Using not determinized HLG

[zed...@gmail.com]

Hi,

I would like to know if there is any issue in the inference results if I don't determinize my decoding graph. When I say "issue", I mean pretty much WER (and recognition time, if possible).

To give you some context: in my application, when doing inference, I use a hand made grammar and when I try to apply determinizeStar(LG) it takes some hours.
I understand determinzation as an optimization step that, among other things, saves time when decoding. But if I don't determinize, should I expect any difference in the recognition result? I think I might experience some slower decoding but the final result won't change. Is it correct?

Thank you,
Eduardo

[Daniel Povey]

The result should be the same in principle.

Normally people add disambiguation symbols before determinizing and later remove them, to stop determinization from blowing up.  

--
Go to http://kaldi-asr.org/forums.html to find out how to join the kaldi-help group
---
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+unsubscribe@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/kaldi-help/0de57460-df8b-4554-a9ad-5477e1088f8dn%40googlegroups.com.