Computing confidence scores using HTK toolkit recognition

Speech Recognition Engines

Flat

User: thilinid
Date: 1/22/2013 5:01 am

Views: 7724
Rating: 9

Hi,

I'm trying to find a way of computing the confidence score of the HTK recognition output. So far haven't had any luck with that.

In the HTK decoder HVite, it outputs the total log probability of each word in the result file. What is the meaning of this term and it is calculated?

Can I use this log probability to compute the confidence scores?

Thank you in advance.

--- (Edited on 1/22/2013 5:01 am [GMT-0600] by thilinid) ---

Re: Computing confidence scores using HTK toolkit recognition

User: nsh
Date: 1/23/2013 8:18 am

Views: 689
Rating: 7

> I'm trying to find a way of computing the confidence score of the HTK recognition output. So far haven't had any luck with that.

There are multiple confidence scores - whole-sentence confidence scores, word confidence scores, phone confidence scores. Most of them could be calculated from a lattice. HTK doesn't have internal tools for that, but you can dump the lattice and use external tools like SRILM:

http://www.speech.sri.com/projects/srilm/manpages/lattice-tool.1.html

http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.33.3196

> In the HTK decoder HVite, it outputs the total log probability of each word in the result file. What is the meaning of this term and it is calculated?

This is an estimate of the probability of the observation sequence given the HMM model, you can find some theory on that in any HMM tutorial, like Rabiner's tutorial in HMM.

Some information is also in HTK book in chapter 1 "The Fundamentals of HTK".

--- (Edited on 1/23/2013 17:18 [GMT+0300] by nsh) ---

Re: Computing confidence scores using HTK toolkit recognition

User: manju
Date: 7/22/2013 6:41 am

Views: 47
Rating: 2

Hi thilind,

I also want to calculate confidence score of the HTK recognition output. Did you found some way. Please let me know. I would be grateful to you.

thanks in advance.

--- (Edited on 7/22/2013 6:41 am [GMT-0500] by Visitor) ---

Re: Computing confidence scores using HTK toolkit recognition

User: manju
Date: 7/22/2013 7:28 am

Views: 110
Rating: 2

Hi thilind,

please reply to my mail id [email protected].

--- (Edited on 7/22/2013 7:28 am [GMT-0500] by Visitor) ---

Re: Computing confidence scores using HTK toolkit recognition

User: thilinid
Date: 7/23/2013 2:29 am

Views: 70
Rating: 3

Hi manju,

As nsh said there aren't any internal tools in HTK to get the confidence score of the outputs.

What I did was: got the word lattice of the recognized output and calculated word posterior probabilities and used them as the word confidence scores to compute the utterance confidence. I got the idea from the following thesis:

http://www.mmk.ei.tum.de/publ/pdf/diss/fabian.pdf

--- (Edited on 7/23/2013 2:29 am [GMT-0500] by thilinid) ---

Re: Computing confidence scores using HTK toolkit recognition

User: Visitor
Date: 7/23/2013 4:57 am

Views: 1576
Rating: 3

Hi thilinid,

Thanks for the response.

Could you please explain the implementaion part of your work.

---->got the word lattice of the recognized output=========> i think you did it using HVITE. could you Please provide me the command.

--->calculated word posterior probabilities ==> i think you did it using SRILM lattice_tool. If you have used some other approach please tell me the same. could you Please provide me the command.

Thanks a lot,

--- (Edited on 7/23/2013 4:57 am [GMT-0500] by Visitor) ---

Re: Computing confidence scores using HTK toolkit recognition

User: Visitor
Date: 10/28/2015 8:50 am

Views: 10
Rating: 0

This may help you

http://www.ee.columbia.edu/ln/LabROSA/doc/HTKBook21/node145.html

--- (Edited on 10/28/2015 8:50 am [GMT-0500] by Visitor) ---

Re: Computing confidence scores using HTK toolkit recognition

User: Visitor
Date: 10/28/2015 8:56 am

Views: 2388
Rating: 0

This may help you

http://www.ee.columbia.edu/ln/LabROSA/doc/HTKBook21/node145.html

--- (Edited on 10/28/2015 8:56 am [GMT-0500] by Visitor) ---

Previous • Next •


Username	Password