Finally we hand labelled the Tilt events in the database. Note labelling Tilt events is much easier than labelling ToBI parameters as the type of accent and type of boundary tone does not need to be assigned, only its presence or absence.
As with experiment two, only half of the individual models show improvement over experiment one. In this experiment, we get RMSE of 33.9Hz and correlation of 0.57 over the 28 test utterances. These results fall in between those of the first and second experiments.