errors with fixed point

[Pankaj]

With the new release of pocketsphinx 0.6.1 I am finding 100% errors on Windows
XP. For every utterance the error message is
ERROR: "fsg_search.c", line 1092: Final state not reached in frame 248.
When sphinxbase is compiled for floating point then everything works fine.

Pankaj

[Pankaj]

Am I missing something? When the new release of Pocketsphinx 0.6.1 is compiled
for fixed point support with VS2008, I am getting 100 % errors.

The error message is "ERROR: "fsg_search.c", line 1092: Final state not
reached in frame 248. ".

The decoding happens properly when pocketsphinx is compiled for floating
point. Ver 0.6 worked properly with both floating point as well as fixed
point.

Regards
Pankaj

[Nickolay V. Shmyrev]

It might be the bug. First of all you need to check the feature extraction
code (sphinx_fe) to return same values with fixed point as with floating.

[Pankaj]

Should I run sphinx_fe.exe or should I use the option -mfclogdir while running
pocketsphinx_batch.exe to compare the results in floating point and fixed
point mode.

Pankaj

[Nickolay V. Shmyrev]

Doesn't matter, you can try both. Which one you like :) It's only important
that feature values must be the same or at least look similar both in fixed
and floating points.

[Pankaj]

Hi

The comparison of mfc files shows vast difference between floating point and
fixed point results.
Fixed point mfc files contain a large no of zeroes. I have attached the links
for the mfc files.
The decoding is happening correctly in floating point mode.

http://www.mediafire.com/download.php?g66ylghbo519wi3
http://www.mediafire.com/download.php?1qbjk8wpcwq1t3f

Pankaj

[Pankaj]

Has anybody come across errors with fixed point implementation in the latest
release of pocketsphinx 0.6.1 on Windows XP.

My observation is that floating point version is working properly but fixed
point version is giving 100 % errors.
I am using hub4wsj_sc_8k. When the features generated are compared a vast
difference is seen between the results of fixed point mfcs and floating point
mfcs.

I have given the links for a sample audio and the corresponding mfc files.

http://www.mediafire.com/?j0s4858mkc73zec
http://www.mediafire.com/download.php?g66ylghbo519wi3
http://www.mediafire.com/download.php?1qbjk8wpcwq1t3f

Thanks in advance for the help

Pankaj

[Nickolay V. Shmyrev]

Sorry pankaj, I didn't have time to check the problem this week. I will look
soon.

[Nickolay V. Shmyrev]

Hello Pankaj

I've got one idea. It will take ages for me to reboot on windows and debug
this issue. Can you just checkout few versions from subversion in 0.6 branch
between 0.6.1 and 0.6 and test which change broke the fixed poiint? That will
give very good hint and allow to fix the issue.

[Pankaj]

Hi nsmhyrev,

Yes I will do that.

[Pankaj]

Hello nshmyrev,

I am wondering how to access the interim versions. Can they be checked out
through a web browser.

Pankaj

[Nickolay V. Shmyrev]

Yes, go here

http://cmusphinx.svn.sourceforge.net/viewvc/cmusphinx/branches/pocketsphinx-0
.6/

See "Download GNU tarball" link in the bottom

See revision history here:
http://cmusphinx.svn.sourceforge.net/viewvc/cmusphinx/branches/pocketsphinx-0
.6/?view=log

To checkout specific revision visit

http://cmusphinx.svn.sourceforge.net/viewvc/cmusphinx/branches/pocketsphinx-0
.6/?pathrev=<number>

For example

http://cmusphinx.svn.sourceforge.net/viewvc/cmusphinx/branches/pocketsphinx-0
.6/?pathrev=9984

[Pankaj]

Hi nshmyrev,

I dont know whether I am correct or not but my observation is as under:
When the rev 10344 (which seems to be the latest revision) is checked out,
then no errors are observed in fixed point implementation.
http://cmusphinx.svn.sourceforge.net/viewvc/cmusphinx/branches/pocketsphinx-0
.6/?pathrev=10344
The errors were seen only when the sources were downloaded from the release
i.e https://sourceforge.net/projects/cmusphinx/files/pocketsphinx/0.6.1
& https://sourceforge.net/projects/cmusphinx/files/sphinxbase/0.6.1

[Nickolay V. Shmyrev]

Sorry, but the sources in this version and released sources are identical,
aren't they? Can you check that release has this bug? Probably there was some
issue with build.

[Pankaj]

Hi nshmyrev,

The bug is found only when the release version is compiled. When the sources
obtained from SVN are compiled, fixed point error is not observed.

[Nickolay V. Shmyrev]

Hello Pankaj

Can you try to unpack the sphinxbase release 0.6.1 and remove the file
include/sphinx_config.h. I have an idea that this file conflicts
with include/win32/sphinx_config.h and because of that build is broken.

[Pankaj]

Hi Nshmyrev,

Yes, after removing the file include/sphinx_config.h the problem got removed.
Thank you very much.

Pankaj

[Nickolay V. Shmyrev]

Thanks, this issue is now fixed in trunk and in 0.6 branch

[Pablo]

hi nikolay,

when I load spanish acoustic model (H4.cd_continuous_8gau) and the dictionary
(H4.dict) with both comands: -hmm and -dict and the path, and I load jsgf too
like this:

#JSGF V1.0;

    grammar gram;
    public <gram> = [<rob>] <order> | [<rob>] <order> [<la>] <location> [<hacia> <la>] <side> <quantity>;

<rob> = ROBOT
;

<order> =  MUEVE
        | MUEVETE 
        | HOLA
        | SALUDOS
        | BUENAS
        | ADIOS
;
<la> = LA
;
<hacia> = HACIA
;
<location> = AVANZA
        | ATRAS
        | CABEZA
        | GIRANDO
;
<side> = IZQUIERDA 
        | DERECHA
        | ARRIBA
        | ABAJO
;

<quantity> = <digit>| <below30>| <tens> <y> [<digit>];

<y> = Y
;

<digit> = UNO 
        | DOS
        | TRES
        | CUATRO    
        | CINCO
        | SEIS
        | SIETE 
        | OCHO 
        | NUEVE
;
<below30> = DIEZ 
        | ONCE 
        | DOCE 
        | TRECE
        | CATORCE
        | QUINCE 
        | DIECISEIS 
        | DIECISIETE 
        | DIECIOCHO
        | DIECINUEVE
        | VEINTE
        | VEINTIUNO
        | VEINTIUN
        | VEINTIDOS
        | VEINTITRES
        | VEINTICUATRO 
        | VEINTICINCO
        | VEINTISEIS
        | VEINTISIETE
        | VEINTIOCHO
        | VEINTINUEVE
;
<tens> = TREINTA
        | CUARENTA
        | CINCUENTA
        | SESENTA
        | SETENTA
        | OCHENTA
        | NOVENTA
;

and load with this -jsgf and the path.

I run my program and I get that:

READY....
Listening...
Stopped listening, please wait...
INFO: cmn_prior.c(121): cmn_prior_update: from < 10.72 -0.07 -0.21 -0.10 -0.27 -0.11 -0.23 -0.07 -0.24 -0.15 -0.11 -0.12 -0.08 >
INFO: cmn_prior.c(139): cmn_prior_update: to   < 10.63 -0.16 -0.25 -0.11 -0.26 -0.10 -0.21 -0.06 -0.22 -0.13 -0.12 -0.10 -0.10 >
INFO: fsg_search.c(1030): 101 frames, 2124 HMMs (21/fr), 4575 senones (45/fr), 584 history entries (5/fr)

ERROR: "fsg_search.c", line 1099: Final state not reached in frame 101
000000000: (null)
READY...
Listening...
Stopped listening, please wait...
INFO: cmn_prior.c(121): cmn_prior_update: from < 10.36  0.24 -0.17 -0.12 -0.26 -0.10 -0.19 -0.16 -0.22 -0.06 -0.10 -0.17 -0.17 >
INFO: cmn_prior.c(139): cmn_prior_update: to   < 10.46  0.24 -0.19 -0.14 -0.27 -0.10 -0.19 -0.16 -0.23 -0.05 -0.10 -0.17 -0.17 >
INFO: fsg_search.c(1030): 103 frames, 3992 HMMs (38/fr), 8387 senones (81/fr), 1082 history entries (10/fr)

ERROR: "fsg_search.c", line 1099: Final state not reached in frame 103
000000001: (null)

I think that it is posible who was the same problem, I read all of this thread
but I can't understand anything

any idea that what must I do?

[Nickolay V. Shmyrev]

Please start a new thread. I don't think your problem is related. You see that
problem was solved already

Provide command line you are using
Provide infromation about pocketsphinx and sphinxbase versions
Provide full pocketsphinx output.
Make sure you properly configured frontend for spanish model (it uses -feat
s3_1x39 feature type)