Train HMM without fillers?

[osman b]

Hello,
I want to train one, long (which have many states) HMM using end point detected wave files with Sphinx. Since there wont be any silences at the end and beginning of end point deteceted input wave files, I dont need to include filler models in my HMM.

I would like to ask, is there a way to train an HMM without silence models in Sphinx?

Best regards,

Osman

[Nickolay V. Shmyrev]

Hi

I think it's easier to have a silence around. From the source code investigation it doesn't look you can train models without them. Also, you need to have silence model for decoding, don't you?

[osman b]

Thanks for your answer
I also think that training and decoding without SIL is not possible with
Sphinx. However I need to have such a system; just one HMM for a specific
sentence. Before training HMM, I run an aligner and clip silence parts from
the wave file.
Now I am trying to implement the system in different way since first thing
that came to my mind seems to be impossible.