Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

Maximum a Posterior Linear Regression with Elliptically Symmetric Matrix Variate Priors

Wu Chou

Bell-Labs - Lucent Technologies, Murray Hill, NJ, USA

In this paper, elliptic symmetric matrix variate distribution is proposed as the prior distribution for maximum a posterior linear regression (MAPLR) based model adaptation. The exact close form solution of MAPLR with elliptically symmetric matrix variate priors is obtained. The effects of the proposed prior in MAPLR are characterized and compared with conventional maximum likelihood linear regression (MLLR). The proposed priors are significant informative priors, through which a well-founded Bayesian theoretical framework is formulated to incorporate prior information in model adaptation. Moreover, an efficient approach of hyperparameter estimation in MAPLR is described. Experimental results indicate that significant gain can be obtained when adaptation data are sparse.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Chou, Wu (1999): "Maximum a posterior linear regression with elliptically symmetric matrix variate priors", In EUROSPEECH'99, 1-4.