See ticket 153.

Ken

--- (Edited on 5/17/2007 11:33 am [GMT-0400] by kmaclean) ---

Hi Ken,

Thanks for quick response. Since this proved to be not an actual bug, I'm posting here from now on.
Your advice proved to be useful; i ignored the last steps (hmm13-hmm15) and used hmm12/hmmdef and hmm12/macros to be processed by julian. I commented "-hlist tiedlist" option in julian.jconf, having no tiedlist.
Now... julius seems to be working normally but the error rate is very high ... (could this be because I didn't used a tiedlist???)

My question now is how can I improve my results; is tiedlist mandatory for an accurate result?
As far as I understand tree.hed is a language specific file... can it be customized?
Any experienced user with this???

Thanks,
Lucian

--- (Edited on 5/18/2007 2:02 am [GMT-0500] by Visitor) ---

Hi Julian,

You said (in the bug report) that you are trying to create a Romanian Acoustic Model. 

What are you using for your phonemes?  Are you trying to use the English Phoneme alphabet with Romanian words?  Did you create your own phonemes?

Ken 

--- (Edited on 5/18/2007 2:57 pm [GMT-0400] by kmaclean) ---

What are you using for your phonemes?

I'm using the phonemes found in the examples (that is, as far as I know, IPA alphabet)

Are you trying to use the English Phoneme alphabet with Romanian words?

yes, although Romanian is a phonetic language (with the exceptions of diphtongs and some other group we pretty much read words as they are written) I considered at the first stage  that I can use pretty much the same phoneme alphabet as in English. After all a speech recognition engine can be trained for recognizing different pronuntiations of a same word (as in accents)

Did you create your own phonemes?

Yes, but without success maybe I don't know where to put them correctly. For instance I tried to add "ii" which doesn't exists in English Phoneme Alphabet, to denote a voiceless a close (high) unrounded central vowel specific to Romanian language.

 In other order of ideas, apparently I figured out why I got that message "ERROR [+2662]  AssignStructure: cannot find tree for b state 2".

It seems that at TB level HHED is searching for an entry like

TB 350 ST_b_s2 {b-*,....} (I don't remember the exact form but I modified an existing one)

After that, the parser will complain about "cannot find tree for b state 3"

All I did was adding commands for all  the missing states and (hooray!!!) now I'm getting a new error.

But this time it's an easy one(I hope...)

 FindProtoModel: no proto for "some phoneme" in hSet

As much as I could understand this is an error that usually comes out if you did trained an unsufficiently balanced vocabulary. Well, it might as well be something but this error, as much as I can see now is pretty well documented on the net.

Any suggestion is highly appreciated, 

Lucian

--- (Edited on 5/18/2007 4:45 pm [GMT-0500] by Visitor) ---

I realized that this "missing state 2 for b" problem is due to the double chop in HTK_Scripts/mkclscript.prl script, as instructed by the tutorial instructions.

After I change it back to a single "chop;" command, everything works.

--- (Edited on 2/26/2008 1:24 am [GMT-0600] by Visitor) ---