[Subject Prev][Subject Next][Thread Prev][Thread Next][Date Index][Thread Index]

[hts-users:02605] Re: About the parameter generation algorithm


Hi, Heiga

I read the source code in HGen.c and I found that the coefficient is calculated one dimension after another even if the covariance matrix is full.

That's because the Calc_WUW_and_WUM, Cholesky Decompesition, Forward and Backward substitution algorithms are called inside the loop M ( vector length), and the covariance matrix is check full or diag. inside the Calc_WUW_and_WUM function. Therefore, the waveform synthesis stage still have to wait for the whole sentence feature vector sequences generated completely. Otherwise, the filter parameter of first frame will not be available.
Is that right?

Na Xingyu

At 2010-08-03 00:14:46,"Heiga ZEN (Byung Ha CHUN)" <heiga.zen@xxxxxxxxxxxxxxxxx> wrote: >Hi, > >(02/08/2010 14:53), 那兴宇 wrote: > >> I am reading the source codes of hts_engine and I met up with this question. >> Why does the numerical solution of WUW*C=WUM work in the order of time length instead of vector length? >> Does the decomposition and substitution method work out that way? > >Because hts_engine assumes diagonal covariance matrices, we can perform the decomposition and >substitution for each dimension separately.  So WUW and WUM work in the order of time length (T) >instead of vector length (M*T).  This is more memory efficient than solving all dimensions >simultaneously. > >HMGenS supports both diagonal & full covariance matrices.  If you use HMGenS with full covariance >matrices, WUW and WUM work in the order of vector length. > >Regards, > >Heiga ZEN (Byung Ha CHUN) > >====================================================================== >NOTE: The information in this email and any attachments may be >confidential and/or legally privileged. This message may be read, copied >and used only by the intended recipient. If you are not the intended >recipient, please destroy this message, delete any copies held on your >system and notify the sender immediately. > >Toshiba Research Europe Limited, registered in England and Wales (2519556). >Registered Office: 208 Cambridge Science Park, Milton Road, Cambridge >CB4 0GZ, England. Web: http://www.toshiba-europe.com/research/crl >====================================================================== > > >______________________________________________________________________ >This email has been scanned by the MessageLabs Email Security System. >For more information please visit http://www.messagelabs.com/email  >______________________________________________________________________ >



您想拥有和网易免费邮箱一样强大的软件吗?

References
[hts-users:02572] Re: About the parameter generation algorithm, Heiga ZEN (Byung Ha CHUN)
[hts-users:02571] About the parameter generation algorithm, 那兴宇