[hts-users:01335] Re: How to tie duration length of sil and sp

["Nicholas Volk" <nvolk@xxxxxxxxxx>]

Hi,

I would do this by modifying the HHEd.c file.
First force a silence node as the start node of your tree (forcing
questions has been discussed on the forum a bit earlier), then force that
branch to use a meaningless dummy question, so that that branch will not
keep branching...

I have modified ValidProbNode() to

void ValidProbNode(Node *node, const double thresh, const char
*question_name)

where question_name (if not NULL) will be used.

Then in BuildTree() something like

/* for dur root */
ValidProbNode(node,threshold,"is_pau");

/* inside the loop: */
if ( macRoot[0] == 'd' && node == tree->root ) { /* d for dur */
  ValidProbNode(node->yes,threshold,"dummy_pause_question");
}
else {
  ValidProbNode(node->yes,threshold,NULL);
}

Note that these are only guidelines, not a full solution.

br,
  Nicholas


> Hi, all:
>
> I trained a system with HTS, and find that there are a lot of duration
> models for states of 'sil', this waste many tree nodes and pdfs to store
> them. This is
> because my wav files has silence parts in the begining and ending, and the
> length are variable. Obviously, making complicated silence duration model
> is useless.
>
> So I want to tie the states of 'sil', making the duration fixed. But I'm
> not familiar with HTK, could anyone give me a hint how and where to write
> the hhed command?
>
> Another question: I want to limit the variance floor for the mgc, for
> example to 1e-4, how could I do this? Although the config.pm file has
> variance limit
> statement:
>   %vflr=('mgc'=>'000",...
> the trained pdf for mgc still has small variances:
> 0.000236, 0.000629, 0.000187...
>
> Thanks
>
>
>
>