[hts-users:00280] Re: about dynamic f0

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

Hi,

liulei_198216@xxxxxxxxxxx wrote:

> for the mel parameter O={O_1,..O_t...O_T} ,there is delta for every O_t.

Yes.

> And also there are some dynamic feature for  f0.
> I mean that can the parameter generation algorithms get them?

It's yes and no.
We can calculate dynamic features within voiced regions.
However, we cannot compute them on the boundaries of voiced and unvoiced regions because of dynamic feature window.
To address this problem, we set variance values of dynamic features to infinity (inverse variance -> 0) as follows:

   for (s=1;s<=f0pst.vSize;s++) {
      f0pst.sm.mseq[pt][s]  = hmm->svec[j].info->pdf[s+1].info->spdf.cpdf[1].mpdf->mean[1];
      f0pst.sm.ivseq[pt][s] = (bound && s!=1) ? 0.0 : hmm->svec[j].info->pdf[s+1].info->spdf.cpdf[1].mpdf->cov.var[1];
   }

It means that statistics of dynamic features are not used during the parameter generation process.
As a result, f0 values on the voiced/unvoiced boundaries are determined so as to maximize their output probabilities 
according not only to statistics of neighboring voiced frames but also to their statistics of static features.

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://kt-lab.ics.nitech.ac.jp/~zen
------------------------------------------------