The Verbmobil System currently under development in Germany would input speech in a source language dialogue, translate it, and output synthesized speech in a target language, all in real time under the conditions of face-to-face dialogue. This preliminary study is an assessment of the state of the art of the speech recognition and machine translation fields, and a frank discussion of the challenges the developers of such a system face. A speech-based machine translation system in effect combines all the technical problems of understanding and generation systems and adds the very special difficulties of translation. The difficulties of translation are illustrated with a variety of examples culled from a number of languages. A general architecture called translation-by-negotiation is proposed. This book is of value not only to students of speech recognition and machine translation, but to anyone interested in natural language processing in general, since computational problems and linguistic approaches at various levels of linguistics and analysis are discussed.