Problem with lattice-align-words

[dakl...@gmail.com]

Hi,

 I'm currently working on a french speech recognizer using LIUM models and the ESTER 2 corpus (broadcast data) but I have a problem with lattice-align-words.
I used TEDLIUM recipe for my first experiment and everything was ok. From this point, I wanted to try nnet3 and some improvements (rescoring, mbr decoding, different alignments, etc) so I went to wsj recipe since this seems to be the last updated recipe.

The problem is I'm stuck during the word-lattices alignment when using word_align_lattices.sh. The command I used is :

    sil_label=`grep '!SIL' data/lang_nosp_test_tgpr/words.txt | awk '{print $2}'`

    steps/word_align_lattices.sh  --cmd "$train_cmd"  --silence-label $sil_label  data/lang_nosp_test_tgpr   exp/tri1/decode_nosp_dev_tgpr  exp/tri1/decode_nosp_dev_tgpr_ali || exit 1;

which ends up with an error :

    run.pl: 3 / 8 failed, log is in exp/tri1/decode_nosp_dev_tgpr_ali/log/align.*.log

When looking in the log files we can see that I have several invalid arcs :

# lattice-align-words --silence-label=1 --test=true data/lang_nosp_test_tgpr/phones/word_boundary.int exp/tri1/final.mdl "ark:gunzip -c exp/tri1/decode_nosp_dev_tgpr/lat.3.gz|" "ark,t:|gzip -c >exp/tri1/decode_nosp_dev_tgpr_ali/lat.3.gz"
# Started at Wed Apr  6 08:19:42 CEST 2016
#
lattice-align-words --silence-label=1 --test=true data/lang_nosp_test_tgpr/phones/word_boundary.int exp/tri1/final.mdl 'ark:gunzip -c exp/tri1/decode_nosp_dev_tgpr/lat.3.gz|' 'ark,t:|gzip -c >exp/tri1/decode_nosp_dev_tgpr_ali/lat.3.gz'
ERROR (lattice-align-words:TestArc():word-align-lattice.cc:749) Invalid arc in aligned CompactLattice: 0 0 1 7.24455,1890.09,2_1_1_1_1_1_1_1_1_1_1_1_1_1_8_5_5_5_5_5_18
WARNING (lattice-align-words:Close():kaldi-io.cc:496) Pipe gunzip -c exp/tri1/decode_nosp_dev_tgpr/lat.3.gz| had nonzero return status 13
ERROR (lattice-align-words:TestArc():word-align-lattice.cc:749) Invalid arc in aligned CompactLattice: 0 0 1 7.24455,1890.09,2_1_1_1_1_1_1_1_1_1_1_1_1_1_8_5_5_5_5_5_18

[stack trace: ]
kaldi::KaldiGetStackTrace()
kaldi::KaldiErrorMessage::~KaldiErrorMessage()
kaldi::WordAlignedLatticeTester::TestArc(fst::ArcTpl<fst::CompactLatticeWeightTpl<fst::LatticeWeightTpl<float>, int> > const&)
kaldi::WordAlignedLatticeTester::Test()
kaldi::TestWordAlignedLattice(fst::VectorFst<fst::ArcTpl<fst::CompactLatticeWeightTpl<fst::LatticeWeightTpl<float>, int> > > const&, kaldi::TransitionModel const&, kaldi::WordBoundaryInfo const&, fst::VectorFst<fst::ArcTpl<fst::CompactLatticeWeightTpl<fst::LatticeWeightTpl<float>, int> > > const&)
lattice-align-words(main+0x51f) [0x67ca45]
/lib64/libc.so.6(__libc_start_main+0xf0) [0x7f2269d79700]
lattice-align-words(_start+0x29) [0x67c459]

# Accounting: time=23 threads=1
# Ended (code 255) at Wed Apr  6 08:20:05 CEST 2016, elapsed time 23 seconds

I understand the problem but to be honest I can't figure out where it went wrong nor how to fix this since all my previous stages (in order : data preparation, features extraction, subset creation, train mono + tri1, rescoring w/ non-pruned trigram) didn't have any errors in log.
Do you have any ideas how to solve this problem?

Thank you in advance. Best regards,

Florian B.

[Daniel Povey]

You seem to be running it with slightly different-than-normal options, e.g. by adding silence.  I suspect it is a bug, triggered by options that haven't previously been used.

If you send me the files that will allow me to reproduce one of these failures, I can try to fix it.  I can't figure out just from looking at the code, what the problem might be.

Dan

--
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Florian BOYER]

Thank you for your answer Dan! You're right, the problem was with the silence addition.

The command I wrote in my previous message is directly taken from the wsj recipe. The only difference in the previous steps between my and their recipe is that I didn't compute the mono training alignments with --boost-silence. Could this be the cause of the errors?

Thanks again. Best regards,

Florian B.

[Daniel Povey]

The --boost-silence wouldn't have been the cause of the errors.  It's likely a bug in the program- please send me the files (model and alignments, and lang/ directory) so I can reproduce it.  You can do it off the list.

Dan

[Daniel Povey]

This issue is fixed now.
Dan