[hts-users:00885] Re: Training by imposing decision trees

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

Hi,

Simon King wrote (2007/10/27 15:11):

> So, how about this:
>
> - you have a decision tree previously learned; in the tree, there is a 
> macro name at each leaf
>
> - load an MMF into HHEd which contains only the definitions of these 
> macros, but no actual HMMs (i.e. you have deleted the HMM definitions 
> from the MMF, manually or using a script); I guess you also have to give 
> it an empty model list file on the command line
>
> - use HHEd's "AU" command to synthesise *all* the models you need (the 
> ones you previously deleted from the MMF, plus any unseen ones)
>
> - the models are now defined in terms of the macros, just like any 
> tied-parameter model set
>
> - save the MMF

In this case parameters of generated MMF are equal to those of MMF used to construct decision trees.
So you should re-estimate model parameters by HERest using speech data which you want to train.
However, it may cause some segmentation errors at the first iteration of HERest if you use different training data to construct these two MMFs.

For example, if you want to use male speech to construct decision trees and female speech to train MMF with imposing decision tree.
At the first iteration, you have to run HERest with female speech using male MMF.
This may lead segmentation errors.
Furthermore, models which do no appear in the female speech data are not re-estimated.
If you try to synthesize speech using this MMF, sometimes synthesized speech may become *mixture* of male and female speech.
To avoid this kind of problem, Yamagishi et al. proposed a technique called "Shared Tree Construction."

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