Multilingual Web Conferencing Using Speech-to-Speech Translation

摘要

It is now commonplace to use web conferencing technology in order to hold meetings between participants situated in different physical locations. A drawback of this technology is that nearly all of the interaction between these participants is monolingual. Here, we demonstrate a novel form of this technology that enables cross-lingual speech-to-speech communication between conference participants in real time. We model this translation problem as a combination of incremental speech recognition and segmentation, addressing the question of finding which segmentation strategy maximizes translation accuracy while minimizing latency. Our demonstration takes the form of a web conferencing scenario where a presenter speaks in one language while talk participants listen to or read the speaker's translated texts in real time. This system is flexible enough to allow realtime translation of technical talks or speeches covering broad topics.