score asr system?

[K.R]

Is it possible within kaldi to score ones asr?.. 

I seem to be running the score.sh and the decode.sh but cant seem to see any stats, or any scores about the asr. 

what kind of test can be done using kaldi, and how do i extract the data?

[K.R]

* I am running the score.sh and decode.sh...

[Daniel Povey]

local/score.sh is a setup-dependent soft link to the real scoring
script, or a setup-dependent script, so hard to say in general.
the WER will generally be obtainable either by doing

grep WER decoding-dir/wer_* | utils/best_wer.sh

or (if sclite)

grep Sum decoding-dir/scor*/*ys | utils/best_wer.sh

there are more detailed stats available, but they depend which scoring
script you are using. If you poke around in the decoding directory
they won't be hard to find.
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.

[K.R]

I seem to have found some results related to the WER in /decodefolder.. 

Some help interpreting them could be usefull.. 

ls scoring_kaldi/wer_detalis has these files: 

ls scoring_kaldi/wer_details
lmwt  ops  per_spk  per_utt  wer_bootci  wip

Which seem to be include information which words it has detected correct, deleted or substituted with another file. 

And data related for each speaker..

But that is purely based on how things are set up, and the scripts generates more files than those, could be nice to know what those are as well... 

Here is a tree structure of the files created..

http://pastebin.com/pLK2MbeU

Any details on how it scores/decodes.. 
Why are there so many wer, but only one wer details?

[Daniel Povey]

I don't have time to help you with this type of question-- someone
else might choose to.
Most people figure this type of thing out for themselves.

[Danijel Korzinek]

"scoring_kaldi/best_wer" contains the final result of the system

- lmwt is the language model weight of the best result

- wip is the word insertion penalty of the best result

- ops contains a list of errors for each word with the counts for the whole experiment - this is useful for debugging pronounciation lexica

- per_spk  contains a list of scores per speaker - this is useful for determining if certain speakers/sessions are worth investigating/removing

- per_utt  contains the most detailed list of errors for each recorded utterance

- wer_bootci is the confidence interval for statistical significance testing - tells how robust your score actually is (good for publications)