[hts-users:00026] Re: problems with HTS learning

[Heiga Zen / Byung-Ha Chun <zen@xxxxxxxxxxxxxxxx>]

Hi Osowski, 

On Tue, 06 Apr 2004 15:20:23 +0200

> /home/luke/mowa/festival/voices/polish/HTS-demo_IVO-PL-JL$ 
> ../../english/HTS-demo/scripts/Training_ivo_pl_jl.pl > train1.log
>   ERROR [+2126]  ViterbiAlign: No path found in 1'th segment
>  FATAL ERROR - Terminating program 

This error means that there are no segments longer than the number of HMM states.
And I think that the name of this model, "#", isn't valid in the HTK.
Is there any segments for model "#" longer than 5 frames in your training data ? 

> There is also other thing - When i start learning from 1 phase i've this 
> error, but when i'm trying to start learning from 3 phase (in which this 
> error occures, from scripts/ Train....) i had other error - i don't 
> understand why. Effects of finished 1 and 2 stage should be written to 
> files, and it should be possible to run 3 phase dirrectly.

In 1st phase, global variance is calculated to determine the variance flooring value.
In 2nd phase, context-independent models (monophone HMMs) are trained 
by segmental k-means algorithm (HInit) and EM algorithm (HRest) with segment boundaries.
In 3rd phase, the context-dependent models (fullcontext HMMs) are generated 
by copying the context-independent models which trained in 2nd phase.
3rd phase requires the result of 2nd phase.
Therefore, you cannot start this training script from 3rd phase directly.

Best regards,

Heiga Zen / Byung-Ha Chun

-- 
 ------------------------------------------------
  Heiga Zen     (in Japanese pronunciation)
  Byung-Ha Chun (in Korean pronunciation)

  Department of Computer Science and Engineering
  Graduate School of Engineering
  Nagoya Institute of Technology
  Japan

  e-mail: zen@xxxxxxxxxxxxxxxx
     web: http://kt-lab.ics.nitech.ac.jp/~zen
 ------------------------------------------------