[hts-users:00007] japanese demo

[Nobuaki MINEMATSU <mine@xxxxxxxxxxxxxxxxxxxx>]

Dear hts-users,

I downloaded HTS_1.1.1 and Japanese speech samples, then built HMMs
for nit_m001. Demo script automatically generated synthetic speech
samples of j01 to j53.

I compared the nit_m001 synthesized speech samples that we made
with those on the web, which I think were created by NIT people.
(http://kt-lab.ics.nitech.ac.jp/~demo/gtalk/demo.php)

I found that there is acoustic difference between m001 we made and m001
you (NIT people) made. The m001 we made sounds more like buzzer.
# I know that the two methods use different linguistic information
  for synthesis. The acoustic difference in the synthesized speech
  might be due to the linguistic difference...

I'd like to know whether I did a correct thing or not. The questions
are follows.

1. Demo script of m001 generates speech samples less natural than m001
   on the web.  Right ?

2. If so, would you let us know some tips to improve the naturalness ?
   Cepstrum dimensions, generation of source, linguistic information in
   labels... and so on.

Best wishes;

nobuaki minematsu.

##===================================================================##
 | Nobuaki MINEMATSU, Associate Professor                            |
 | Department of Information and Communication Engineering           |
 | School of Information Science and Technology, University of Tokyo |
 | E-mail : mine@xxxxxxxxxxxxxxxxxxxx, TEL/FAX : +81-3-5841-6662     |
##===================================================================##