Intonation Synthesis with AM models

We now turn our attention to this issue of intonation synthesis itself. In Chapter 6 we described a variety of techniques for generating the prosodic form representations of phrasing and prominence from text. We only made passing reference to the issue of generating intonational form fi-om text as this really required a thorough discussion on the issue of intonational form first. With this job now done, we can describe techniques for both generating intonational form descriptions fi om text and for generating FO controls from intonational form. 

While it is possible to build synthesis algorithms with any of the models outlined in Section 9.3.3, we will focus our attention on the most popular techniques, which divide into those based on AM models (described here) and those based on acoustic models (described next). 

There is a basic distinction in most models as to whether intonation is an inherently tone or pitch level based system (as in AM models), or is based on pitch shapes or pitch dynamics (as in the Dutch model, FujisaM model and Tilt model). In theoretical linguistics this debate is seen as particularly important since researchers are attempting to find the true nature of intonation, with conclusive evidence occasionally being produced for one view over the other. The issue is hard to resolve, and it may be the case that intonation is actually a mixture of both phenomena. 

Some models have a superimpositional nature (Dutch, Fujisaki) such that pitch accents are seen as being relatively short-term, riding on top of phrases, which are seen as being relatively long. By contrast, in linear models (AM/Tilt) contours are composed of linear sequences of intonational units. Often a grammar that states which units can follow each other is used, but, so long as the sequence lies within the grammar, any unit can foUow any other. 

In brief, the main reason for supporting the superimpositional model is thatFO contours do seem to exhibit patterns of global behaviour, whereby phrases define particular FO patterns. The main reason for supporting linear models comes from speech-production concerns. Here, we see that if we support a superimpositional model then the speaker has to pre-plan the utterance some way ahead of time. This doesn t seem to correspond to known facts in speech production, where we know for example that a speaker can change the course of the FO contour at any time. 

Finally we should note that some papers have shown how labels in one model can be translated into labels in anotiier [374], [436]. 

---