[hts-users:01212] Re: Problems in algorithm of parameter generation in HGen.c and HTS_mplg.c

["Heiga ZEN (Byung Ha CHUN)" <zen@xxxxxxxxxxxxxxx>]

Hi,

QHE wrote (2008/03/10 22:04):

> 1. In line 836 of HGen.c:
>
> pst->WUM[M*(t-1)+m] += ((double)pst->win.coef[d][-j]) * pst->mseq[t+j][d*pst->order+m+bias];
>
> there is no 'inverse variation' term to calculate WUM, while in line 85 of HTS_mplg.c:
>
> WU = pst->dw.coef[i][-j] * pst->sm.ivseq[t + j][i * pst->order + m];
> pst->sm.WUM[t] += WU * pst->sm.mseq[t + j][i * pst->order + m];
>
> there is the 'inverse variation' term sm.ivseq.
> Is the term ommited intentionally in HGen.c or just a mistake?

Inverse covariance matrices have already been multiplied to mean vectors in SetupPdfStreams() and UpdatePdfStreams() in HGen.c.

> 2. In line 849 of HGen.c:
> /* W'*U^{-1}*W */
> for (k=0; (WU!=0.0) && (k<pst->width) && (t+k<=pst->T); k++)
> if ((pst->win.width[d][WLEFT]<=k-j) && (k-j<=pst->win.width[d][WRIGHT]) && (M*k+n-m+1>0) && (pst->win.coef[d][k-j]!=0.0))
> pst->WUW[M*(t-1)+m][M*k+n-m+1] += WU*(double)pst->win.coef[d][k-j];
>
> and the line 88 of HTS_mplg.c:
>                for (k = 0; (k < pst->width) && (t + k <= pst->T); k++)
>
> We see k is from 0 to width (for exmple 0 to 4), but I think the k should be from win.width[d][WLEFT] to win.width[d][WRIGHT] (for exmple -2 to +2). 
> In the current code, I think only half of the window is applied when calculating WUW by WU * W.
>
> Am I right?

No.
Because WUW becomes (2*width+1)-band diagonal matrix and we store its upper-triangular part only.

Regards,

Heiga ZEN (Byung Ha CHUN)

--
------------------------------------------------
Heiga ZEN     (in Japanese pronunciation)
Byung Ha CHUN (in Korean pronunciation)

Department of Computer Science and Engineering
Nagoya Institute of Technology
Gokiso-cho, Showa-ku, Nagoya 466-8555 Japan

http://www.sp.nitech.ac.jp/~zen
------------------------------------------------