ABSTRACT

Note that the MAP estimates of the (hyper-) parameters, 4,, 1VIi , 4i, and Pi, for the prior distributions given by Eqs. 12.52-12.55 are actually obtained only in some intermediate stages of the M step in the EM algorithm. The objective of the M step is to estimate the actual parameters (i.e., not hyper-parameters) of the trended HMM — 13, and 61. In the implementation of the MAP-based speaker-adaptive speech recognition system, Eqs. 12.52-12.57 are programmed as such, which constitutes a full iteration of the MAP adaptive training procedure.