Identifying the regularities underlying speaker decisions to emphasize or de-emphasize an item intonationally has long been the subject of speculation and controversy. This paper describes a study of accent assignment based upon the analysis of natural recorded (read) speech. Results are being incorporated in NewSpeak, an interface to the Bell Laboratories Text-to-Speech System, which varies intonational features based upon syntactic structure and higher-level discourse information inferred from unrestricted text, in order to generate more natural synthetic speech. Implications of the work for the evaluation of discourse models, for automatic labeling of prosodic features, and for speech synthesis are discussed.