Speex dev - Feb 2010 - Fwd: Re: Fixed Point on wideband-mode: Single Frame loss on 2000 Hz sine causes "freak off"

Jean-Marc Valin <Jean-Marc.Valin at USherbrooke.ca> hat am 1. Februar 2010
um 17:29 geschrieben:
> Can you look at the residual signal? i.e. what it looks like after the
pitch is
> removed. Any significant difference there? Also, the pitch gain can be
> informative.
>
I catched the "orignal" residual (st->exc from line 400 inside
nb_celp.c) plus the "synthesized" one (exc on line 880 inside
nb_celp.c) both for fixed and floating point.

There is a significant difference in the residual signal -- the fixed point
signal shows some kind of "glitches" (see attachments) while the
floating point version looks more or less smooth. But it is interesting that the
huge overdrive you can observe in the re-synthesized output signal isn't
present in the synthesized residual. This leads to an instable LPC synthesis
filter, or am I wrong?

cheers,
Frank
_________________________________________
NEU: Mit WEB.DE DSL ?ber 1000,- ? sparen!
http://produkte.web.de/go/02/

-------------- next part --------------
A non-text attachment was scrubbed...
Name: residual_fixed.png
Type: application/octet-stream
Size: 36386 bytes
Desc: not available
Url :
http://lists.xiph.org/pipermail/speex-dev/attachments/20100202/923c65cf/attachment-0005.obj
-------------- next part --------------
A non-text attachment was scrubbed...
Name: residual_fixed_detail.png
Type: application/octet-stream
Size: 113846 bytes
Desc: not available
Url :
http://lists.xiph.org/pipermail/speex-dev/attachments/20100202/923c65cf/attachment-0006.obj
-------------- next part --------------
A non-text attachment was scrubbed...
Name: syn_residual_fixed.png
Type: application/octet-stream
Size: 65383 bytes
Desc: not available
Url :
http://lists.xiph.org/pipermail/speex-dev/attachments/20100202/923c65cf/attachment-0007.obj
-------------- next part --------------
A non-text attachment was scrubbed...
Name: residual_float.png
Type: application/octet-stream
Size: 38973 bytes
Desc: not available
Url :
http://lists.xiph.org/pipermail/speex-dev/attachments/20100202/923c65cf/attachment-0008.obj
-------------- next part --------------
A non-text attachment was scrubbed...
Name: syn_residual_float.png
Type: application/octet-stream
Size: 58492 bytes
Desc: not available
Url :
http://lists.xiph.org/pipermail/speex-dev/attachments/20100202/923c65cf/attachment-0009.obj

Just had a look and the fixed-point residual *does* look odd. That's worth
investigating (though I don't think it's related to LPC stability).

    Jean-Marc

Quoting Frank Lorenz <Frank_wtal at web.de>:
> Jean-Marc Valin <Jean-Marc.Valin at USherbrooke.ca> hat am 1. Februar
2010 um
> 17:29 geschrieben:
>
> > Can you look at the residual signal? i.e. what it looks like after the
> pitch is
> > removed. Any significant difference there? Also, the pitch gain can be
> > informative.
> >
>
> I catched the "orignal" residual (st->exc from line 400 inside
nb_celp.c)
> plus the "synthesized" one (exc on line 880 inside nb_celp.c)
both for fixed
> and floating point.
>
> There is a significant difference in the residual signal -- the fixed point
> signal shows some kind of "glitches" (see attachments) while the
floating
> point version looks more or less smooth. But it is interesting that the
huge
> overdrive you can observe in the re-synthesized output signal isn't
present
> in the synthesized residual. This leads to an instable LPC synthesis
filter,
> or am I wrong?
>
> cheers,
> Frank
> _________________________________________
> NEU: Mit WEB.DE DSL ?ber 1000,- ? sparen!
> http://produkte.web.de/go/02/
>
>