International Symposium on Chinese Spoken Language Processing
August 23-24, 2002
A Comparative Study of Quickprop and GPD Optimization Algorithms for MCELR Adaptation of CDHMM Parameters
Jian Wu, Qiang Huo
Department of Computer Science and Information Systems,
The University of Hong Kong, Hong Kong
In our previous work, we have presented an approach of minimum
classification error linear regression (MCELR) for adaptation of
Gaussian mixture continuous density HMM (CDHMM) parameters.
It is shown that a stochastic approximation approach known
as the GPD (generalized probabilistic descent) can be used to optimize
the MCE objective function. However, it is relatively diffi-
cult to set an appropriate value for the learning control parameter
to achieve a fast yet stable GPD optimization process. In this paper,
we study another batch-mode approximate second-order optimization
approach, namely Quickprop, aiming at speeding up the
convergence of the objective function of MCELR while making
the learning more robust. It is demonstrated by a series of experiments
for supervised speaker adaptation that Quickprop is a better
alternative to GPD for MCELR.
Wu, Jian / Huo, Qiang (2002):
"A comparative study of quickprop and GPD optimization algorithms for MCELR adaptation of CDHMM parameters",
In ISCSLP 2002, paper 23.