Just completed my first run on creating a speaker dependent acoustic model using the voxforge tutorial.  I ran into a problem with step 10 with a missing proto error so I decided to script everything in bash.  Its pretty rough but it  seems to work. With a small vocabulary set and with a pruned lexicon, it completes in seconds.  If anyone is interested, let me know where to upload the files.  As always, use at your own risk - cannot stress that enough. Don't blame me if something goes wrong.

Regarding the proto error in step 10, I figured the best way to deal with it with a limited set of training samples is to remove words from the lexicon that contain phonemes that are not in my monophones0. This process takes quite a while so run it once and save the modified lexicon.  Of course, going the other way, adding training samples that contain the missing phonemes would also work. 

This is such a fun project.

--- (Edited on 4/15/2010 3:14 pm [GMT-0500] by cmisip) ---

Here's an idea.  Why not simply make each word its own class and then spell out each possible combination of words in the grammar file.  Then it would be easy to script the .voca and .grammar file creation based on these.  Then julius will use the contexts of the grammar files to return even complex sentences and if these sentences are unique enough, accuracy should be high even on speaker independent models.

--- (Edited on 4/17/2010 10:25 am [GMT-0500] by Visitor) ---