Statistical speech-to-speech translation with multilingual speech recognition and bilingual-chunk parsing

Initiated mainly from speech community, researches in speech to speech (S2S) translation have made steady progress in the past decade. Many approaches to S2S translation have been proposed continually. Among of them, corpus-dependent statistical strategies have been widely studied during recent years. In corpus-based translation methodology, rather than taking the corpus just as reference templates, more detailed or structural information should be exploited and integrated in statistical modeling. Under the statistical translation framework that provides very flexible way of integrating different prior or structural knowledge, we have conducted a series of R&D activities on S2S translation. In the most recent version, we have independently developed a prototype Chinese-English bi-directional S2S translation system with the supports of multilingual speech recognition and bilingual-Chunk based statistical translation techniques to meet the demand of Manos – a multilingual information service project for 2008 Beijing Olympic Games. This paper introduces our works in the research of multilingual S2S translation.