[hts-users:03728] Re: A problem with decision tree-based context clustering

[Fahime Bahmany <fahime.bahmany@xxxxxxxxx>]

Hi,

I intended to create tied-state triphones using HMM Toolkit (HTK) in order to train a context-dependent HMM.
Attached tree results from my experiment which is generated by running following command:

HHEd -H myhmms/hmm3_tri3/newMacros -M myhmms/hmm3_tri4(8) tree(8).hed triphones1 > log

This command is written based on the explanation given by Section 3.3.2 of HTK-book.

The initial part of this tree for 2-nd state is as follow:
   0            'F4_sonorant'      -1         -3   
  -1            'F3_sonorant'      -2         -6   
  -2        'F3_non-sonotant'      -4         -8   
  -3           'F4_vowel_low'      -5         -7   
  -4               'F5_Vowel'  "tri_s21"     -45
 
It is obvious that the terminal node "tri_s21" is unreachable; since, for reaching this leaf F4_sonarant, F3_sonorant, and F3_non-sonotant must be true. However, F3 cannot be both sonarant and non_sonorant.
I am really surprised at this problem.

Thanks a bunch in advance,
Fahimeh

 

On Sun, Apr 21, 2013 at 5:38 PM, Yongxin Wang <fefe.wyx@xxxxxxxxx> wrote:

Would you provide us your clustered tree and show us the path that you
think is unreachable?

On Sun, Apr 21, 2013 at 3:44 AM, Fahime Bahmany
<fahime.bahmany@xxxxxxxxx> wrote:
> Hi,
>
> I have a problem with decision trees which are resulted in context
> clustering.
> Some leaves of the tree may be unreachable when we cluster the contexts. In
> other words, when we trace some paths of the tree we may encounter questions
> that are in contrast to one of the questions in that path. This problem
> causes the terminal node in that path to be unreachable.
>
> I would be really grateful if someone could help me through this problem.
> Thanks in advance
>
>

Attachment: trees(8)
Description: Binary data