Improved speech recognition word lattice translation by confidence measure

In conventional speech translation systems, Automatic Speech Recognition (ASR) produces a single hypothesis which is then translated by the SMT system. The translation results of SMT system are impaired by the word errors of the first best hypothesis in this approach more or less. To improve speech translation, we use a new word lattice translation approach which integrates multiple information sources from the speech recognition word lattice to discount the misrecognition. Furthermore, in order to improve speech translation and to reduce computation, we used N-bests cutoff, merging of identical word ids, and confidence measure. Experiments of Japanese-to-English speech translation showed that the proposed word lattice translation outperforms the conventional single best method.