Word-level TIMIT Monophone Model

19 views
Skip to first unread message

TY_A_61_atharva deshpande

unread,
Apr 1, 2026, 1:34:17 AMApr 1
to kaldi-help
I was trying to train a word-level monophone model on TIMIT dataset. However, on doing the scoring on the test data of TIMIT, I am getting a WER of 77%. 
Could anyone please guide as to where I could be possibly going wrong?

Thanks.

Jan Yenda Trmal

unread,
Apr 1, 2026, 3:48:13 AMApr 1
to kaldi...@googlegroups.com
Hi, I'm not sure what you mean by word-level monophone model?
Anyway, timit is tiny, so if you are attempting something complex, that might be the reason. Or, the model is too small. Or, your abstraction is invalid.
You are not providing any technical information for me to be more specific :/
y.

--
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+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/kaldi-help/19ea9e01-68dd-41b3-9ebd-17657578efb5n%40googlegroups.com.

Jan Yenda Trmal

unread,
Apr 6, 2026, 4:23:53 AMApr 6
to kaldi...@googlegroups.com
Hi, sorry for the delay 
Have you tried the official Kaldi example? You should try running it.

_I think_ that the results in https://github.com/kaldi-asr/kaldi/blob/master/egs/timit/s5/RESULTS are actually phone error rates, not word error rates, so you might not be far off -- 37% PER is pretty bad and usually means the WER will be much worse.
But I might be wrong; please keep it in mind.

Also, as a general rule in these recipes, no-one cared too much about the performance of the monophone system, its purpose was only to provide initialization for triphone systems. 
So you can play with the number of gaussians (--totgauss) which defaults to 1000, to improve the score, if you for some reason care.

I cannot provide more help now (I cannot even run timit now), but perhaps/hopefully someone else will be able to.
y,

Atharva S V Deshpande

unread,
Apr 6, 2026, 8:49:10 AMApr 6
to kaldi...@googlegroups.com
Thank you for your reply. I'll look into it.

Reply all
Reply to author
Forward
0 new messages