Robust Knowledge Discovery from Parallel Speech and Text Sources

Johns Hopkins University, Baltimore MD

Investigators

Abstract

The goal of this project is to enable information retrieval systems to handle speech-based information by considerably enhancing speech recognition technology. The PI's approach is to align collections of stories from text and speech sources in multiple languages, and to develop methods which exploit the resulting parallelism. The PI will improve acoustic and language models, and will develop iterative recognition methods to facilitate capture of cross-lingual information resident in text and speech. Voice-of-America broadcasts in English, Greek, and a third language to be determined, will provide the data source for this research.