In the second experiment the same Tilt events derived from the ToBI labelling are used but this time we include the ToBI labels as features used to predict the individual parameter values. This is done to test what contribution the finer accent and boundary tone distinctions would make to our model.
The individual model results are similar to those of the first experiment. Six of the parameters score higher than the first set, while six score lower.
As with the results of the first experiment, the generated F0 is generally similar to the smoothed original. For this experiment, we get slightly worse overall results with an average RMSE of 34.0Hz and lower correlation of 0.55 over the 28 test utterances.