WER 100% (monocode decoding)

[Muhammad Salem]

Hi ALL
I am getting a WER 100% when i run monophone decoding , i tried to go throught the log file
This is example of 1 log file out of 4 jobs
10-track1 SENT-START SENT-END
LOG (gmm-latgen-faster:DecodeUtteranceLatticeFaster():decoder-wrappers.cc:285) Log-like per frame for utterance 10-track1 is -7.62512 over 581 frames.
11-track1 SENT-START SENT-END
LOG (gmm-latgen-faster:DecodeUtteranceLatticeFaster():decoder-wrappers.cc:285) Log-like per frame for utterance 11-track1 is -7.80905 over 389 frames.
12-track1 SENT-START SENT-END
LOG (gmm-latgen-faster:DecodeUtteranceLatticeFaster():decoder-wrappers.cc:285) Log-like per frame for utterance 12-track1 is -7.87264 over 426 frames.
13-track1 SENT-START SENT-END
LOG (gmm-latgen-faster:DecodeUtteranceLatticeFaster():decoder-wrappers.cc:285) Log-like per frame for utterance 13-track1 is -8.08871 over 259 frames.
LOG (gmm-latgen-faster:main():gmm-latgen-faster.cc:175) Time taken 3.76832s: real-time factor assuming 100 frames/sec is 0.227693
LOG (gmm-latgen-faster:main():gmm-latgen-faster.cc:178) Done 4 utterances, failed for 0
LOG (gmm-latgen-faster:main():gmm-latgen-faster.cc:180) Overall log-likelihood per frame is -7.80461 over 1655 frames.

Do you have a clue what could be the problem in?

[Daniel Povey]

It's hard to tell what went wrong without looking at the training logs.  For example, what does the 'text' file in your training directory look like?  Does it have words in it?  And are those words in the words.txt in the lang directory? If not you'll get a bunch of warnings about OOVs in compile_graphs.*.log and/or align.*.log.
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.

[Muhammad Salem]

in compile_graphs i get following messages
sym2int.pl: replacing m_a with 1
sym2int.pl: replacing hA with 1
sym2int.pl: replacing m_i with 1
sym2int.pl: replacing . with 1
sym2int.pl: replacing m_a with 1
sym2int.pl: replacing hA with 1
sym2int.pl: not warning for OOVs any more times
** Replaced 13373 instances of OOVs with 1
I checked that those words are not in my  lexicon nor my words.txt the others are included

In align.*.log i get lots of warnings like

WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance 092011001006-110 with beam 40
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:475) Did not successfully decode file 092011001006-110, len = 568
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance 092011001006-203 with beam 40
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance 092011001006-243 with beam 40
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance 092011001006-350 with beam 40
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance 092011001006-454 with beam 40
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance 092011001006-514 with beam 40
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance 092011001006-561 with beam 40


I want to give you a quick picture of what i am doing may be there is a step am doing wrong
First: in training I got lexicon of (phone to phone lexicon), a text files its transcripts are by phone , I used those to do the monophone training and the language directory for training
Second: for testing i have a different lexicon (words to phone ) so i created another lang_test which based on this lexicon, The G.fst is a converted file from a standard lattice format i used in HTK
Third : in decoding i have 16 waves for testing each is transcripted by phones those i used to create the dircetory data_test

[Muhammad Salem]

FYI

Those Warning in align i have similar warnings in another Kaldi project and it gives reasonable WER so i guess its not the problem as far as i my small understanding

[Daniel Povey]

Make sure that the phones.txt files in the 'lang' and 'lang_test' directories are exactly the same (except possibly for one having extra disambiguation symbols).  They need to match.
Dan

On Tue, Nov 24, 2015 at 6:34 AM, Muhammad Salem <mhmd.sl...@gmail.com> wrote:

FYI

Those Warning in align i have similar warnings in another Kaldi project and it gives reasonable WER so i guess its not the problem as far as i my small understanding

--