Hi:

I have followed the HTKbook tutorial to create the HMMs with the train part of the VoxForge corpus. Now, I'd like to test these HMMs with the test part.

Is there any way of reference to do this? I mean if there is a reference grammar or wdnet (wordnet) in this web, in order to compare differents HMMs. Do you use LM, or using a word loop is enough?

Thank you,

--- (Edited on 5/12/2011 4:19 am [GMT-0500] by igor_odriozola) ---

>I'd like to test these HMMs with the test part.

I just use HResults as described on this page: Testing Your Acoustic Model with HTK & Julius

There is some good AM testing info in this post: Acoustic model testing

--- (Edited on 6/7/2011 12:40 pm [GMT-0400] by kmaclean) ---

Ok, thank you, I will read it in depth. Nevertheless, there is a question I cannot solve, and I would like to explain it here.

In a first evaluation, I have created a word loop in a bnf file, with the 5062 words corresponding to the test part of the corpus:

({sil} <$word {sil}>)

And I have obtained these results:

====================== HTK Results Analysis =======================
  Date: Wed Jun  8 11:49:41 2011
  Ref : testref.mlf
  Rec : recout.mlf
------------------------ Overall Results --------------------------
SENT: %Correct=2.07 [H=57, S=2702, N=2759]
WORD: %Corr=47.16, Acc=16.87 [H=12137, D=651, S=12947, I=7796, N=25735]
===================================================================

I think that, of course, a LM could improve these results, but first of all I would like to improve the acoustic probabilities. That is why I have thought about using the word-network-expansion options of HTK.

If I execute the HVite with a monophone dict with:

ALLOWXWRDEXP = FALSE
ALLOWCXTEXP = TRUE
FORCECXTEXP = TRUE
CFWORDBOUNDARY = TRUE

I obtain the same result as when I use a triphone dictionary, because the system only performs context expansion in the word.

If I set ALLOWXWRDEXP = TRUE to allow context expansion across words, I obtain this error:

ERROR [+8230]  NewPronHolder: Every word must define some context [sil=-1/0/-1]

I suppose the system manages the "sil" model like the other phoneme-models and is looking for models like "a-sil+b". I do not know how to indicate that this model should be treated as a context-free one (note that I have set CFWORDBOUNDARY = TRUE).

Anybody knows how I can use these settings?

Thank you,

Igor Odriozola

--- (Edited on 6/8/2011 5:37 am [GMT-0500] by igor_odriozola) ---

--- (Edited on 6/9/2011 9:03 am [GMT-0500] by igor_odriozola) ---

>using the word-network-expansion options of HTK.

This is a question for the HTK users list

--- (Edited on 8/2/2011 10:03 am [GMT-0400] by kmaclean) ---