This paper describes NewExpress, the new text-to-phonetic-representation component of the AT&T Bell Laboratories Text-to-Speech system (TTS). To the best of our knowledge, NewExpress represents the first extensive use of corpus-based linguistic techniques in a text-to-speech program. We discuss the use of such techniques in the system in four main areas: general pitch accent assignment, prosodic phrasing, pitch accent assignment in noun compounds, and homograph disambiguation. We demonstrate that these techniques afford an improvement in the performance of TTS.