[hts-users:00575] Re: F0 contours become flat

["Xavi Gonzalvo" <gonzalvo@xxxxxxxxxxxxx>]

Hi all,

Let me add to this that we carried out some tests with Spanish Language
comparing the f0 contour from HTS 1.1.1 and the one obtained from a CBR
(Case Base Reasoning) algorithm. HMM contour were absolutely flatter than
the CBR one, especially when phrases were interrogative and short.
Furthermore, we concluded that a bigger tree doesn't definitely resolve the
problem of a flat f0 because it even produced undesired peaks in the f0
curve. 

The contextual factors used to cluster are quite important and so the lambda
must be chosen accordingly with the real effect of contextual features in
training data. As stated, a flat f0 also produced for us a stable
synthesized speech which I think is one of the positive points of this kind
of synthesis.

Greetings
Thank you

-----Mensaje original-----
De: Heiga ZEN (Byung Ha CHUN) [mailto:zen@xxxxxxxxxxxxxxxx] 
Enviado el: martes, 13 de marzo de 2007 13:34
Para: hts-users@xxxxxxxxxxxxxxxxxxxxxxxxx
Asunto: [hts-users:00574] F0 contours become flat

Hi all,

Today I attended a Japanese domestic conference and had a presentation about
HTS 2.0.
At that time, some people reported me that F0 contours generated from the
HMMs trained by HTS 2.0 are flatter than those generated from the HMMs
trained by HTS 1.1.1.

I think this is caused by variance flooring in tree-based clustering.
In the previous versions, we didn't care about variance flooring during
clustering procedure.
However, at the Blizzard Challenge 2005, we found that variance flooring in
clustering made synthesized speech more stable.
So we incorporated this into HTS 2.0.
But it makes tree a bit small and generated F0 contours flat.

If you want to avoid this problem, please set %lamb in scripts/Config.pm

%lamb = ('mcp' => "1.0",  # tree size control param. for MDL
         'lf0' => "1.0",
         'dur' => "1.0");

to less than 1.0.
This makes trees larger.

Thank you very much,

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