I'm posting this on Voxforge since the issue might be with me, HTK, Julius or elsewhere. Here's what happens :
In my speaker-dependent grammar I have about 8 sentences (of about 100) that start with WORD and have a second part QUIT, STATUS, TIME, etc. Seven of the eight are recognized perfectly 100% of the time and one of them WORD RESTART bombs out with "hypothesis stack exhausted" at least 50% of the time. Julius never thinks it is something else, it just runs out of suggestions. Recognition of the grammar as a whole is very close to 100%.
Now it gets strange. If my prompt is WORD RESTART and I say WORD ESTART or WORD START (neither of these is in my grammar) then it returns WORD RESTART 100% right all the time.
It seems something is happening to my Rs. I also have a problem occasionally with ZERO and ROMEO.
I'm trying to develop a theory/hypothesis list.
1. I am not saying R at all even though I think I am (some French rolling Rs might come in handy). I don't think I am saying W.
2. My mike (bluetooth) is not hearing R even though I am saying it
3. The recording tries to decipher the R but it gets mixed in with background noise
4. It is recorded but HTK misses it
5. HTK gets it but Julius misses it.

Does R already have a rap sheet? Any suggestions how I can narrow this down? I have the workaround, just omit the R while enunciating, but it would be good to have an explanation here. It's pretty hard to design a sensible grammar constantly trying to avoid Rs.

--- (Edited on 1/13/2010 8:09 am [GMT-0600] by colbec) ---

--- (Edited on 1/13/2010 8:15 am [GMT-0600] by colbec) ---