[Subject Prev][Subject Next][Thread Prev][Thread Next][Date Index][Thread Index]

[hts-users:00372] Re: Questions about the latest released HTS


Hi

Heng Lu wrote:

Question 1:
I have used both HTS2.0-RC1 and HTS2.0-RC2 to synthesize Chinese sentences, and
have found that although there may be lots of improvements and fixs in the code of
RC2 , the quality of the synthesized sentences using RC2 remains nearly the same
as the sentences from RC1. Is there anything wrong ?

Modifications made in RC2 are mainly bug fixes (segmentation fault, memory leaks, etc).
Two critical bugs in RC1 were voiced/unvoiced handling and EM-based parameter generation in HMGenS.
First one affects only on voiced/unvoiced boundaries, so its effect may be small.
Another one affects entire speech but I have found that its effect has been small too.

Another question: Compared with the synthesizing without EM algorithm(HMGens c 0) , synthesizing with EM(HMGens c 1) seems to make the quality of the synthesized speech decrease .
And I am really puzzled about it .

This algorithm iteratively maximizes the output probability with respect to the speech parameter sequence via the EM algorithm. As you know, increasing the output probability does not guarantee improving the speech quality.
However, in Dr. Tokuda's ICASSP 2000 paper, improvements were reported.
So now I'm wondering if the current implementation still contains some bugs (therefore it is still a Release Candidate). Everyone, please look at it and help me to find bugs in the current release!!

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://kt-lab.ics.nitech.ac.jp/~zen
------------------------------------------------


References
[hts-users:00371] Questions about the latest released HTS, Heng Lu