Modelling pitch accent types for Polish speech synthesis

We describe a Polish prosody modelling module for the Festival speech synthesis system. The module uses classification and regression trees for accent type prediction and a linear regression technique for F0 contour generation for these contours. The techniques used to attempt to overcome problems with the only available data are shown. We demonstrate how improvements were achieved by the use of a modified F0 stylisation, accent type clustering and language specific features. Results of a formal perception study show a significant preference for the new intonation model over the original one.