[Speex-dev] white noise problem when using -DFIXED_POINT

[Jean-Marc Valin]

Please check without VAD and DTX. It hasn't been tested much with fixed-
point, so maybe that's the problem.

	Jean-Marc

On mer, 2004-09-15 at 07:41 -0700, wrote:
> It seems that with the file I posted, the decode noise persists all the way back to 1.1.1, when using --enable-fixed-point.
> 
> When we configure with --enable-fixed-point-debug, the decode seems to be correct, even on 1.1.6.
> 
> Hope this helps, and thanks for your time!
> 
> Regards,
> Owen
> 
> Jean-Marc Valin wrote:
> 
> > Hi Owen,
> > 
> > I have little time right now to look at the details, but it would help
> > if you could test with all 1.1.x versions and see when the problem
> > appeared. Also, compile with --enable-fixed-point-debug so messages will
> > be printed in case of overflows.
> > 
> > 	Jean-Marc
> > 
> > P.S. Please also post the original file.
> > 
> > On mar, 2004-09-14 at 22:42 -0700, wrote:
> > 
> >>Jean-Marc Valin wrote:
> >>
> >>>On mar, 2004-09-14 at 16:31 -0700, wrote:
> >>>
> >>>
> >>>>Hi there.  We're trying to build a speex decoder for Palm OS (i.e. ARM7).
> >>>>
> >>>>No major problems porting thus far, but when compiling the speex modules with -DFIXED_POINT all the silent parts of the output are replaced with what sounds like high-power white noise.  The spoken parts sound great, but all the silent parts in between speaking just sounds like white noise.
> >>>>
> >>>>We're using speex 1.1.6, and current libogg sources pulled down from subversion.
> >>>
> >>>
> >>>It's difficult to say with only that info. Have you tried the same file
> >>>on an x86 (with and without FIXED_POINT) and comparing the results. If
> >>>you still can't find the problem, please post the file somewhere (along
> >>>with the encoded version) and I'll have a look at it. Oh, and are you
> >>>using speexenc/speexdec or you own encoder and decoder?
> >>>
> >>>	Jean-Marc
> >>
> >>Jean-Marc,
> >>
> >>Thanks for the quick reply, and sorry for eliding some of the details.  Our test builds thus far have indeed been on x86.  I reproduced the white noise condition described above using unmodified speexdec from speex 1.1.6, configured with --enable-fixed-point.  The output sounds okay when configured without fixed point support.
> >>
> >>As I mentioned, we're using speex 1.1.6, current svn libogg, and gcc 2.95.4.
> >>
> >>Here is one of the files that seems to exhibit the problem:
> >>http://owen.theemrys.com/speex/0102.spx
> >>
> >>Thanks in advance for any information.  I've spent the last few hours combing the list archives, but nothing seems to stand out.
> >>
> >>Best regards,
> >>Owen
> > 
> > 
> > 
> > 
> > 
>