Models of FO contours

In practice standard regression algorithms are difficult to use directly for FO prediction, because of a serious mismatch in the nature of the input and output representations. In general the input representation is a representation of intonational form as just discusses (for example, pitch accent types and positions) whereas the output is a continuous list of real valued numbers. In particular, the feature combination needs to generate not one, but a sequence of FO values, which is further complicated because the number of values in this sequence can vary from case to case  

Furthermore, when we come to performing a bottom up analysis, where we wish to describe the continuous contour with a discrete set of linguistic units, we find the mapping even more complicated. Because of these problems, it is nearly universal to use some sort of acoustic model which acts as an interim representation between the abstract intonational form and the FO contour. This model has a fixed number of parameters per unit, and so we can use one of the standard mapping algorithms (such as CART) to generate these parameters from the intonational form  

As a second step we can use the specialised model to generate the FO values themselves  

These models usually work on a principle whereby they encode or stylize the real FO contour. In other words, they give a tractable simpUfication of the FO contour in which the salient linguistically important aspects of the contour are preserved, while less-necessary parts are discarded. At its simplest, the parameters would be a discrete listof target points in the contour. From these, a continuous FO contour can be synthesized. If we keep the number of points constant for each linguistic unit, then we have a situation wherein the parameters of the model align lucely with the linguistic description, while, it is hoped. 

---