Project home

MedSLT is a multilingual spoken language translation system tailored for medical domains. The system is designed to help in situations where no common language between the diagnosing doctor and the patient exists.

In the typical case the communication by using the medSLT system is as follows:

  • The doctor selects sub-domain and language pair
  • The doctor asks diagnosis questions
  • The diagnosis questions are translated by the system into the target language
  • The patient answers the questions saying "yes" or "no", by nodding or by pointing the body part (unidirectional version)
  • The patient answers by speaking directly to the system (bidirectional version).

Currently the system supports speech recognition and translation for the headache, chest pain and abdominal pain domains. Supported languages are English, French, Japanese, Spanish, Catalan and Arabic.

The project is financed by the Swiss National Science Foundation.

MedSLT in a nut shell

  • Open Source speech-to-speech translation system for medical domain
  • One way translation of yes/no questions
  • Interlingua based translation
  • Translation on headache, chest pain and abdominal pain sub domains
  • Supported Languages: English, French, Japanese, Spanish, Catalan and Arabic
  • Exploits the same grammar for recognition, parsing and generation
  • Tools used in the system development: Regulus, Nuance

System Architecture

The diagram depicts Regulus compile and runtime components (green items) and their interaction with the NuanceTM Voice Platform (blue items). At compile time the Regulus grammar compiler reads in domain independent grammars along with application specific input files, and generates a recognition grammar. The recognition grammar is in turn compiled into a recongition package.

At runtime the medSLT application creates instances of a platform interface process and a translation server. The translation server loads translation rules along with the generation grammar for the target language. The generation grammar can be compiled from the same domain independent grammar and corpus used for recognition.

  © Copyright 2007-2008 ISSCO/TIM/ETI. All Rights Reserved