ABSTRACT
According to [Jia 99], the Bayesian predictive classification (BPC) decision rule makes a speech recognizer minimize the overall recognition error when the expectation is taken with respect to the model parameter uncertainty described by its prior PDF. Assuming that the functional form of the parameter transformation is exactly known and that all available information about the transformation parameters is completely contained in the prior PDF p(,), given unknown observation o, then we can express such an optimal recognition result W as
arg max Pr(W) • max f Pr(o, s, /, I Tn(A,), W) • p(,) dn, (12.78) W 3,1 ?I
where 5,1 denote the HMM state sequence and the Gaussian component label sequence, respectively. In the above equation, the Viterbi approximation [Jia 99] has been adopted to make the integral tractable. This approximate BPC is called Viterbi BPC or VBPC. Based on the work in [Jia 99], where prior PDF's are defined for all HMM parameters and the integral is taken with respect to the HMM parameters, we now introduce a new transformation-based structure constraint into the BPC method. Therefore, Eq. 12.78 can be thought as a kind of "constraint-based" VBPC.
