[hts-users:01716] Re: a variance of duration model

[Alexis Moinet <alexis.moinet@xxxxxxxxxx>]

As you say, those are variance values. They are used only if you ask hts 
to stretch the sound (option -r of hts_engine). Otherwise durations are 
computed according to mean values.

Bigger values for silences simply means that silence durations will 
change more than phoneme durations when you use -r option. (aka 
non-uniform duration-stretch)

Alexis


ILHWAN KIM wrote :
> Hi. 
>
> I experiment HMM-based Korean speech synthesis system using HTS 2.01 version and I study about state-duration model control. 
>
> When I check the mean and variance of duration model, I find a special case. 
>
> only,  5th state`s variance of silence model have a number of frames, (400~ 800 frames = 2~4sec) 
>
> I check traning speech DB, but I don`t find a long silence such a 2~4 sec 
>
> why?   Is it bugs? 
>