ENGLISH NOUN PHRASE ACCENT PREDICTION FOR TEXT-TO-SPEECH

This paper describes np, a component of the AT&T Bell Laboratories English text-to-speech system that computes accentuation for constructions such as apple cake, apple pie and sump pump factor. Np uses both rule-based and statistical 'corpus-based' methods. These methods are discussed, and their benefits and shortcomings enumerated. The various components of np are evaluated, and it is shown that the overall performance of np significantly reduces the error rate in accent assignment over some simple-minded 'baseline' approaches. The paper concludes by outlining some future areas of research.