Prediction of intonation patterns of accented words in a corpus of read Swedish news

This paper describes an initial attempt at the construction of a data-driven model of Swedish intonation. The study is mainly concerned with model building and prediction of the intonation patterns of accented words in a corpus of read news in Swedish. Extraction of pitch information is achieved by performing a stylization of the pitch contours. The information is used to build a model for the prediction of pitch patterns using linguistic features such as accent type and position of stress. The model is tested against unseen data from the same corpus. The evaluation is done by numerical comparisons. The RMSE between predicted and original contours for the different categories ranges between 3.7 and 31.4 Hz. The results are quite promising for future studies.