Acoustic Model Discussions

Flat
Acoustic model problem...
User: orfeas
Date: 9/29/2008 3:11 am
Views: 5654
Rating: 3

Hi everyone,


I use HTK for about 2 years so i already have an acoustic/language model ready that works perfectly with HVite.

So i try to decode with Julius and messages like this come out:

  line 59992: triphone "G-J+A" not found

and then comes a list of biphones and triphones that are not found:

 Missing phones:
*-D+d or biphone D+d
*-D+p or biphone D+p
*-D+s or biphone D+s
*-g+c or biphone g+c
*-p+D or biphone p+D
*-p+G or biphone p+G
*-p+m or biphone p+m
.

.

.

So far, i get that julius is searching my pronunciation dictionary, tranform each entry into triphones and outputs an error for each triphone i have not trained (and consequently is not in the acoustic model)..

So my question is: Julius handles only acoustic model with cross-word triphones???

What about word-internal triphones? In HTK i never had such a problem.

Every help would be appreciated!!!!


Orfeas

--- (Edited on 9/29/2008 3:11 am [GMT-0500] by orfeas) ---

Re: Acoustic model problem...
User: kmaclean
Date: 9/29/2008 5:13 pm
Views: 2453
Rating: 3

Hi Orfeas,

>acoustic/language model ready that works perfectly with HVite

Are you using Julius for dictation? I'm more familiar with Julius for command and control (i.e. using grammars)...

>Julius handles only acoustic model with cross-word triphones???

>What about word-internal triphones?

The best place to ask this would be on the Julius forum.

>In HTK i never had such a problem.

HTK is a research toolkit, so you can experiment with lots of different things.  Based on my experience, Julius tends to perform better than HVite on the same acoustic model.  It is the old features vs performance trade-off...

Ken

--- (Edited on 9/29/2008 6:13 pm [GMT-0400] by kmaclean) ---

PreviousNext