[hts-users:01359] Re: State Duration HVite Decode/Recognize

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

Hi,

Moritz Beyreuther wrote (2008/04/26 0:44):

> I am working on earthquake recognition using HMM. Therefore I created my
> own feature-vectors and then use HTK for training and recognition. For an
> two class ("reg" and "noise") example I use a pass of:
>
> HInit -S traindata.txt -T 1 -i 20 -M hmm0 -H proto/reg reg 
> HRest -S traindata.txt -T 1 -i 60 -M hmm1 -H hmm0/reg reg
> ... (also for noise, 5 iterations)
> HVite -T 1 -H hmm5/reg -H hmm5/noise -w net.slf dict.txt def/hmmlist.txt
> file_to_recognize
>
> I was looking for an easy way to include state duration in the models.
> That is where I came across the HTS website. Thus I installed HTS and
> tried the above example with state duration in HTS. There is the option -g
> for HRest or HERest to output a duration model; however, I could not find
> an option for HVite or HDecode to include the duration model produced by
> HERest or HVite for the recognition.
>
> My question is:
> - Is it possible to decode/recognize using HTS and state duration models?

It has been discussed in this mailing list.
Please see the thread hts-users:01358 on http://hts.sp.nitech.ac.jp/hts-users/spool/2008/threads.html

> - So far I could only run one iteration of HRest and HERest, is it possible
>   or useful to do it iterative (as in the above example)?

HERest supports iterative reestimation.
You can find how to perform this via HTS-demo.

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