Acoustic Model Discussions

Flat
A little help needed
User: Rio
Date: 11/13/2009 7:55 pm
Views: 7210
Rating: 6

Hi!

I have a few simple questions:

a) In order to be able to use the Julius SRE with a vocabulary and grammar that is not simply from the "dialing a phone number" scenario (say I want to start recording other things), must I create a brand new language model? I have successfully tested Julius using the sample.voca and grammar from the nightly releases, but it's also constrained to the voca and grammar- if I were simply to change the voca and grammar, would it work or is that not correct?

b) Consequently, does the Speaker-Independent Acoustic Model available for download only useful for the language model of "dialing a phone number"? Or does it incorporate other wonderful language models like CSLU?

--- (Edited on 11/13/2009 7:55 pm [GMT-0600] by Visitor) ---

Re: A little help needed
User: kmaclean
Date: 11/16/2009 12:24 pm
Views: 2236
Rating: 7

Hi Rio,

>must I create a brand new language model?

Yes, for dictation type application you need to create a language model.

>but it's also constrained to the voca and grammar- if I were simply to

>change the voca and grammar, would it work or is that not correct?

See the VoxForge tutorial Step 1 - Task Grammar for details on how to do this

>does the Speaker-Independent Acoustic Model available for

>download only useful for the language model of "dialing a phone

>number"? Or does it incorporate other wonderful language models

>like CSLU?

Julius supports the use of language models - however, you recognition rates won't be so good...

You might want experiment with Keith Vertanen's Language models: CSR LM-1 language model training recipe; or the English Gigaword language model training recipe.

Ken

--- (Edited on 11/16/2009 1:24 pm [GMT-0500] by kmaclean) ---

PreviousNext