[hts-users:01102] Re: HTS-2.1beta

[Junichi Yamagishi <jyamagis@xxxxxxxxxxxx>]

Hi,

On 2008/01/15, at 12:07, Junichi Yamagishi wrote:

> > > HTS-2.1beta:
> > [...]
> > > - Random generation of state transitions, state durations, and  
> > > mixture components.
> >
> > What does this random generation refer to? Initialization
> > of the models, or some alternative parameter generation logic?
> > Is that described in a particular article?
>
> Probably, there is no particular articles, I guess.
>
> Random generation[sampling] of Gaussian distributions for  
> Trajectory HMM (see chapter 5)
> http://www.sp.nitech.ac.jp/~zen/english/index.php? 
> plugin=attach&refer=Publications%2FInternational% 
> 20conferences&openfile=tokuda-beyondHMM.pdf
> has already supported in the previous version 2.0.
>
> This extension newly supports random sampling of state transitions,  
> state durations, and/or
> mixture components in the parameter generation algorithm using the  
> Trajectory HMMs.
>
> This function would be good when you want to randomly change  
> synthesized trajectory or
> duration for the synthesized trajectory somewhat every time.
>
> Please imagine human-motion synthesis. If the generated motion  
> trajectory is
> always completely the same, people would think it is unnatural.
> The above function can add some random effect to the generated  
> trajectory.

Another example would be a dialogue system.

Synthetic speech for the dialogue systems is very monotone
because they output the same speech for the same input sentences.

By using the above random generation, you can change e.g. prosody or  
tone
of synthetic speech every time even for the same sentences/inputs.