[hts-users:03926] Re: I have a question about acceleration coefficients.

[Matt Shannon <sms46@xxxxxxxxx>]

Hi Hea,

Yes, typically an HTS system (e.g. the HTS demo) uses the window 
coefficients (up to an arbitrary scale factor):

-1 0 1 (delta)
1 -2 1 (delta-delta)

where the centre coefficient is the one for the current frame.  I 
believe the HTK default is to use the window coefficients (up to an 
arbitrary scale factor):

-2 -1 0 1 2 (delta)
4 4 1 -4 -10 -4 1 4 4 (delta-delta)

Presumably the HTS-style window coefficients were adopted for synthesis 
because they were found to give better synthesis performance.  In 
informal experiments I've conducted in the past the HTS-style 
coefficients did result in clearer-sounding speech, and had a bit better 
MCD score and *much* better predictive log probabilities than the 
HTK-style coefficients.

Matt


On 15/11/13 01:17, Hea Young Park wrote:
> Hi, Matt Shannon.
> Do you mean that it is different acceleration coefficients using HTK and
> acceleration coefficients using HTS?
>
>
> 2013/11/15 Matt Shannon <sms46@xxxxxxxxx <mailto:sms46@xxxxxxxxx>>
>
>     Hi Hea,
>
>     Formulae like the ones you presented are finite difference
>     approximations to the first and second derivative, and are exact for
>     polynomials of some order but inexact for polynomials of higher
>     order and for general non-polynomial functions.  I would say that
>     one formula is not really more "correct" than another, and it
>     depends on the application which one will lead to a better model.
>
>     The specific window coefficients you proposed probably won't work
>     very well for speech synthesis, though, since in your proposed
>     scheme the even-indexed frames and the odd-indexed frames are
>     independent, and so you're likely to get very discontinuous speech
>     parameter trajectories.
>
>     Matt
>
>
>
>     On 11/11/13 00:53, Hea Young Park wrote:
>
>         Dear all.
>
>         I’m trying to build HTS with my custom voice using HTS-2.2 demo
>         script.
>
>         I have aquestion about acceleration coefficients.
>
>
>         According to HTS book equation (5.16),
>
>         The delta coefficients in time t-1, t, t+1 are computed as below.
>
>         본문 이미지 1
>
>         Just then, the acceleration coefficients are computed as below.
>
>         본문 이미지 2
>
>         However, according to /data/script/windows.pl
>         <http://windows.pl> <http://windows.pl> and
>
>         /data/win/*.win3 included in HTS-2.2 demo,
>
>         the acceleration coefficients are computed as below.
>
>         본문 이미지 3
>
>         So, If it’s correct thatI’m thinking,
>
>
>         /data/win/*.win3 file changed as below.
>
>         Current : 3 1.0 -2.0 1.0
>
>         New : 5 0.25 0.0 -0.5 0.0 0.25
>
>         I just wonder whether or not what I think is correct.
>
>         Please comment on my calculation mentioned above.