Comparing two different principles of parametric F0 modeling

A number of data-based approaches to intonation modeling represent F0 movements using continuous parameters. This is contradictory to most intonation theories, which suggest that intonation can be modeled with a set of distinct phonological entities that are phonetically realized as F0 movements. This principle has rarely been incorporated into data-based intonation modeling. In this study we compare two data-based intonation models following the two different principles described above. The first approach uses an F0 parametrization with 6 continuous parametric intonation event (PaIntE) parameters. These parameters are derived by approximating the F0 curve with an appropriate model function. In the second model we apply vector quantization (VQ) to the PaIntE parametrization, resulting in a number of distinct F0 shapes. We found that the VQ model has advantages over the PaIntE model, because the intonation events as a whole can be predicted, rather than individual parameters. Furthermore, phonetic analysis can be performed on the basic shapes represented in the codebook.