Hello, I was following the tutorial as listed at http://fliiby.com/file/269825/zxkixccizp.html for creating a yes/no recognition system. It ends up creating a single phoneme for "yes" and "no", and doesn't do the multi-phonemic approach. I thought this approach would be fine just as long as the number of possible recognizable words remained low.

The recognition accuracy was really excellent with julius and so I created a small program that does the above steps in the tutorial automatically. I trained the word "example" and added it to the .voca and .grammar, and cranked those through the mkdfa.pl script to get the appropriate output files and also added the appropriate hmm to the hmmdefs file but... it doesn't seem like Julius will ever recognize "example", and instead interprets it as "yes".

Is there a problem with using a long words as a single phoneme with julius? Do I have to adjust some specific set of parameters?

I suppose I could go the multi-phonic route but it seems pretty complex to programmatically implement and I haven't actually successfully got it to successfully run (Following the voxforge tutorial) due to some "missing tree" complaints that remain unresolved.

(BTW, if anyone attempts to follow the above linked tutorial, beware that some of the command line statements are incorrectly ordered, so you may have problems.)

--- (Edited on 10/24/2010 3:55 pm [GMT-0500] by swbluto) ---