The model we use is written in Soft Template Mark-up Language (Stem-ML)[7,8], and depends upon its underlying mathematical model of prosody control. Stem-ML calculates an intonational contour from a set of tags that describe the intonation.

Our view of intonation starts from a small collection of tone classes, each of which implements some linguistic function. From these, we calculate the surface realization of the pitch as a function of the strengths and classes of nearby accents.

We write a Stem-ML description in terms of a set of undetermined parameters, then find the parameter values that accurately reproduce a training corpus. The fitting is done automatically.

Greg Kochanski, Chilin Shih 2002-08-03