The first experiment consisted of generating parameter prediction models from the automatically labelled Tilt events. An optimised prediction model was created for each parameter (start F0, amplitude, duration, tilt, peak position) of each event type. That is, not all features were used in each model and hand experimentation was used to find an optimal set of features. The CART method can deal with a certain amount of noise in the input features but will be misled by too much noise (even with cross validation). Tables 1 - 3 show the RMSE and correlation of each of the twelve optimised models.
Table 1: RMSE and Correlation of accent models
Table 2: RMSE and Correlation of boundary models
The generated F0 is generally similar to the smoothed original. As an overall measure of accuracy for the 28 test utterance, we get an average RMSE of 32.5Hz and correlation of 0.60.
Table 3: RMSE and Correlation of c and sil models