WER for each speaker

[Sidrah Azhar]

Dear all,

Is it possible to calculate WER for every speaker separately?

I have 12 speakers in test set and 13 speakers in training data and I have trained monophone, tri1, tri2a and tri2b model (with WER to 49% in tri2v) and I want to calculate WER for every 13 speakers separately for these 4 models.

[Daniel Povey]

If you scored using the Kaldi scripts that information will be present in 

scoring_kaldi/wer_details/per_spk 

if you scored with sclite it will be in the .sys file I think.

--
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.
To post to this group, send email to kaldi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/83df3a7f-8e39-4210-a3cf-b739f3e4f3ae%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Sidrah Azhar]

i am using kaldi script local/score.sh. Actually i have put my data in voxforge setup/

[Sidrah Azhar]

And one more question. What should be the WER for this size of data?

Training data consist of 372 wav files, each wav file consist of 14 to 16 seconds and total size of training data is 171.4 MB. And most importantly, 12 speakers are deaf children and only 1 speaker is normal hearing person.

I have got 100% WER on monophone while 48% WER on tri2b.

[Daniel Povey]

WER will depend on the data.  That WER sounds pretty good for the speech of deaf children.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/55a5ccbd-2d0d-4074-85fe-129edb6d03c8%40googlegroups.com.

[Sidrah Azhar]

Thank you Dan.

[Sidrah Azhar]

One last thing Dan. Does it make a difference or how much my WER effects when all the speakers of my test data are also in training data.

I meant to say that there are 13 speakers ( 12 deaf and 1 normal hearing) in training data and same 13 speakers are in test data but the test data is obviously different from training data (Like I select some portion for test data and rest for the training data for each speaker).

How much WER effects by this?

Thanks for your help Dan.

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/seqltfNbTro/unsubscribe.
To unsubscribe from this group and all its topics, send an email to kaldi-help+unsubscribe@googlegroups.com.

To post to this group, send email to kaldi...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/2ac4c84a-9e62-46fa-8e8b-140921350b7e%40googlegroups.com.

[Daniel Povey]

It will definitely make a difference to have speakers in training and test-- how much difference depends how many speakers there are.  It's normally considered more correct to have different speakers in training and test.

To unsubscribe from this group and all its topics, send an email to kaldi-help+...@googlegroups.com.

To post to this group, send email to kaldi...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/2ac4c84a-9e62-46fa-8e8b-140921350b7e%40googlegroups.com.

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

--
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.
To post to this group, send email to kaldi...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CABF2gO9z4QKpyKiFzLU5225ZoJQcW%3DLyJk0VGPq9YAPKUs5H9Q%40mail.gmail.com.

[Sidrah Azhar]

Thank you!