[hts-users:01064] Re: Problem about HMGenS

["wei chen" <cwsunshine@xxxxxxxxx>]

Hi,Heiga Zen,
       Sorry to interrupt you! I'm confused now.Yesterday,I asked you how to reduce the vector size of delta, delta-delta,e.g. original feature size is 25,but delta is 10,and delta-delta is 15?you told me one method,
       If I were you, I do as follows:
     A feature vector consists of static, delta, delta^2 features as
       c(1),..,c(25)
       delta-c(1),..,delta-c(10)
       delta^2-c(11),..,delta^2-c(25)
     First I train HTS using this feature vector.
    And then I add nuisance dimensions as
      c(1),..,c(25)
     delta-c(1),..,delta-c(10),

     *delta-c(11),..,delta-c(25)*
    *delta^2-c(1),..,delta^2-c(10)*,delta^2-c(11),..,delta^2-c(25)
     where *...* denotes nuisance dimensions.
       If I set all inverse variances and means in nuisance dimensions to 0,
   they are not taken into account in the parameter generation algorithm.
  By running HMGenS with appropriate window coefficients,
  you can obtain proper outputs equivalent to what you really want to do.

   I have trained them,but I found the size of tiedlist and re_clustered_all.mmf.1mix reduce a lot : 
the space of  re_clustered_all.mmf.1mix reduces from 11M to 1M ,tiedlist also reduces from 9M to 1M,    why?
      And If I want to modify HMGenS  code,Could  I  just  set  pst->mseq[t] and pst->vseq[t] of the vector I have not used in delta and delta^2 to zero in the function SetupPdfStreams()? 


Thanks a lot!

2007/12/18, wei chen <cwsunshine@xxxxxxxxx>:

Thanks a lot! I will do it as what you suggest!

2007/12/18, Heiga ZEN (Byung Ha CHUN) < zen@xxxxxxxxxxxxxxx>:

Hi,

wei chen wrote (2007/12/18 10:18):

> 1) how could I design the window, is there any method?

If I were you, I do as follows:

A feature vector consists of static, delta, delta^2 features as

c(1),..,c(25)
delta-c(1),..,delta-c(10)
delta^2-c(11),..,delta^2-c(25)

First I train HTS using this feature vector.
And then I add nuisance dimensions as

c(1),..,c(25)
delta-c(1),..,delta-c(10),*delta-c(11),..,delta-c(25)*
*delta^2-c(1),..,delta^2-c(10)*,delta^2-c(11),..,delta^2-c(25)

where *...* denotes nuisance dimensions.
If I set all inverse variances and means in nuisance dimensions to 0,
they are not taken into account in the parameter generation algorithm.
By running HMGenS with appropriate window coefficients,
you can obtain proper outputs equivalent to what you really want to do.

> 2) do I just modify Calc_WUM_and_WUW() to support the window to
> achieve my goal?

Yes, but you may need to turn off some consistency checks for dimensionality in HGen.c and HMGenS.c.

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
------------------------------------------------