Improving speech-to-speech translation using word posterior probabilities
Vicente Alabau, Alberto Sanchís, Francisco Casacuberta · 2007
Nowadays, speech translation is a research problem in machine translation. The problem arises as to how to combine speech recognition and machine translation in a suitable way. Some authors have shown that the speech translation can be improved by using word lattices as input of the translation system. The acoustic recognition scores from the word lattice are used for improving the translation quality. However, word lattices do not consider word co-occurrences between different hypothesis and those probabilities are not real probabilities but merely Viterbi approximations. In this work, we propose an improved word lattice representation for using posterior probabilities instead of acoustic scores. We present preliminary results of this approach compared against other common approaches on two different corpora. Although the results are not strongly conclusive, they show that this approach is worth exploring more deeply. 1