WSJ scoring + NIST compatibility

[Sean Robertson]

Hello,

Thanks for taking the time to read.

I'd just like some verification w.r.t. scoring for the WSJ recipe. AFAICT, it looks like Kaldi uses its own internal functions to compute WER in the WSJ corpus rather than using sclite. Also AFAICT, Kaldi uses a fixed insertion, deletion, and substitution penalty when aligning reference to hypothesis in compute-wer. My understanding of the NIST standard (such as sclite) is to use the weights 3, 3, and 4 when aligning. So it looks like Kaldi will generate different word error rates than sclite. Am I correct? Did I mess up somewhere?

Thanks again for your time,
Sean

[Jan Trmal]

You might be right. But depending on the flags sclite is called with(esp. Fragment scoring) it tends to be more forgiving than kaldi scores.

Y.

--
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 view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/c4043aaf-b883-4a81-a409-8f12b0938dd6%40googlegroups.com.