Any results on pocolm?

[Daniel Galvez]

I remember that pocolm was developed a while back, and now I see it's been used only in the tedlium_r2 corpus,even though the pocolm repo on github mentions the swbd and swbd_fisher recipes.

Were there ever any results on pocolm's language models vs a toolkit like IRSTLM or SRILM?

Would it be recommended to use pocolm at this point?

--

Daniel Galvez

[vinc...@yahoo.com]

We worked with Pocolm on Tedlium_r2. We also use it for our internal production systems before rnnlm rescoring (which is now available and much better)
It's really close to Srilm.
The main thing to me is the license term, which makes it freely available for production.

Vincent

[Daniel Povey]

The perplexity results on plain estimation are just a smidgen better, but the difference is small enough that it's not worth talking about.  (like 120 to 120.2 or something like that--order of magnitude).  There are also gains when doing pruning, but again, not vast differences.  (With SRILM, if you are going to prune heavily you just have to remember to use non-kneser-ney smoothing, I think good-turing might have been the recommended one).

It's also arguably more convenient for doing interpolation, than SRILM; but on the other hand, it's not as fast.  I ended up deciding not to put much more effort into it because there really wasn't a killer difference versus SRILM.

Dan

--
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+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/da762828-ded7-440c-8250-75ec4ab3fa07%40googlegroups.com.

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