Interchange Format-based Language Model for Automatic Speech Recognition in Speech-to-Speech Translation

This paper relates a methodology to include some semantic information early in the statistical language model for Automatic Speech Recognition (ASR). This work is done in the framework of a global speech-to-speech translation project. An Interchange Format (IF) based approach, representing the meaning of phrases independently of languages, is adopted. The methodology consists in introducing semantic information by using a class-based statistical language model for which classes directly correspond to IF entries. With this new Language Model, the ASR module can analyze into IF an important amount of dialogue data: 35% dialogue words; 58% speaker turns. Among these 58% turns directly analyzed, 84% are properly analyzed.