Beginner's question - "spk2utt and utt2spkk do not seem to match"

[Rolando Coto]

Hi!

I'm starting to use Kaldi and I want to train it on my own data from a Central American Indigenous language. Because I'm just starting, I recorded myself saying a few words in the language and try to train Kaldi on that. But I am getting a few errors, the first of which is:

utils/validate_data_dir.sh: spk2utt and utt2spk do not seem to match

When I created spk2utt, I just transposed utt2spk and pretty much pasted it in there, so it should match. Does anyone have any suggestions? Thank you so much!

[Daniel Povey]

I guess you must have made a mistake converting one to the other, see
utt2spk_to_spk2utt.pl and spk2utt_to_utt2spk.pl for how to convert one
into another without the possibility of error, but as the error
messages point out, it's not a good idea to put all the utterances as
the same speaker-- make it a separate speaker per utterance.

> --
> Go to http://kaldi-asr.org/forums.html find out how to join
> ---
> 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.

[Rolando Coto]

Hi, Dan!

Thank you so much for your quick reply! I really appreciate it.

Unfortunately, utt2spk_to_spk2utt.pl are producing the exact same output I already have. I already tried copy-pasting their output into the files just to make sure it was working well, but run.sh gave me the same error. (I'm attaching a screen capture of the output of utt2spk_to_spkutt.pl).

(As for the single speaker problem, I will for sure give it more input data once I have at least one working demonstration model. Wouldn't the problem be independent from having just one speaker?)

Thanks!

[Daniel Povey]

Well I can see just from glancing at it that it's not right-- there is
a missing underscore (it's a space) 3 characters from the end. I
think you need to learn the basics of UNIX scripting. You may be
relying too much on copy-pasting and viewing and too little on UNIX
concepts like redirection, diff, cmp, things like that.

[Rolando Coto]

Hi, Dan!

The missing underscore was actually just me capturing the screen when the cursor was on; sorry about that. I wrote the utt2spk_to_spk2utt output directly to a file and then ran diff and cmp, and the commands said that the files were identical. I tried running run.sh on the autogenerated file, and it still didn't work (I'm attaching screen captures and relevant files). Unless there's some problem in how I named the speaker and the utterances, I don't know why I'd continue to issue the error message that spk2utt and utt2spk don't seem to match.

Thanks again for the help!

[Daniel Povey]

Well there are other errors too, e.g. about the sorting order. That
might be the cause. You could try running fix_data_dir.sh on the
data-dir and see if it resolves the problem.
And as I said before, making all the utterances have the same speaker
if you don't have the speaker info is generally a bad idea.

[Rolando Coto]

I ran fix_data_dir.sh on both the training and the test data and it worked! The fix for the test data gave me the following output:

     rolando@Bender:~/kaldi/egs/bribri01$ ./utils/fix_data_dir.sh data/test
     ./utils/fix_data_dir.sh: file data/test/spk2gender is not in sorted order or not unique, sorting it
     fix_data_dir.sh: kept all 225 utterances.
     fix_data_dir.sh: old files are kept in data/test/.backup

After that there were a few more issues, but I managed to make it recognize one word (capture attached). Thanks for your help!