Speex dev - Oct 2005 - Noisy sound quality with Blackfin in WB-mode

Hello all,

I'm testing the Speex codec for my diploma thesis on a BF-533 Blackfin under
uCLinux (2005R3 RC3 release). I successfully compiled the Speex (1.1.11-svn)
and I can encode/decode wav-files on my STAMP-board using the speexenc/speexdec
sample apps. But I encountered that the decoded file sounds strange/noisy, when
compiling with "--enable-blackfin-asm" +
"--enable-fixed-point", while it was
comparable to a PC/x86/Linux encoded/decoded file, when Speex was compiled with
"--enable-fixed-point" only. The strange sound seems to occur only in
16 kHz
WB-mode, not in NB-mode (I didn't test 32 kHz).

I encoded using for WB-mode:
# speexenc -w --quality 5 --comp 5 input-wav-16khz outputA.speex
and for NB-mode
# speexenc --quality 5 --comp 5 input-wav-8kHz outputB.speex

I decoded with
# speexdec outputA.speex
# speexdec outputB.speex

The decoded files can be downloaded at (156k each):

http://www.stud.tu-ilmenau.de/~bege-mt/maennl1-16000-16-mono-speex-c5-q5-noblackfinopt-decoded.wav

http://www.stud.tu-ilmenau.de/~bege-mt/maennl1-16000-16-mono-speex-c5-q5-blackfinopt-decoded.wav

(Text is german, hope that doesn't matter...)

Am I doing something wrong? Or is there a bug in the blackfin assembler code?

Thanks for answers,


Bernhard Gerlach
Technical University Ilmenau
Germany





----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.

Hi Bernhard,

Can you confirm I'm understanding everything correctly? You encode with
the same encoder and then decode with either A) blackfin assembly and
fixed-point or B) fixed-point only on Blackfin. Then A) sounds bad and
B) sounds good. If you do the same in narrowband, it sounds OK. Is that
correct? If that's the case, it's *probably* some kind of bug and/or
invalid assumption in the blackfin assembly. Could you check if that
happens on all wideband modes (and none of the narrowband modes)? Also,
testing with 1.1.10 would help (in case I broke something in svn).

	Jean-Marc

Le lundi 24 octobre 2005 ? 11:22 +0200,
bernhard.gerlach@stud.tu-ilmenau.de a ?crit :
> Hello all,
> 
> I'm testing the Speex codec for my diploma thesis on a BF-533 Blackfin
under
> uCLinux (2005R3 RC3 release). I successfully compiled the Speex
(1.1.11-svn)
> and I can encode/decode wav-files on my STAMP-board using the
speexenc/speexdec
> sample apps. But I encountered that the decoded file sounds strange/noisy,
when
> compiling with "--enable-blackfin-asm" +
"--enable-fixed-point", while it was
> comparable to a PC/x86/Linux encoded/decoded file, when Speex was compiled
with
> "--enable-fixed-point" only. The strange sound seems to occur
only in 16 kHz
> WB-mode, not in NB-mode (I didn't test 32 kHz).
> 
> I encoded using for WB-mode:
> # speexenc -w --quality 5 --comp 5 input-wav-16khz outputA.speex
> and for NB-mode
> # speexenc --quality 5 --comp 5 input-wav-8kHz outputB.speex
> 
> I decoded with
> # speexdec outputA.speex
> # speexdec outputB.speex
> 
> The decoded files can be downloaded at (156k each):
> 
>
http://www.stud.tu-ilmenau.de/~bege-mt/maennl1-16000-16-mono-speex-c5-q5-noblackfinopt-decoded.wav
> 
>
http://www.stud.tu-ilmenau.de/~bege-mt/maennl1-16000-16-mono-speex-c5-q5-blackfinopt-decoded.wav
> 
> (Text is german, hope that doesn't matter...)
> 
> Am I doing something wrong? Or is there a bug in the blackfin assembler
code?
> 
> Thanks for answers,
> 
> 
> Bernhard Gerlach
> Technical University Ilmenau
> Germany
> 
> 
> 
> 
> 
> ----------------------------------------------------------------
> This message was sent using IMP, the Internet Messaging Program.
> _______________________________________________
> Speex-dev mailing list
> Speex-dev@xiph.org
> http://lists.xiph.org/mailman/listinfo/speex-dev
>

Hi Jean-Marc,
> Can you confirm I'm understanding everything correctly? You encode  
> with
> the same encoder and then decode with either A) blackfin assembly and
> fixed-point or B) fixed-point only on Blackfin. Then A) sounds bad and
> B) sounds good. If you do the same in narrowband, it sounds OK. Is  
> that
> correct? If that's the case, it's *probably* some kind of bug
and/or
Yes, that's right, except one thing: I always use the blackfin 
+fixedpoint optimized encoder *and* decoder or the fixedpoint-only  
optimized encoder *and* decoder at one time.  And I do every encoding/ 
decoding on the Blackfin.

> invalid assumption in the blackfin assembly. Could you check if that
> happens on all wideband modes (and none of the narrowband modes)?  
> Also,
> testing with 1.1.10 would help (in case I broke something in svn).
I rechecked the issue with the following combinations with the 1.1.11- 
svn version:
NB-Mode (8kHz):
Complexity = 1, Quality = 1, 5, 9
Complexity = 5, Quality = 1, 5, 9
Complexity = 9, Quality = 1, 5, 9

I also checked the above combinations for the WB (16 KHz) and UWB- 
Mode (32kHz). And I checked everything with 1.1.10 version, too, as  
you suggested.  I checked by encoding+decoding a test speech sample  
for each mode with fixed+blackfin optimizations and by comparing the  
resulting file to the fixed-only-optimized encoded+decoded version of  
the speech sample.

Results: The strange sound persists in all WB and UWB combinations,  
no matter if its the 1.1.11-svn or the 1.1.10 version. No difference  
between blackfin+fixed and fixed-only in NB modes. The sound is  
getting even worse when comparing WB to UWB modes. Also sound seems  
to get worse with raising quality values.


Hope that helps,

Bernhard Gerlach
Technical University Ilmenau
Germany