The detection of emphatic words using acoustic and lexical features

In this study, we describe an automatic detector for prosodically salient or emphasized words in speech. Knowledge of whether a word is emphatic or not could improve Text-to-Speech synthesis as well as spoken language summarization. Previous work on emphasis detection has focused on the automatic recognition of pitch accents. Our model extends earlier research by automatically identifying emphatic pitch accents, a subset of pitch accents that mark special discourse functions with extreme degrees of salience. The overall best performance achieved by our system was 87.8% correct, 8.0% above baseline performance. The results of a feature selection algorithm show that the top-performing features in our models are primarily acoustic measures. Our work identifies important cues for emphasis in speech and shows that it is possible for an automated system to distinguish between two levels of perceived prominence in pitch accents with a high degree of accuracy.