help with creation of RTTM files

[Yash]

Hi,

I have been working with the kaldi-KWS scripts and now I want to evaluate it. 

It looks like all I have to do is prepare an RTTM file and put skip-scoring to false in the kws_score.sh (please correct me if I am wrong here.)

The kaldi page on KWS says that, "you can generate the rttm file by force aligning the search collection with a trained model". 

Can anyone tell me the details of how exactly we can do this? I see there are some align_*.sh scripts in steps. Am I supposed to use these? further more, how am I supposed to interpret the output? A step by step guide would be very helpful here.

Thanks

Yash

[Yash]

More importantly, I am evaluating KWS only for OOV words. Can I force align a search collection even if its has OOVs?

Thanks!

[Jan Trmal]

There is a script  in egs/babel/s5c/local named ali2rttm.sh  (or something like that).
Ad the OOV words -  you would have to create a new lang directory  using a lexicon  that contains all the words.  If figuring their pronunciation would pose an issue,  you could always define the pronunciation  to be the same as UNK word.  You don't need the g.fst
Y.

--
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.