Speex dev - Jan 2005 - different output for opt/non-opt on win32 platform

hi all,

thanks Jean for the codec. 

I am new to speech coding. Am interested to implement 
speex on a TI dsp, say C54. just downloaded speex v 1.1.6.
 when I removed optimisation flags for win32 in libspeex.dsp, 
I get a slightly different output for decoding using speexdec.exe 
(debug version). no prob for speexenc.exe except for some 
differences in the ogg header. 

for anyone interested, i have zipped some wave files in speexlib.zip
and posted on geocities.com/dunjie17/

- libspeex.dsp is the modified version.
- male.spx is the encoded version of male.wav
on the speex website, coded using bitrate 4000
- origdec_male.wav
- origdec_male.wav
- unoptmale.wav (the result after removing the
optimisation flags)

config of my system: winxp sp2, vs6 sp5, libogg1.1

from the archives, i read fixed-point is implemented.  wonder
whether there are any sample results that I could compare
with?

i have the understanding each speex frame is self-contained,
i.e., the decoder could decode each frame independently. is that
right?

thanks for the time.
cheers,
tk

Le mercredi 05 janvier 2005 ? 13:37 +0800, tunkeat a ?crit
:
> thanks Jean for the codec. 
My first name's Jean-Marc BTW :-)
> I am new to speech coding. Am interested to implement 
> speex on a TI dsp, say C54. 
I think I've heard someone was working on that before...
> just downloaded speex v 1.1.6.
>  when I removed optimisation flags for win32 in libspeex.dsp, 
> I get a slightly different output for decoding using speexdec.exe 
> (debug version). no prob for speexenc.exe except for some 
> differences in the ogg header. 
What difference in the ogg header? It's quite normal that compiler
options will (slightly) change the result. The reason is that the
rounding depends on when the compiler loads/stores the float values. 
> i have the understanding each speex frame is self-contained,
> i.e., the decoder could decode each frame independently. is that
> right?
No. Like any other CELP codec, Speex has a memory. The only non-PCM
speech codec I know that has independent frames is iLBC and it's
suboptimal for most uses because of that.

	Jean-Marc

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

----- Original Message ----- 
From: "Jean-Marc Valin" <Jean-Marc.Valin@USherbrooke.ca>
To: "tunkeat" <tunkeat@gmail.com>
Cc: "speex" <speex-dev@xiph.org>
Sent: Wednesday, January 05, 2005 2:48 PM
Subject: Re: [Speex-dev] different output for opt/non-opt on win32 platform
>No. Like any other CELP codec, Speex has a memory. The only non-PCM
>speech codec I know that has independent frames is iLBC and it's
>suboptimal for most uses because of that.
Hmmm... i was unaware of this... so how many frames/packets of preroll are 
required to ensure "perfect" (where perfect = as it would have been 
reproduced at this point had the file played from the beginning) output ?

ie basically what i'm asking is what impact does this have on seeking  or 
editing/cutting/recombining ? Because when you just seek to some random 
point... obviously it doesn't have the same memory state as if it had played
to there. I assume the difference is either very small or over such a short 
duration that the audio change at a seek essentially masks it ?

Zen.

hi,

Jean-Marc, thanks for the comments. 
> > just downloaded speex v 1.1.6.
> >  when I removed optimisation flags for win32 in libspeex.dsp,
> > I get a slightly different output for decoding using speexdec.exe
> > (debug version). no prob for speexenc.exe except for some
> > differences in the ogg header.
> 
> What difference in the ogg header? It's quite normal that compiler
> options will (slightly) change the result. The reason is that the
> rounding depends on when the compiler loads/stores the float values.
diff in ogg header: oh, due to a random number passed to the
second argument of ogg_stream_init().

without a good understanding of the celp and haven't delved into the 
codes, I am surprised that there is no such issue with the encoder.
whether the optimsation flags are included or excluded, the encoded
bits are the same. I assumed the decoder shared quite a number of 
functions as the encoder except maybe the perceptual enhancement part.
strange.
> 
> > i have the understanding each speex frame is self-contained,
> > i.e., the decoder could decode each frame independently. is that
> > right?
> 
> No. Like any other CELP codec, Speex has a memory. The only non-PCM
> speech codec I know that has independent frames is iLBC and it's
> suboptimal for most uses because of that.
> 
how big is this "Speex's memory"?  I am thinking perhaps naively:
at
the encoder side, is it possible to get the states for the decoder
from the encoder's memory without doing the actual decoding? if yes,
the decoder's states could be sent over.
I hope I am making sense here.

thanks in advance,
tk