First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

The Making of a Speech-to-Speech Translation System: Some Findings from the DmDialog Project

Hiroaki Kitano

Center for Machine Translation, Carnegie Mellon University, Pittsburgh, Pennsylvania, USA

In this paper, we discuss some important problems which we encountered in developing the DmDialog real-time speech-to-speech translation system. DMDlALOG is a real-time Japanese-English speech-to-speech dialog translation system that accepts speaker-independent continuous speech input and produces audio output of translated sentences. It has been publicly demonstrated at the Center for Machine Translation at Carnegie Mellon University since March 1989. Major problems are seen in the areas of (1) integration of linguistic and statistical nautral language processing, (2) use of discourse knowledge to improve speech recognition rates, and (3) attainment of simultaneity in translation. We chose these problems to discuss here because they are very difficult and unique to speech-to-speech translation systems; they are not encountered in text-based machine translaton systems. We describe our approaches to these problems and describe problems left unresolved, so that this paper can serve as a basis for further discussions on these issues.

Full Paper

Bibliographic reference.  Kitano, Hiroaki (1990): "The making of a speech-to-speech translation system: some findings from the dmdialog project", In ICSLP-1990, 1241-1244.