[hts-users:00396] Re: Generation using phonemes duration in HTS2

[Ossama Abdel-Hamid <ossama_a@xxxxxxxxx>]

Hi,

I have found the cause of the problem and solved it, it seems to be a 
bug in HTS-2.0RC2.
Here are the errors in the code and corrections:
in the file HMGenS in the function CountDurStat, the line 291 it's written:
     *sqr += ivar[j];
It should be:
   *sqr += 1.0f / ivar[j];

and in line 336 it's written:
   rho = (SpeakRate*sum-sum)*sqr;
It should be:
   rho = (SpeakRate*sum-sum)/sqr;

And also in line 356
   rho = (modeldur-sum)*sqr;
IT should be:
   rho = (modeldur-sum)/sqr;

Best Regards,
Ossama


Ossama AbdelHamid wrote:
> Hi,
>
> First, thanks for your previous replies.
> I have a problem, that in HTS2 when I use HMGens with the option
> -m, and put the desired phonemes durations in the label file, it
> generates all states having only one frame, although it worked
> well with hts1.1.1 (using -g option ).
>
> Regards,
> Ossama
> ---------------------------------------------
> Free POP3 Email from www.Gawab.com 
> Sign up NOW and get your account @gawab.com!!
>
>