[hts-users:03442] Re: Issue with GV

[Veera Raghavendra <raghavendra@xxxxxxxxxx>]

I have tried many parameter tuning. But, the spikes are still present in the synthesized file.

The analysis shows that the spikes are happening when vowel phone is followed by nasal sound. Can I change any thing in the context dependent modeling to avoid these spikes?

 
Thanks,
Raghavendra.

On Fri, Oct 19, 2012 at 8:01 PM, Keiichiro Oura <uratec@xxxxxxxxxxxxxxx> wrote:

Hi,

You can try to change many settings for GV in the HTS demo scripts.

  MAXGVITER   maximum number of iterations of GV-based parameter
generation algorithm (default=50)
  GVEPSILON   convergence factor for GV iteration (default=0.0001)
  MINEUCNORM  minimum Euclid norm for GV iteration (default=0.01)
  STEPINIT    initial step size (default=1.0)
  STEPINC     step size acceleration factor (default=1.2)
  STEPDEC     step size deceleration factor (default=0.5)
  HMMWEIGHT   weight for HMM output prob. (default=1.0)
  GVWEIGHT    weight for GV output prob. (default=1.0)
  OPTKIND     optimization method (STEEPEST, NEWTON, or LBFGS) (default=NEWTON)
  NOSILGV     turn on GV without silent and pause phoneme (0:off or
1:on, default=1)
  CDGV        turn on context-dependent GV (0:off or 1:on, default=1)

Regards,
Keiichiro Oura


2012/10/19 Veera Raghavendra <raghavendra@xxxxxxxxxx>:

> Dear All,
>
> I found the difference in the quality with and without GV.
>
> If GV is turned-on, the voice quality is very clear but there are sudden
> spikes. These spikes destroys the synthesis quality.
>
> If GV is turned-off, there are no spikes but the content is not clear.
>
> Do I need to change any settings in HSMMAlign.
>
> Thanks,
> Raghavendra.