[hts-users:01396] Re: Questions about "reg.base" when doing adaptation

["Heiga ZEN (Byung Ha CHUN)" <zen@xxxxxxxxxxxxxxx>]

Hi,

QiXiong Tu wrote (2008/05/20 22:11):

> In the file "reg.base", it's written:
>  
> <CLASS> 1 
> {dh^ih-s+p.......state[2].stream[4].mix[2] 
> ..........
>  
> here, let's assume "dh^ih-s+p........state[2]._stream[4].mix[2]" as 
> the second mixture of  the stream "~p "lf0_s2_47-4".
> Actually, there're a lot of "~p "lf0_s2_47-4" in the model file "MMF".
>  
> Then my questions are listed as follows:
>  
> 1.when classing (generate the "reg.base" file), I think HHEd tool using 
> "dh^ih-s+p........state[2]._stream[4].mix[2]"  as a delegate for so 
> many "~p "lf0_s2_47-4" in the model file "MMF", just because it's 
> found firstly when classing, is it so?

Yes.

> 2.Does HTS do adaptation of the second mixture of MSD?

No, they are 0-dimensional distribution and they are skipped:

HAdapt.c:

Boolean GenXForm(RegNode *node, XFInfo *xfinfo, IntVec classes, int xfindex)
{
  AccStruct *accs;
  int class=1,b;
  IntVec blockSize,bandWidth;

  AdaptXForm *xform = xfinfo->outXForm;
  AdaptXForm *diagCovXForm = xfinfo->diagCovXForm;

  if (node->valid && node->vsize>0) {
                     ^^^^^^^^^^^^^^

node->vsize is the dimensionality of distributions clustered in this node.

Regards,

Heiga ZEN (Byung Ha CHUN)

--
------------------------------------------------
Heiga ZEN     (in Japanese pronunciation)
Byung Ha CHUN (in Korean pronunciation)

Department of Computer Science and Engineering
Nagoya Institute of Technology
Gokiso-cho, Showa-ku, Nagoya 466-8555 Japan

http://www.sp.nitech.ac.jp/~zen
------------------------------------------------