Too many alignment errors in Librispeech recipe?

[Xavier Anguera]

Hi,

I am running the Librispeech recipe and I am getting reasonable WER results (by reasonable I mean they are quite close/equal to those in RESULTS file), therefore I suppose my setup is correct.

Looking at the logs closely I notice that for the alignment step tri4b_ali_clean_460 (i.e. right after pronunciation and silence probabilities and recreating the lang directory) there is an insane number of warnings indicating that a file could not be decoded/aligned successfully. For example:

align_pass1.10.log:WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance 4152-12926-0010 with beam 40

align_pass1.10.log:WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:475) Did not successfully decode file 4152-12926-0010, len = 1377

I tried raising the retry_beam to 60 but still get many of those.

I am wondering whether this is "normal" in the Librispeech recipe or else it must be something in my setup that I messed up (in the second case I would not see where, as WER's seem still good).

It is a shame to be throwing away so much data, so any help would be appreciated!

X.

[Daniel Povey]

It's hard to know from your email what proportion were not aligned
successfully (and therefore how problematic it is). The last line of
the log file should tell you.
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.

[Xavier Anguera]

About 30% can not be aligned.

[Daniel Povey]

I found an old directory of this, and it was only retrying a handful
(see below).
You might want to check for earlier errors, in other log files in that
directory.
And check that the phones.txt is not changed between 'lang_nosp' and 'lang'.

Dan


grep 'Retried'
/home/dpovey/kaldi-svn-clean/egs/librispeech/s5/exp/tri4b_ali_clean_460/log/align_pass?.1.log

b_ali_clean_460/log/align_pass?.1.log

/home/dpovey/kaldi-svn-clean/egs/librispeech/s5/exp/tri4b_ali_clean_460/log/align_pass1.1.log:LOG
(gmm-align-compiled:main():gmm-align-compiled.cc:180) Retried 12 out
of\

6511 utterances.

/home/dpovey/kaldi-svn-clean/egs/librispeech/s5/exp/tri4b_ali_clean_460/log/align_pass2.1.log:LOG
(gmm-align-compiled:main():gmm-align-compiled.cc:180) Retried 23 out
of\

6511 utterances.

On Thu, May 12, 2016 at 1:49 PM, Xavier Anguera <xav...@elsanow.io> wrote:
> About 30% can not be aligned.
>

[Xavier Anguera]

I see, thanks!

I will try to find what I broke.

You received this message because you are subscribed to a topic in the Google Groups "kaldi-help" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/kaldi-help/VcdrcSnJhTE/unsubscribe.
To unsubscribe from this group and all its topics, send an email to kaldi-help+...@googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

--

Xavier Anguera

CTO & CSO

ELSA Corp.

[Xavier Anguera]

Quick update, your hunch was totally accurate. I just realized that I recently changed the initial prepare_lang.sh call to add an external phones.txt but did not change the second call. This will definitely be the reason.

thank you very much!