[hts-users:03618] Re: dynamic window specification

[那兴宇 <nxy-yzqs@xxxxxxx>]

Hi Matt,

Thank you for info.
> We can see this is we compare the synthesis results for a standard
> window such as "1.0 -2.0 1.0" versus a very long one such as "0.0 0.0
> 0.0 0.0 0.0 0.0 0.0 0.0 1.0 -2.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0".
> The trajectories synthesized by HMGenS differ slightly towards the very
> start and very end of the utterance.
I didn't try that long window, but I agree with you on this.
What I am interested is, when you use a real long window of the history, is it comparable with autoregressive HMM?

> Personally I have altered my HTS code to use -1e10 as a special value
> distinct from 0.0 in the window definition files to allow using windows
> of truly arbitrary sizes.
Do you mean -1e10 or 1e-10?

Xingyu

At 2013-02-15 21:03:09,"Matt Shannon" <sms46@xxxxxxxxx> wrote:

>Hi,
>
>Adding a zero to a window in this way gives an almost identical result,
>which is the same for all practical purposes.  However it is not
>strictly speaking identical.
>
>In fact there is a small bug in Calc_WUW_and_WUM (see
>http://hts.sp.nitech.ac.jp/hts-users/spool/2009/msg00024.html) which is
>still present in HTS-2.3alpha which may cause problems for windows of
>even length, so it is definitely a good idea to pad with a zero to make
>your window of odd length.
>
>Matt Shannon
>
>
>On 14/02/13 17:25, Heiga ZEN (Byung Ha CHUN) wrote:
>> They are the same.  However, to avoid confusions, I recommend you to use
>> 
>> 5 1.0 1.0 [-2.0] 1.0 0.0
>> 
>> for this case.
>> 
>> I personally think that it was my mistake to introduce the following code.
>> 
>>       if (fsize % 2 == 0)
>>           win->r_width[i]--;
>> 
>> Regards,
>> 
>> Heiga
>> 
>> 
>> 2013/2/14 那兴宇 <nxy-yzqs@xxxxxxx>:
>>> Hi all,
>>>
>>> In current MLPG implementations, either HMGenS or hts_engine_API, the default dynamic window is symmetrical.
>>> But I wonder what would happen if I set unsymmetrical window.
>>> So I checked the HGen(in HTS-2.2) and HTS_model(in hts_engine_API-1.06), found similar codes below;
>>> +     /* set pointer */
>>> +     length = fsize / 2;
>>> +     win->coefficient[i] += length;
>>> +     win->l_width[i] = -length;
>>> +     win->r_width[i] = length;
>>> +     if (fsize % 2 == 0)
>>> +         win->r_width[i]--;
>>>
>>> I suppose it means I can set unsymmetrical window like
>>> +   4   1.0   1.0   [-2.0]   1.0
>>> with  '-2.0' be the coeff. for current frame, right?
>>> But what if I set
>>> +   5   1.0   1.0   [-2.0]   1.0   0.0
>>> as unsymmetrical window in a symmetrical form? Is it the same?
>>> Please drop in any opinion.
>>> Thank you.
>>>
>>> Best,
>>> --
>>> Xingyu Na (那兴宇)
>>> http://www.idiap.ch/~xna/
>>> Beijing Institute of Technology [naxy(at)bit.edu.cn]
>>> Idiap Research Institute [xingyu.na(at)idiap.ch]
>>>
>> 
>> 
>> 
>> --
>> Heiga ZEN (in Japanese)
>> Byung Ha CHUN (in Korean)
>> <heigazen@xxxxxxxxxx>
>> 
>