[hts-users:03172] Re: Should we enable MGE and GV at the same time?

["Heiga ZEN (Byung Ha CHUN)" <heigazen@xxxxxxxxxx>]

Hi,

2012/2/17 Yuan-Fu Liao <yfliao@xxxxxxxxxxx>:

> Should we combine MGE training (using HMgeTool) and GV parameter generation
> (option in HMGens)? I'm wondering if GV will distort the effect of MGE
> training or not.

For me, using the MGE-trained model with the current implementation of GV in HMGenS sounds wrong, as there is a severe mismatch in their objective functions:

MGE: minimize \sum_n (c_n - ^c_n(q, \lambda))^2 w.r.t. \lambda

GV: maximize \log N(Wc; \mu_q, \Sigma_q) + \alpha \log N(v(c); \mu_gv, \Sigma_gv) w.r.t. c

The MGE-trained model will not give proper N(Wc; \mu_q, \Sigma_q) as its parameters are stimated so as to minimize the Euclidean distance rather than likelihood.

Regards,

Heiga

--
Heiga ZEN (in Japanese)
Byung Ha CHUN (in Korean)
<heigazen@xxxxxxxxxx>