Segmentation Fault when running PocketSphinx_batch with -mllr option

[Haim Zysberg]

Hello,

I have followed http://cmusphinx.sourceforge.net/wiki/tutorialadapt for adapting the default PocketSphinx acoustic model using MLLR adaptation (I do understand that MLLR adaptation is not very useful for semi-continuous models..).

I used the hub4wsj_sc_8k acoustic model distributed with PocketSphinx and the weather.lm language model distributed with the PocketSphinxAndroid Demo. For adaptation I used the full model referred-to in the tutorialadapt tutorial.

None of sphinx_fe, bw, and mllr_solve had errors.

When invoking PocketSphinx_batch with the –mllr option, I got a segmentation fault (without this option it worked fine). pocketsphinx_batch.exe.stackdump has the following stack trace info:

Exception: STATUS_ACCESS_VIOLATION at eip=00418F57
eax=3EAC6BD3 ebx=80072FF0 ecx=0000000D edx=FFFFFFFF esi=00000000 edi=80D246E0
ebp=800730C8 esp=0028A520 program=C:\CMUSphinx-11983\CMUSphinx\pocketsphinx\src\programs\.libs\pocketsphinx_batch.exe, pid 21932, thread main
cs=0023 ds=002B es=002B fs=0053 gs=002B ss=002B
Stack trace:
Frame     Function  Args
End of stack trace

The last line in the generated log file is:

INFO: cmn.c(175): CMN: 61.96  0.73  0.43  1.01 -0.76 -2.35 -2.43  0.03  0.86 -0.94 -0.23  0.34 -0.50

Outputs of sphinx_fe, bw, and mllr_solve as well as the PocketSphinx_batch run output and log file, and the stackdump file are located in https://www.dropbox.com/sh/lss0ggdh0970wjv/yCJK8Y4vq3

Any idea what is the problem?

Thanks,
Haim

EDIT: I downloaded the latest version (tagged with 12117), and now running PocketSphinx_batch with mllr_matrix (using the hub4wsj_sc_8k acoustic model included in this version) completes with no problems.

I tried also using this version with the "full" hub4wsj_sc_8k model, referred to in http://cmusphinx.sourceforge.net/wiki/tutorialadapt . I understand both models are supposed to be the same, but with the "full" model I still get the same segmentation fault.

Haim

[Nickolay V. Shmyrev]

So can you please share the matrix?

[Haim Zysberg]

Hi Nickolay,

Well, I tried to run again, and I don't get the exception anymore.. I'm not sure what was changed.. Sorry.

Regards,
Haim

BTW, I do get somewhat different accuracy results in comparison with what I get when decoding with the model included in the 12117 version.

[Nickolay V. Shmyrev]

Hi

Ok, great, let me know if there are futher issues. There was a known issue with MLLR related to locale, i'm not sure if you ever changed that.

Have a nice holidays

[Cubeler]

I'm using the same acoustic model and experience the same problem.
Here are the output of sphinx_fe, bw and mllr_solve:
http://nopaste.info/3554ce9871.html

The mllr_matrix:
http://nopaste.info/8ef563de80.html

And here is the log of the pocketsphinx call and the stacktrace:
http://nopaste.info/96c55bcff7.html

The segfault occurs in line 861 in s2_semi_mgau.c :
mgau_dist(s, frame, i, featbuf[i]);

Without the mllr switch everything runs just fine.

[Cubeler]

I updated to the svn version in the trunk but still no luck.
I really don't know what else to do.

[Nickolay V. Shmyrev]

This issue has been fixed in trunk today. Please update, thanks for the report!

[Nickolay V. Shmyrev]

Hello Cubeler

MLLR transform doesn't make much sense for semi-continuous model like hub4wsj_sc_8k (this issue is covered in tutorial) and it's not fully supported by pocketsphinx. We'll probably fix crash some day in the the future but overall I suggest you to use MAP instead of MLLR.