Minimum Word Error Rate Decoding

From the theory of Bayesian pattern recognition it is well known that the maximum a posteriori decision criterion yields a recogniser with the minimum probability of assigning the incorrect label to a pattern, if the correct probability distributions are used. This MAP criterion is also routinely employed in automatic speech recognition system. The problem addressed in this thesis is the fact that in a continuous speech recognition task the MAP criterion only guarantees to minimise the sentence error rate, but for many application the by far more relevant error metric is the word error rate. In the general framework of lattice rescoring an approach is investigated that tries to directly use the word error metric in the objective function that is the basis of the decision in the recogniser. An implementation of two variants of this approach is presented and its performance is evaluated on two very different continuous speech corpora. Closely related to this is the problem of replacing the a posteriori probabilities based on whole sentences that are used in conventional decoders by posterior probabilities on the word level. A new technique of estimating these word posteriors is presented and applied with the aim of improving the decoding result and also providing accurate confidence scores.