Speech Recognition Engines

Flat
Making two different acoustic models' performance comparable?
User: asaaki
Date: 1/13/2011 1:26 am
Views: 4010
Rating: 5

Hi,

I have two different acoustic recognizers trained on some data, for the same language. I'm doing monophone recognition.

One of my models uses 32 phones, the other one was built using 62 phones. But both were trained on the same corpus. I thought I could simply run HResults to find out the rate of recognition for both recognizers, but I realized that these are not comparable, because they both rely on different numbers of HMMs.

How can I use one of the models (say the one with 32 phones) as a baseline model with which to compare the performance of acoustic models built on other numbers of phones? I understand that theoretically I should "map one set of phones onto the baseline set", and then use HResults, but I really don't understand how to do that.

I hope my question makes sense?

--- (Edited on 1/13/2011 1:26 am [GMT-0600] by Visitor) ---

Re: Making two different acoustic models' performance comparable?
User: kmaclean
Date: 3/21/2011 12:47 am
Views: 1622
Rating: 4

>but I realized that these are not comparable, because they

>both rely on different numbers of HMMs.

You should still be able to use HResults since it only checks for words and sentence results... the number of hmms is not relevant.

 

--- (Edited on 3/21/2011 1:47 am [GMT-0400] by kmaclean) ---

PreviousNext