Domain Unlimited Speech Translation

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

All speech-to-speech translation systems to-date have focused on domain-limited tasks such as travel planning, scheduling, and hotel reservation. STR-DUST focuses on the exploration of new algorithmic approaches that break through this limitation to enable domain-unrestricted speech translation (as would be needed for translating telephone conversations, lectures, meetings, broadcasts etc.) These approaches build on learning approaches and multiple layers of reductions and transformation to extract the desired content from an utterance in one language and render it in another language. The project builds on progress in statistical text translation, spoken language summarization as well as conversational speech recognition and merges these critical components in a jointly trainable framework to achieve unrestricted speech translation. All components are developed under a statistical framework so as to jointly optimize overall speech translation performance. The project also involves collaboration between research partners in Europe and the US. It will carry out an aggressive, bold research plan while developing a strong multinational educational program. The collaboration develops joint courses, seminars and workshops to train young scientists in Machine Translation, Speech, and Language Processing in an international and multilingual setting. In addition, the prject has impact in its educational mission by training a new generation of MT savvy, multilingual and internationally adept scientists.