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

[hts-users:03792] input and parent transforms


Hi,

I am building a cascade adaptation system using CMLLR. "xform1" is used as the parent transform in the estimation of "xform2". After three iterations, I got "xform2.1", "xform2.2" and "xform2.3". In the estimation of "xform2.1", "xform1" is used as both input and parent transform. In the estimation of "xform2.2", "xform2.1" is used as input transform and "xform1" is used as parent transform, etc. This is system A.

Then I get system B, which doesn't have parent transform. But "xform1" is used as the input transform in the estimation of "xform2.1". What I found in the log is that, the average log prob of the adaptation data of system B is higher then that of system A, which has parent transform. What does this imply? The data that used to estimate "xform1" and "xform2" are from the same speaker in the same language.

Please let me if the procedure is correct or not.
--
Xingyu Na (那兴宇)
http://www.idiap.ch/~xna/
Beijing Institute of Technology [naxy(at)bit.edu.cn]
Idiap Research Institute [xingyu.na(at)idiap.ch]