[speex-dev] Frozen upper spectrum in WB VBR CNG

[Jean-Marc Valin]

Yep, you found a bug and here's the patch. It should solve your problem
(using VBR without DTX), but there many still be some tuning do to on
the DTX.

        Jean-Marc

<p>diff -u -r1.118 sb_celp.c
--- sb_celp.c   18 Mar 2003 06:13:30 -0000      1.118
+++ sb_celp.c   4 May 2003 04:58:16 -0000
@@ -351,6 +351,8 @@
          int modeid;
          modeid = mode->nb_modes-1;
          st->relative_quality+=1.0*(ratio+2);
+        if (st->relative_quality<-1)
+            st->relative_quality=-1;
          while (modeid)
          {
             int v1;
@@ -361,7 +363,7 @@
             else
                thresh = (st->vbr_quality-v1)   *
mode->vbr_thresh[modeid][v1+1] +
                         (1+v1-st->vbr_quality) *
mode->vbr_thresh[modeid][v1];
-            if (st->relative_quality > thresh)
+            if (st->relative_quality >= thresh)
                break;
             modeid--;
          }

<p><p>Le sam 03/05/2003 à 15:48, Tom Grandgent a écrit :
> Jean-Marc Valin (jean-marc.valin at hermes.usherb.ca) wrote:
> > 
> > > I've been using Speex in my voice-over-IP program on Win32, in 
> > > wideband (16kHz) mode.  I just starting using VBR recently and 
> > > have run into something that might be a problem within Speex:
> > 
> > Are you turning on DTX in addition to VBR? Also, what version are you
> > using. As of 1.0, DTX is no longer implied by VBR.
> 
> I'm using 1.0.  DTX is definitely turned off.  (I tried turning it on 
> and saw the bitrate go even lower, but the sound still sounds the 
> same as when DTX is off.)
> 
> > > If someone hasn't spoken for a little while, and the bitrate drops 
> > > to very low, sometimes the high half of the spectrum becomes frozen 
> > > with a looping sound.  The bottom half of the spectrum is always 
> > > ok.  (I have a linear spectral analysis view that makes it very easy 
> > > to see this behavior.)  The frozen sound is usually much louder than 
> > > the background noise and sounds like a strange buzzing.  It seems 
> > > like it could be a frame from recent speech activity.  Occasionally 
> > > (like every few seconds) the sound will change to something 
> > > different, but still frozen.  As soon as the speaker starts speaking 
> > > again, the problem goes away and doesn't come back until another 
> > > period of silence.
> > 
> > The only way the spectrum can be "frozen" is when DTX is on and the VAD
> > detects no speech. The idea is to reproduce the noise without
> > transmitting any data. Can you send me some samples (original and .spx)
> > so I can check what's happening?
> 
> Ok.  I reproduced the problem using speexenc and speexdec (the Win32 
> binaries posted at http://www.speex.org/download.html) and have put 
> my sample input, the encoded .spx, and a .txt with commands used here:
> 
> http://www.grandgent.com/tom/temp/clickExample.zip  (220KB)
> 
> You will hear some background noise from the room and me clicking the 
> mouse a few times.  The first buzz is brought to a halt by the sound of 
> a soda can being opened.  Then there is a second buzz later on.  It 
> should be pretty noticible.
> 
> This is kind of a dumb example but I think it will demonstrate the 
> problem sufficiently.  If you want, I can make an example involving 
> conversation to prove that this problem is troublesome during typical 
> conversation...
> 
> > > Now, to describe the audio data I'm sending through the codec...  
> > > There's always some background noise but it's kept at reasonably 
> > > low levels, and is what you would expect to pick up with a mic in 
> > > a typical room with a computer in it.  I do a high-pass filter to 
> > > reduce bass below around 400Hz before encoding.  (Without the filter, 
> > > if there is significant bass background noise, VBR struggles and 
> > > the quality is very poor.)
> > 
> > Strange, the wideband more is designed for the 50-7000 kHz band. Sure if
> > you lower the bit-rate a lot, there may be problems with the bass
> > though.
> 
> I think the bass was just too strong.  My friend had his mic sitting 
> on his computer case... :)  At first I wondered why his voice sounded 
> so strange.  Then I had him turn on the filter on his end and it was 
> perfect after that.
> 
> > I think most VoIP applications so far have used CBR because it's more
> > predictable.
> 
> Yes, I was pretty happy with CBR but it is kind of wasteful to be 
> transmitting at full bitrate during silence.  So, I decided to at 
> least make VBR an option in my program, since Speex supports it.
> 
> Thanks!
> 
> Tom

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

<p>
-------------- 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/20030504/5a6262ea/signature-0001.pgp