Incremental Speech Translation: A Layered Chart Approach

Human speech understanding works incrementally. We begin to process acoustic input before the speaker's utterance has ended. A system capable of performing sophisticated communication in a natural dialogue or simultaneous interpreting, has to work incrementally, too. The architecture of such a system should be modular, uniform and integrated. We present an architectural framework that suits these three requirements by implementing layered charts, a multipurpose data structure intended to represent several competing hypotheses about linguistic content of utterance intervals based on hypergraphs. We demonstrate the feasibility by presenting results from an actual interpreting system.