Click here to register.

General Discussion

Flat
Use of the Quickstart nightly package with a HTK_AcousticModel nightly
User: okoloth
Date: 2/27/2008 5:40 am
Views: 2585
Rating: 25
Hello there, I have been using the quickstart linux nightly package for a few days and have ran into problems trying to use the HTK_AcousticModel nightly. From what I can see, the nightly Julius quickstart package contains the hmmdefs and tiedlist from the HTK_AcousticModel already. But julius is still using the sample.dict and sample.dfa of just 23 words. I would like to use the larger 11,000 word dictionary from the HTK_AcousticModel package but I’ve struggled to find any information on preparing Julius-friendly dict & dfa files.

Can you point me in the right direction?

Kind Regards,

Oko

--- (Edited on 2/27/2008 5:40 am [GMT-0600] by okoloth) ---

Re: Use of the Quickstart nightly package with a HTK_AcousticModel nightly
User: kmaclean
Date: 2/27/2008 7:53 am
Views: 292
Rating: 23

Hi Oko,

>From what I can see, the nightly Julius quickstart package contains the

>hmmdefs and tiedlist from the HTK_AcousticModel already.

Yes, they are the same 

>I would like to use the larger 11,000 word dictionary from the

>HTK_AcousticModel package but I’ve struggled to find any information on

>preparing Julius-friendly dict & dfa files.

See step 1 of the acoustic model creation tutorial. 

Note:If you are using the dict file included with the nightly build, you need to remove the "sp" phone from the pronunciations for each word you use for the Julian grammar.  Ticket #294 provides some details.  

Since the Nightly build uses tied-state triphone acoustic models, you can use the VoxForge pronunciation dictionary too:

[   ] VoxForge.tgz 13-Feb-2008 03:59 2.6M

Hope that helps, 

Ken 

--- (Edited on 2/27/2008 8:53 am [GMT-0500] by kmaclean) ---

Re: Use of the Quickstart nightly package with a HTK_AcousticModel nightly
User: okoloth
Date: 3/3/2008 5:41 am
Views: 1106
Rating: 27

Thanks Ken,

A quick bit of shell scripting and I removed the "sp"s from the pronunciations and built my own sample.dfa / sample.dict

My initial run using the full 11,000 word dictionary caused a SEGV but cropping the list to about 1,000 words worked well.

Thanks for the help!

Oko 

--- (Edited on 3/3/2008 5:41 am [GMT-0600] by okoloth) ---

PreviousNext