[hts-users:00690] Re: HTS Training

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

Hi,

Carlos Mendes wrote (2007/05/29 21:31):

> About the context dependent trees.... yes... I didn't explained myself 
> clear .... or maybe I'm not seeing this correctly, :), I'm relatively new to 
> HMM based synthesis, so sorry if I'm saying something really stupid. But as 
> I see it .... usually units are clustered according to their context 
> information, generally according to their placement in the phrase, word, 
> syllable, triphone, diphone etc... and other signal features like mcep, 
> etc....(at least in unit selection algorithms) depends on what you want. In 
> HMM synthesis, HMM's are similarly clustered in context trees ...  
> correct...?

Yes.

> I would like to do some analysis on the units that are being used on the 
> training, meaning their information relative to the assigned cluster or node. 
> For this, I need to know were exactly HTK placed a certain unit, during the 
> training....
> I could use the questions file and then apply it to the corpus, but I would 
> rather prefer to use the information that HTK can give me. So the question 
> would be .... Is it possible to obtain this information during the training?

HHEd outputs lists of models which are clustered to each leaf node by turning on the trace flag T_CLUSTERS.
You can turn on this flag as 

% HHEd -A -B -C trn.cnf -D -T 0x0101 -p -i ... cxc.hed full.list
                              ^^^^^^
According to this information, I think you can know segments corresponding to each leaf node.
Note that this output is so hard to read :-(

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