[hts-users:01818] Re: HTS voice built with Festvox 2.1 - HHed failure during duration model clustering
Hi,
Daniel Tihelka wrote:
first of all I have found small bug on 64-bit platform: in
src/hts_build/data/mkdata.pl (written by Heiga Zen), line 114, the format of
first two items stored into tmp.head file are formatted by -al in command
x2x. It will write two long numbers - in 32-bit platform they are 4B long, as
expected by HTK. However, when SPTK (I use 3.1 version) is compiled at 64bit
platform, the numbers are 8B long and HTK (also 64bit version) cannot of
course read it. Quick workaround is to use -ai switch; it should work on both
32 and 64bit.
It doesn7t happen in the current HTS-demo. The HTS-demo no longer uses
SPTK to append HTK headers. It uses addhtkheader.pl to append this.
You can find the following statement in addhtkheader.pl:
# number of frames in long
$NFRAME = pack("l", $nframe);
it packs number of frames in "l". In Perl, packing a value in "l"
peforms packing the value in "signed 32-bit interger value". So it
doesn't depend on 32/64-bit.
What troubles me, however, is that Training.pl script from src/hts_build/ does
not build HTS voice correctly (checked on CMU ARCTIC AWB 0.90 and CMU US SLT
ARCTIC 0.95, on both the same type of failure occurs). No changes were made
both in scripts and voice!
Festvox's training script is out of date. I wrote it about 6 years ago.
I don't recommend you to use that. It cannot use the latest
technologies and fixes.
What may be wrong? I do not expect a problem be a 64bit platform. All scripts
and voice files were also used without any change (except -x switch in the
build script and -ai switch for x2x, which, I think, are unlikely to cause
the error). And the awb and slt HTS voices are available for Festival (and
working), so they were build somehow - and I suppose that by build_hts
script. Is the script up-to-date (the last date there is May 2003)? Or is
there another preferred way how to build HTS voice for Festival?
Please do not use training script included in Festvox.
Any advice will be really appreciated. May I give you any additional
information?
Please use the latest one released on the HTS website.
Regards,
Heiga ZEN (Byung Ha CHUN)
--
--------------------------
Heiga ZEN (Byung Ha CHUN)
Speech Technology Group
Cambridge Research Lab
Toshiba Research Europe
phone: +44 1223 436975
______________________________________________________________________
This email has been scanned by the MessageLabs Email Security System.
For more information please visit http://www.messagelabs.com/email
______________________________________________________________________
- Follow-Ups
-
- [hts-users:01819] Re: HTS voice built with Festvox 2.1 - HHed failure during duration model clustering, Esther Klabbers
- References
-
- [hts-users:01817] HTS voice built with Festvox 2.1 - HHed failure during duration model clustering, Daniel Tihelka