I found some of the lines misleading.
1) en-us should have mdef otherwise it wont work was not mentioned. (I think its in adaptive model mentioned how to create mdef.)
2) word_align is in pocketsphinx/test and the script of the decoder should have been perl (path)/word_align.pl test.transcription test.hyp
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
It is getting difficult to follow your question. You asked how to compute WER. For that you decode a speech file and get test.hyp hypothesis. Then you use word_align to count the errors in hypothesis with respect to the reference transcript
Did you actually want to know how to runpocketsphinx to decode the audio. Please clarify what is the problem that you are trying to solve...
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
for that I needed to know where could I find the WER calculation in pocketsphinx. That resolved.
But while following the steps in (http://cmusphinx.sourceforge.net/wiki/tutorialtuning) I found some problems which I thought to share.
So currently, I am trying to understand what is implemented for WER.
I have one more doubt, how to incorporate sclite.
Are we looking forward to make a bridge which connects the pocketsphinx and sclite tool by creating the different format of hyp, ref, ctm,stm files which is easily recognized by sclite?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
OK, now it's clear. To be honest, I do not know who was the author of this project idea and what was behind that.
But while following the steps in (http://cmusphinx.sourceforge.net/wiki/tutorialtuning) I found some problems which I thought to share.
yeah, please clarify. Not sure what is this problem
So currently, I am trying to understand what is implemented for WER.
WER is now computed with word_align.pl or sclite. Probably the idea was to implement it inside sphinx to provide reference file as an option and calculate the error without the external script.
I am not sure if that was the idea...
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Can you please guide where in pocketsphinix I could find the modules to calculate WER.
Hi
You need word_align.pl from sphinxtrain (installed separately from pocketsphinx) http://cmusphinx.sourceforge.net/wiki/tutorialtuning
Alternatively, sclite toolkit can be used
Hi Arseniy,
I found some of the lines misleading.
1) en-us should have mdef otherwise it wont work was not mentioned. (I think its in adaptive model mentioned how to create mdef.)
2) word_align is in pocketsphinx/test and the script of the decoder should have been perl (path)/word_align.pl test.transcription test.hyp
It is getting difficult to follow your question. You asked how to compute WER. For that you decode a speech file and get test.hyp hypothesis. Then you use word_align to count the errors in hypothesis with respect to the reference transcript
Did you actually want to know how to runpocketsphinx to decode the audio. Please clarify what is the problem that you are trying to solve...
So the main objective I am trying to achieve is http://cmusphinx.sourceforge.net/wiki/projectideas#implement_wer_evaluation_framework
for that I needed to know where could I find the WER calculation in pocketsphinx. That resolved.
But while following the steps in (http://cmusphinx.sourceforge.net/wiki/tutorialtuning) I found some problems which I thought to share.
So currently, I am trying to understand what is implemented for WER.
I have one more doubt, how to incorporate sclite.
Are we looking forward to make a bridge which connects the pocketsphinx and sclite tool by creating the different format of hyp, ref, ctm,stm files which is easily recognized by sclite?
OK, now it's clear. To be honest, I do not know who was the author of this project idea and what was behind that.
yeah, please clarify. Not sure what is this problem
WER is now computed with word_align.pl or sclite. Probably the idea was to implement it inside sphinx to provide reference file as an option and calculate the error without the external script.
I am not sure if that was the idea...