We present a simplified derivation of the extended Baum-Welchfl

procedure, which shows that it can be used for Maximum Mutualfl

Information (MMI) of a large class of continuous emissionfl

density hidden Markov models (HMMs). We use the extendedfl

Baum-Welch procedure for discriminative estimation offl

MLLR-type speaker adaptation transformations. The resultingfl

adaptation procedure, termed Conditional Maximum Likelihoodfl

Linear Regression (CMLLR), is used successfully for supervisedfl

and unsupervised adaptation tasks on the Switchboardfl

corpus, yielding an improvement over MLLR. The interactionfl

of unsupervised CMLLR with segmental minimum Bayes riskfl

lattice voting procedures is also explored, showing that the twofl

procedures are complimentary.

PDF file

In: Proc. of the Eurospeech Conference

Details