Janus is a multi-lingual machine translation system that translates spontaneously spoken dialogues in a limited domain. The machine translation components of Janus have been designed to handle the
disfluencies of spoken language and the imperfect output of the speech recognizer. These compoents include robust parsers (Phoenix and GLR*) and discourse processors (plan inference and speech-act bigrams) as
well as procedures for disambiguation and segmentation of multi-sentence utterances. This presentation will describe these components along with results of recent end-to-end evaluations of Spanish-to-English translation.