PocketSphinx decoder init failed

[Firoj Alam]

Dear all,
I think this is a known problem albeit I am not getting a good solution. One
of the solutions asked to set the -dictcase "yes" in arguments. I tried to set
it in etc/feat.params but there is no result. Following is my output in shell
and log file link is: http://faculty.bracu.ac.bd/~firoj/asr/crblp_asr-1-1.log

**shell output****
crblp@crblp-firoj:~/ASR/tutorial/crblp_asr$ perl scripts_pl/decode/slave.pl
MODULE: DECODE Decoding using models previously trained
Decoding 2 segments starting at 0 (part 1 of 1)
0% FATAL_ERROR: "batch.c", line 430: *PocketSphinx decoder init failed

This step had 4 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Failed to start /home/crblp/ASR/tutorial/crblp_asr/bin/pocketsphinx_batch
Aligning results to find error rate
Can't open /home/crblp/ASR/tutorial/crblp_asr/result/crblp_asr-1-1.match
word_align.pl failed with error code 65280 at scripts_pl/decode/slave.pl line
173.
crblp@crblp-firoj:~/ASR/tutorial/crblp_asr$

Thanks
Firoj

[Nickolay V. Shmyrev]

One of the solutions asked to set the -dictcase "yes" in arguments. I tried
to set it in etc/feat.params but there is no result. Following is my output in
shell and log file link is: http://faculty.bracu.ac.bd/~firoj/asr/crblp_asr-1
-1.log

Did you try to read the log? It clearly says that dictcase is still no. You
shouldn't add dictcase argument to feat.params, that's senseless. You need to
add it to scripts_pl/decode/psdecode.pl script this way:

-dictcase => "yes",

[Firoj Alam]

Thanks nsh.
Now i facing word alignment problem in some input files when decoding. I'm not
sure whether the word alignment problem is the cause of the following errors
in log file .

1. ERROR: "cmd_ln.c", line 840: Unknown argument: -logfn
2. ERROR: "s2_semi_mgau.c", line 1015: /home/crblp/ASR/tutorial/crblp_asr/model_parameters/crblp_asr.cd_cont_1000/means: #codebooks (1114) != 1
3. WARNING: "ngram_search.c", line 1000: not found in last frame, using আছে, instead
INFO: ngram_search.c(1046): lattice start node .0 end node আছে,.1831
ERROR: "ngram_search.c", line 1116: End node of lattice isolated; unreachable
4. WARNING: "ngram_search.c", line 1000: not found in last frame, using যা instead
INFO: ngram_search.c(1046): lattice start node .0 end node যা.1432
ERROR: "ngram_search.c", line 1116: End node of lattice isolated; unreachable

Several solutions shows several indicator for word alignment problem, which
I'm not getting the any relation with my one. Would be helpful if I get any
reference.

***shell output***
crblp@crblp-firoj:~/ASR/tutorial/crblp_asr$ perl
scripts_pl/decode/slave.plMODULE: DECODE Decoding using models previously
trained
Decoding 2 segments starting at 0 (part 1 of 1)
0%
This step had 4 ERROR messages and 5 WARNING messages. Please check the log
file for details.
Aligning results to find error rate
*word_align.pl failed with error code 65280 at scripts_pl/decode/slave.pl line 173.
crblp@crblp-firoj:~/ASR/tutorial/crblp_asr$

***end******

Thanks in advance.
Firoj

[Nickolay V. Shmyrev]

You need to provide the files required to run the decoding in order to help us
to reproduce your problem

[Firoj Alam]

This is the link of all the files. http://faculty.bracu.ac.bd/~firoj/asr/crbl
p_asr.tar.gz

Thank you

[Firoj Alam]

sphinx package i'm using.
sphinxtrain v1.0
pocketsphinx v0.5
sphinxbase

LM created from CMUlmtk: http://www.speech.cs.cmu.edu/tools/lmtool-
new.html
Dictionary manually created.

[Nickolay V. Shmyrev]

Your errors are numerous I'm not sure I even can list them all

1. Audio files are 44 kHz instead of 16 kHz
2. Phone file has no SIL phone
3. Dictionary file has tabs instead of spaces
4. Transcription file has double spaces instead of single spaces
5. You changed number of states without a reason
6. You changed the feature type without a reason

[Firoj Alam]

I have fixed the first four issues. I had to change the (5) # of states and
(6) feature type due to the following reason:

Can not create models used by Sphinx-II.
If you intend to create models to use with Sphinx-II models, please rerun with: $ST::CFG_HMM_TYPE = '.semi.' or $ST::CFG_HMM_TYPE = '.cont' and $ST::CFG_FEATURE = '1s_12c_12d_3p_12dd' and $ST::CFG_STATESPERHMM = '5'

But my problem has not solved yet rather error rate increases though there may
many reasons of error other than these fixes.

Thanks

[Firoj Alam]

Last updated package:
http://faculty.bracu.ac.bd/~firoj/asr/crblp_asr_v2.tar.gz

[Nickolay V. Shmyrev]

. I had to change the (5) # of states and (6) feature type due to the
following reason:

You didn't read this message precisely, you shouldn't do that. This is only
for sphinx2 which is obsolete. Moreover, you have old SphinxTrain, latest
sphinxtrain will not suggest you such a thing. So please revert your changes.

Ok, the next set of mistakes:

1. After you changed sample rate you didn't reextract the features with make_feats.pl

2. Your fileids file doesn't match utterance ids in transcription. Log file in logdir clearly says you that. In
   your fileids file utt 10001 is missing and in transcription file utt 10000 has
   some garbage in utterance id.

3. For such ridiculously small amount of training data you are using too many senones/gaussians. I suggest
   you 100 senones instead of 1000 and 1 gaussian instead of 8

After all fixed you should get this result

SENTENCE ERROR: 50.0% (1/2) WORD ERROR RATE: 4.3% (1/23)