Hi,
all,
I probably find a bug in HSMM-based speech synthesis system of version 2.2: some variances of state duration model are negative. The finding reals that the current implementation of HTS version 2.2 is not fully correct in a strictly mathematical sense when dealing with state duration modelling.
Although it is feasible to apply variance flooring technique to the problematic variances, as suggested by Prof. Keiichiro Oura through previous private discussion, I think there should be other more rigorous and elegant solution, which I have not figured out until present.
The aim of this report is to let us notice this fact, and to attract attention to perfect the implementation of HTS.
This report is directly available from http://www.escience.cn/system/file?fileId=70728
It can also be found on my homepage http://www.escience.cn/people/yangwang
Any comments/corrections are highly welcome.
Thanks a lot!
Yang Wang