Speech recognition strategy for stress-induced speech

Abstract

The impact of style variations, and in particular stress-induced variations, on speech recognition performance is discussed. The approach is threefold. First, databases of stress-induced speech are used to measure the recognition performance of humans and commercial devices. Second, acoustic-phonetic analysis of the data is done to determine correlations between style variations and recognition performance; such a analysis helps to determine how changes in strategic design might better accommodate prosodic variations. Third, these hypotheses are tested using development recognition systems.<>