A study of JEMA for intonation modeling

In the literature many intonation models are trained using parameters extracted sentence-by-sentence on contours interpolated in the unvoiced segments. This may introduce a bias in the final parameters and a reduction of the generalization of the model due to the increased dispersion of them. Recently, we have proposed JEMA, a joint extraction and prediction approach for intonation modeling that avoids such assumption. The parameter extraction and model training are combined in a loop where i) the model is successively refined, and ii) the parameters are extracted using information provided by the model. In this papers we present experiments based on synthetic data to evaluate this approach in a controlled environment. Both, the results with synthetic data and with natural speech, show that the use of JEMA is clearly superior to the standard estimation approach. The parameters are correctly extracted using several degrees of missing data (0% to 80%) and Gaussian noise. In fact, the study shows that including JEMA in the training algorithm is even more relevant than the selection of a particular representation of the intonation contours, as Fujisaki, Bezier, Tilt, or others.