An improved minimum word error approachto lattice rescoring and system combination

In this paper we show how to combine confusion network (CN) and the Bayes Risk (BR) criterion to realize a close approximated Minimum Word Error (MWE) lattice decoding and system combination. We use the BR criterion as objective function that is optimized by means of selecting a best candidate as the consistent result from CN. To this end, an approximated BR criterion calculation algorithm is described and an approach to searching the optimal candidate sentence from CN is presented. The approximated BR criterion is implemented by performing recursive edit distance computation over word lattice, while the decoding result that is obtained by traversing CN satisfies the criterion minimization. Based on the experiments, it was observed that the proposed scheme achieved better results over the CN method in both single lattice rescoring and system combination work.