Are you using numbers/special characters for the name of phones? it's better to use just english letters.

--- (Edited on 8/9/2007 3:13 am [GMT-0500] by nsh) ---

Yes, we use"?", "@", "-^" and some other characters!

But what could we do? Which characters should we use instead of them?

Oh, and we have removed all numbers!

Bettina

--- (Edited on 8/9/2007 9:21 am [GMT-0500] by Visitor) ---

Does this mean that IPA characters cannot be used as well? Or simply no +, ?, } etc

 Robin

--- (Edited on 8/9/2007 11:13 am [GMT-0500] by Robin) ---

+ and  -  used in triphone names, so it's better avoid them. ? and } I'm not sure, but nobody can be sure.

Use minus and plus instead like "aplus" instead "a+"

--- (Edited on 8/9/2007 11:39 am [GMT-0500] by nsh) ---

--- (Edited on 8/9/2007 11:40 am [GMT-0500] by nsh) ---

But if we let HTK do the segmentation of the phonemes (i.e. no " " between them in the dictionary) - how can we ensure that he recognizes "plus" as "+" and not as "p", "l", "u",...? -- Peter Grasch

--- (Edited on 8/10/2007 2:51 pm [GMT-0500] by Visitor) ---

Hm, so many different people are asking so I'm afraid I'm wrong

Actually there is no problem, since in dictionary phones are separated by spaces:

ONE  [] W AH N

In your case it will be something

POLKA [] P OPLUS L K A

--- (Edited on 8/10/2007 3:22 pm [GMT-0500] by nsh) ---

Hm, and I decided to add this, since I'm still afraid. Of course one can consider the task of segmentation of strings without spaces to phonemes with HMMS by force alignment. But I'm not really sure how is it related to acoustic model training.

 

--- (Edited on 8/10/2007 3:54 pm [GMT-0500] by nsh) ---

The problem is: we use the BOMP dictionary which is formated like this: ONE DIG ?|WAHN (for example) In the training process HTK automatically segments "WAHN" into "W AH N". Is it possible to supply the list of monophones per hand and force the whole HTK toolkit to use it? (Sorry if I'm asking stupid questions - I just don't know any better ^^) -- Peter Grasch

--- (Edited on 8/11/2007 7:22 am [GMT-0500] by Visitor) ---

Hi Peter,

Here's my try at figuring this out ... 

The original error mentioned in this thread occurred when create hmm13 - this corresponds to Step 10 of the Tutorial.  Steps 9 & 10 are where you create triphone acoustic models. 

If you are only interested in monophone acoustic models, you can use the hmmdefs files (in the hmm9 directory) from Step 8, and use monophones1 (rather than tieldlist from Step 10) in your Julian config file.  Note that monophone AMs do not work as well as triphone AMs.

So I think creating your own monophone list (by hand) is not what you are looking for ...  

I am assuming you are trying to use the VoxForge Howto script to create German Acoustic Models. 

Firstly, I need a better idea of what the BOMP dictionary looks like.  I looked it up, but I need to email the creator to get a copy ... Please post a longer excerpt (around 10 entries - copy and paste exactly what it looks like in the dictionary) or you can email to me if licensing prevents you from doing so.  I would also like to see what your fulllist file looks - please post this also.

I  also want to confirm that you are using a German version of the tree.hed script.  These are the "questions" that HTK asks to figure out all the triphones from the complete pronunciation dictionary you are using (not just the words in your prompts file). 

thanks,

Ken

P.S. when dealing with HTK, there are no stupid questions ... HTK is notorious for allowing things at earlier steps that create problems at a later step.

 

--- (Edited on 8/11/2007 8:37 pm [GMT-0400] by kmaclean) ---

Hi Ken! I mailed you a some entries from the original bomp dict. as well as from the imported one (we import the bomp and the wiktionary dicts. to the same format used in your tutorial). As for the tree.hed-skript: I used your english-one for small tests before and it worked flawlessly. A quick google search returned no ready-to-download german one (that would have been too easy I guess)... I looked it up in the HTK Book and I found the mkclscript - which is also included in your distribution - for (semi-)autmatic creation of the tree.hed file. What I didn't find was a step-by-step tutorial on how to use it (just some vague descriptions). Any help would be appreciated. -- Peter

--- (Edited on 8/16/2007 5:49 am [GMT-0500] by Visitor) ---