Statistical Modelling of Stress Groups in Spanish

A statistical corpus-based synthesis strategy has been developed for fundamental frequency contours (##) of Spanish sentences. Input text is assumed to be made of a sequence of intonation groups, each one containing one or more stress groups. The stress group is taken as the basic prosodic unit at acoustic level. For every kind of acoustic unit, we get a set of statistical distributions for the parameters of a Bezier function that generates the ## contour of the unit. This distributions are obtained directly from the corpus and two different models are annalyzed. In one of them, linguistic knowledge forces the grouping of stress groups, while in the other an unsupervised clustering is carried out. Comparison of the results of both methods show the relative importance of different prosodic features found in real data