[Subject Prev][Subject Next][Thread Prev][Thread Next][Date Index][Thread Index]

[hts-users:01975] Re: the DR command of HHEd


Hui LIANG (Tienshyang) wrote:

I'm still looking forward to knowing some details of the process of converting decision trees into a regression tree. My questions are as follows:

1. Not only decision tree structures but also questions are used when AU is carried out. What about DR? Are the questions in decision tree files used when DR is carried out?

No. Because they are not required. What HTK wants to know while using a regression class tree is a set of Gaussians associated to each leaf node of regression tree. This regression class tree is used in bottom-up fashion rather than top-down, questions are not required.

2. There are a lot of class definitions in a *.base file, for instance, "<CLASS> 27 {a~t-lw+iy=dh.state[6].stream[5].mix[1],r~d-lw+iy=g.state[6].stream[5].mix[1]}". How does HTS choose such full-context labels for each regression class? The two full-context labels are shortened for better legibility of this email.

The first one appeared while converting decision tree to regression tree. Please refer to HHEd.c:PrintNodeBaseClass() for details.

3. By the way, I want to confirm one more thing: the two steps, constructing a regression tree (by DR or RC) and adding unseen models of adaptation data, can be swapped, can't they?

Yes and no. If your full_all.list includes all full context models in full.list, (I think) you can swap them. Otherwise, not.


Heiga ZEN (Byung Ha CHUN)

Heiga ZEN (Byung Ha CHUN)
Speech Technology Group
Cambridge Research Lab
Toshiba Research Europe
phone: +44 1223 436975

This email has been scanned by the MessageLabs Email Security System.
For more information please visit http://www.messagelabs.com/email ______________________________________________________________________
[hts-users:01974] Re: the DR command of HHEd, Hui LIANG (Tienshyang)