[hts-users:00356] Re: about " decision trees"

["Nicholas Volk" <nvolk@xxxxxxxxxx>]

Hi,

> Hi,
>
> I  have  read the " tree-dur.inf"  ,"tree-lf0.inf" ," tree-mcep.inf"  in
> the  " HTS voices for Festival" .
>
> and    I  found  there is only one tree for  duratoin ( state 2) ,
>
> but for f0 and mcep , there are 5 trees (state 2 ,3 4, 5, 6 ) .
>
> why?

The leaf contains mean&dev data for all five states.
Not sure about the underlying motivation for only one tree,
but it's by far accurate enough and a whole lot more efficient that going
through five separate trees.

Heika: Making SearchTree() non-recursive and
inlining PMatch and QMatch enhances the overall performance.

> -----------------------
> in  the  "inf "  file,   the tree is as followed:
>
> pau[2]
> {
>      -1 Pos_C-Syl_in_C-Phrase(Bw)==1
> -2               -9
> }
>
> -1 is the index of the node ,
> -2 is the index of node->yes
> -9 is the index of node-> no
>
> why  the  index are all the negative?
> ------------------------------------------
>
> There is only one tree for every state in the "inf" file.
> does it mean the tree is used by all the phoneme.
>

Yes. If you interpret the trees, you'll notice that the first
question/pattern is typically something like vowel/consonant or
voiced/unvoiced.
After a few branches/splits the you'll begin to see patterns similar
to {*-a+*} which mean that the subbranches are about a particular phoneme
after that point. (Not all phonemes are necessarily covered this way,
but it doesn't really matter.)

regards,
  Nicholas