Acoustic Model Discussions

Flat
Re: Model Creation Script
User: bedahr
Date: 4/18/2010 3:18 am
Views: 52
Rating: 1

Hi!

It looks alright at the first glance and yes you should do the same for gmm2 and gmm3.

 

What you just did raised the mixture count to 4. The next step will increment it to 7 (gmm2.hed) and the next one to 10 (gmm3.hed).Depending on the amount of speakers in your model you should experiment on what is the optimal amount of mixtures. We use ten because it doesn't take too long to build and provides a good value for small to medium sized models. Considering the size of the voxforge model you might want to increase this even further.

 

Theoretically, the HTK will remove "empty" mixtures during reestimation so a value too large won't cause much trouble.

 

Increasing the amount of mixtures in small steps (like we do it) is recommended because it produces better overall results.

 

Greetings,

Peter

--- (Edited on 4/18/2010 3:18 am [GMT-0500] by bedahr) ---

Re: Model Creation Script
User: cmisip
Date: 4/22/2010 12:13 pm
Views: 80
Rating: 1

I tried as far as gmm1.hed but my recognition results got worse.  Perhaps because I have only one speaker and 120 prompts only? 

Thanks

Chris

--- (Edited on 4/22/2010 12:13 pm [GMT-0500] by cmisip) ---

Re: Model Creation Script
User: bedahr
Date: 4/22/2010 1:12 pm
Views: 2339
Rating: 1

Your recognition shouldn't get (significantly) worse but increasing the amount of mixtures makes little sense if you only have such a small model.

 

Greetings,

Peter

--- (Edited on 4/22/2010 1:12 pm [GMT-0500] by bedahr) ---

PreviousNext