how can I get the log likelihood of each model unit when decoding

[Wei Li]

Dear all

I have used Kaldi to train word level DNN-HMM acoustic models, such as word "yes", "no". And my grammar is ( yes | no). and used kaldi to decode the given wave files. after decoding, I get the CTM files. but I found the score in the CTM file is confidence score, not the log likelihood of word yes, or no.   

So how can I get the log likelihood of each model unit when decoding? what's the relationship between confidence score and  log likelihood?

Thanks in advance!

[Daniel Povey]

Log likelihoods are not generally that meaningful (this especially used to be true in the days of GMMs).  Confidence score is similar to a posterior computed over the lattice.  It would be possible to get the log likelihoods in the correct place as part of the lattice by rescoring the lattice using programs like nnet-compute | lattice-rescore-mapped.   But we currently don't have a program to output them in ctm-like format. 

Dan

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

[Wei Li]

HI, Povey,

Thanks for your kindly reply. I plan to use log likelihood ratio test to verify my detection result. So, you mean I can use confidence score to take place of Log likelihoods to do verification? am I right? 

在 2015年7月31日星期五 UTC-4下午1:36:08，Dan Povey写道：

[Daniel Povey]

Yes, confidence scores are definitely the most straightforward way to do that.
In future, at some point, we need to do some work on improving the confidence scores and having them based on something more accurate than lattice posteriors alone (e.g. putting a bunch of things into a logistic regression).  But that's not done yet.

Dan