A language model for conversational speech recognition using information designed for speech translation

In this paper, a new language model is proposed for speech recognition in conversational speech translation. In conversation, speech strongly depends on the previous utterance of the other participant. Applying this dependency in language modeling, we can reduce the speech recognition error rate. To this end, we propose the following new language model where the content of the previous utterance is expressed by an interlingual representation which is widely used in the spoken language translation research group C-star. The proposed method reduces word error rate by 6% (from 14.7% to 13.9%), con rming our expectations.