Word duration and Editing the hmm parameters

[Karthik]

I am pretty new to Sphinx and I have a couple of questions. It'll be great if
someone can help me out. Thanks.

First, once you train files (say wav files of connected tidigits) using sphinx
is there some sort of segmentation done internally? Is it possible to get the
average duration of a particular word as it has occurred over a large number
of training files?

And second, how can we define a different number of states individually to
define different digits?

Thanks again :).

[Nickolay V. Shmyrev]

First, once you train files (say wav files of connected tidigits) using
sphinx is there some sort of segmentation done internally? Is it possible to
get the average duration of a particular word as it has occurred over a large
number of training files?

Yes, you can dump word segmentation during forced alignment state with
-wdsegdir option which you can append to sphinx3_align arguments in the
scripts

And second, how can we define a different number of states individually to
define different digits?

No, you can't do that. Instead, you should use word-dependent phones. See
tuturial:

http://cmusphinx.sourceforge.net/wiki/tutorialam

[Karthik]

No, you can't do that. Instead, you should use word-dependent phones. See
tuturial: http://cmusphinx.sourceforge.net/wiki/tutorialam

Thanks :).

Yes, you can dump word segmentation during forced alignment state with
-wdsegdir option which you can append to sphinx3_align arguments in the
scripts

Ah. That clears up a lot. But do you mean I have to edit all the perl scripts
individually?

[Karthik]

I think I got it. Thanks a ton :).