I present the first steps in a research program rethinking how intonation is represented and stored cognitively. My proposal is that intonation contours are stored in the first instance with the utterance they are produced on. The link between the contour and the words decays if it is not sufficiently discriminable and robust; however, if frequent enough, intonation contours can be lexicalised, i.e. associated with particular lexical items or discourse units. If true, this leads to a very different way of conceptualising how intonational meanings, such as implicature, arise.
In the first stage of testing this approach, I present the results of a statistical analysis of the association between tonal pitch accent curves and lexical types in the Switchboard corpus. I firstly used an automatic pitch parameterisation algorithm, PaIntE (Mohler 2001), to define a "phonetic space" of the pitch curves of nuclear accents. I then used statistical clustering techniques to identify distinct types of pitch curves. I then use collocation measures to look at the association between these pitch accent curve types and the words they fall on.