The measurement of the accuracy of the F0 generation process followed previous studies in using the root mean squared error (RMSE) to measure how far away the predicted F0 is from its target (the original F0 of an utterance) at any given time. Correlation is a more general measure which shows how well a contour changes in time with its target (i.e. do they both rise and fall in the same places). The RMSE is tied very closely to the speaker being tested (generally smaller for males than for females, due to overall pitch range), while correlation provides a more independent measure.
In addition to the overall measure, RMSE and correlation were used to judge the accuracy of the individual models used in the development of the decision trees. These individual results aided in the selection of optimized feature sets (i.e. not all features were used for predicting all parameter values) for each parameter, as well as deterimining which parameters were in need of improvement.