Is there a script can get the am score and the lm score of an utterance ?

[付嘉懿]

Hi all,

  I have a simple question. Is there any any scripts can get am and lm score for utterances ? I found that the scores can be got by using lattice-1best and nbest-to-linear.  Is this way right? or is there a combined script do the same thing ?

  Thanks!

[付嘉懿]

I tried to use lattice-1best | nbest-to-linear. to get the ac cost and lm cost. 

The value of ac cost is a negative number like -503.7546 of -4292.171. The value of lm is a positive number like 26.46918 or 111.22.

Are there any introductions about the computation process of these values? How can I compare ac cost with lm cost to find which part make more influence on the recognize result ?

Thanks!

[Daniel Povey]

the lm cost can be interpreted as the negative of a log-prob (but rememberm that there may be transition probabilities and pronunciation and silence probabilities in there as well).

The acoustic cost generally isn't meaningful unless compared against other acoustic costs.

How it was obtained depends very much on the model type.

--
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 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/49907ccb-1aac-40ef-80fc-3b7adb5b4fe7%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[付嘉懿]

Thanks Dan.

在 2018年11月13日星期二 UTC+8上午11:30:37，Dan Povey写道：

[付嘉懿]

Dan,

    When I found got the mis-recognized results of an utterance, how can I found which part (ac or lm) is more likely to cause the error ? 

在 2018年11月13日星期二 UTC+8上午11:30:37，Dan Povey写道：

the lm cost can be interpreted as the negative of a log-prob (but rememberm that there may be transition probabilities and pronunciation and silence probabilities in there as well).