marc.Valin@USherbrooke.ca>
Universit? de Sherbrooke

Hi Jared,

Just to let you know that this behavior got fixed in the 1.1.8 release.

	Jean-Marc

Le jeudi 13 janvier 2005 ? 14:19 -0500, Jared Whitby a ?crit
:
> Well I think a sinusoid shouldn't totally trash the encoder state, so
> I still think theres a bug lurking around in there. The denoiser just
> prevents it from ever making it to the encoder. Just from browsing
> through the speex code and from what I've learned reading through the
> mailing list, it looks like there are a few places where checks are in
> place to prevent sinusoids from destabilizing the stream. Somehow this
> particular sine wave is getting through the cracks I think. I actually
> tried a couple of lower frequency samples and It didn't seem to cause
> the problem I was having.
> 
> But, I seem to be in good shape with the filter in place... so just
> figured I would share what i have learned.
> 
> Jared
> 
> 
> On Thu, 13 Jan 2005 19:40:11 +0100, Claude Brisson
<claude@renegat.net> wrote:
> > On Thu, 2005-01-13 at 12:42 -0500, Jean-Marc Valin wrote:
> > > Le jeudi 13 janvier 2005 ? 10:59 -0500, Jared Whitby a ?crit :
> > > > Interestingly enough.. I started playing around with
preprocessing
> > > > options in 1.1.6 and happened upon the denoise filter
> > > > (SPEEX_PREPROCESS_SET_DENOISE). When i run the test tone
using that
> > > > option it is completely filtered out and I just get
(complete)
> > > > silence. When the test tone is intermixed with regular voice
I only
> > > > get the voice. So while i still don't quite understand
why the test
> > > > tone dorks up the encoder, it looks like I have a way around
it.
> > >
> > > Well, the denoiser is designed to remove any kind of stationary
> > > background noise. You sinusoid fits that definition pretty well
:-)
> > >
> > >       Jean-Marc
> > 
> > The "Is is a bug? No, it's a feature" has just make
another victim!
> > 
> > Claude
> > 
> > 
> 
> 
-- 
Jean-Marc Valin <Jean-Marc.Valin@USherbrooke.ca>
Universit? de Sherbrooke