Thank you very much for out help. I read the GV paper and looks awesome.
I set the GVWEIGHT to 0.5 but now the voice a very small amplitude :)
I'll keep tunning this. It is a pity that the whole training procedure
has to be re-run again when you change this setting.
Thanks again.
--
Jordi Adell
Nicholas Volk escribió:
Just reduce the GV factor from 1.0 to an appropriate level for your
voice,
say 0.5. (Or set it 0.0 and use beta with 0.3 or something like that.)
(Using global variance usually improves the overall quality a lot IMHO.)
br,
Nicholas
Dear all,
I'm a newbie in using HTS. I've been working with HTS for some months
and finally I got some audio files using label information generated by
our own prosodic model. However, the audio files are clipped. You can
guess the expected sentence under the noise, although. Even the alice
files that the demo script generates for testing are clipped. Many
values go further than the integer maximum.
I assumed that *raw files in hts are 16000 Khz and 16 bits signed
integer.
I'm running HTK 3.4 and HTS 2.1 and HTS engine 1.0
in a 64 bits little endian machine with windows server 2003
and cygwin.
Regards
--
----------------------------------------------
Jordi Adell
Signal Processing Section
Communications and Signal Theory Department
Universitat Ramon Llull
Passeig Bonanova 8 08022 - Barcelona (Spain)
tlf: +34 932902452 (ext.277)
----------------------------------------------
--
----------------------------------------------
Jordi Adell
Signal Processing Section
Communications and Signal Theory Department
Passeig Bonanova 8 08022 - Barcelona
tlf: +34 932902452 (ext.277)
----------------------------------------------