[hts-users:00872] Re: demos problems

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

Hi,

Thomas Drugman wrote (2007/10/22 17:42):

> It seems that, though it says it uses MGCEP, 
> it actually uses MCEP since the command in data/Makefile is $MCEP ... and $MCEP = .../SPTK/bin/mcep. 
> Shouldn't it be $MGCEP with $MGCEP = .../SPTK/bin/mgcep so as to use the -g option with non-zero gamma parameter?

In configure.ac, you can find the following lines:


# mgcep
if test -z "$GAMMA"; then
       AC_PATH_PROG([MGCEP], [mcep], [:], [${sptk_search_path}:$PATH])
elif test $GAMMA -eq 0; then
       AC_PATH_PROG([MGCEP], [mcep], [:], [${sptk_search_path}:$PATH])
else
       AC_PATH_PROG([MGCEP], [mgcep], [:], [${sptk_search_path}:$PATH])
...


So "mcep" is used in HTS-demo only if the pole/zero weight parameter, GAMMA, is equal to 0.
"mcep" and "mgcep -g 0" should behave exactly the same, but they don't.
Because spectral parameters are optimized in cepstral domain in "mcep" but they are optimized in MGLSA filter parameter domain in "mgcep".
We observed that "mcep" can extract spectral envelopes more robustly than "mgcep -g 0".
Therefore, we use "mcep" rather than "mgcep -g 0" in HTS-demo if GAMMA=0.

> I also encountered some problems in HTS-demo_CMU-ARTIC-SLT:
> the generated speech is too fast because trained durations are too small (about 1 or 2 frames by HMM state), but I can't fix why yet..

I cannot say anything from the above information.

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
------------------------------------------------