Background

As an initial part of the DARPA Babylon project we were tasked with building an interlingua-based two-way speech-to-speech translation system on a small device in a language that our group had no (significant) previous experience in. This required us to solve three specific problems:

• How to collect sufficient and appropriate data for translation, recognition, and synthesis, in the most efficient way. Using foreign language experts we designed protocols to define the translation domain and collect examples to allow an appropriate interlingua to be designed.

• How to take advantage of both knowledge-based techniques in defining an interlingua; and statistical techniques, in learning the relationship between surface forms and that interlingua; in such a way to make transfer to new domains and languages efficient.

• How to fit two recognizers, two synthesizers and a two-way translation system on a device with only 40Mb of available space and limited CPU power. This required addressing engineering issues: lack of floating point support, synthesis database compression, efficient recognition decoding algorithm; as well as research issues in model design for size and efficient access.

The end result is a working prototype on a Compaq iPaq which can recognize, translate and synthesize bi-directionally between two languages, English and Egyptian Arabic, and do so in a reasonable time. Although this prototype is limited, it was aimed at medical interviews, and deals with only many hundreds of sentence types, it shows the feasibility of such a system.

This particular system was built over a period of six months, using the tools and techniques we have developed over a number years in rapid development for speech-to-speech translations systems [1], [2].