[hts-users:03710] Re: about HERest mixture weight flooring option

["Heiga ZEN (Byung Ha CHUN)" <heigazen@xxxxxxxxxx>]

Hi,

2013/4/16 Sebastien Le Maguer <Sebastien.Le_maguer@xxxxxxxx>:

> in the reestimation stages of the Training.pl script, I've found that the mixture weight flooring
> option is called twice. The first one in the $HERest{'ful'} definition (Training.pl, line 213). The
> second one for each full model reestimation stage in the "$opt" definition (Training.pl, lines 471
> and 544) where it is set to 0.0. Is to to disable the mixture weight flooring for these particular
> stages? If so, is there a specific reason to do this?

This is done just before the clustering stage.  In the decision
tree-based context clustering process, statistics (occupancy counts,
1st and 2nd-order statistics) are first recovered from the stats file,
means and variances (please check Odell's PhD thesis for details).  If
you use flooring (both mixture weights or variances), recovered
statistics are somewhat biased; you cannot recover true, original
accumulated statistics once you perform flooring.  To avoid this
problem, the HTS training script does not perform mixture-weight and
variance flooring just before the tree-based context clustering stage.

I hope this explanation makes sense.

Heiga

--
Heiga ZEN (in Japanese)
Byung Ha CHUN (in Korean)
<heigazen@xxxxxxxxxx>