An accurate computational model of the prosodic patterns which organize utterances into coherent units of sound is essential for good-quality machine synthesis and recognition. Such a model will require an understanding of the abstract prosidic categories of intonation, stress, and phrasing, and a description of how they are implemented in pitch and rhythm patterns. This research is aimed at modeling the timing patterns of English as they reflect abstract prosodic categories. First, experiments will be conducted to explore the types of stresses and phases that affect speech timing, and the precise nature of the effects. A particular concern will be to discover a representation for these effects that will make the specification of the underlying prosodic patterns invariant across changes in speech tempo and segmental content. For example, the representation must allow slow and fast productions of machine recognition to be generated from an invariant specification of the stress and phrasing for this noun phase, and it must allow the use of the same specification in generating the prosodically identical noun phrase sublime delegation. The preliminary experiments will examine the timing patterns of the speech articulators involved in speech production as well as the resulting patterns evident in the acoustic signal. The computational model also will generate both articulatory and acoustic timing patterns.
