next up previous
Next: Conclusion Up: Discussion Previous: Discussion

Context features

The majority of the context features used in this study have been tested previously. [2] However we found that the features used in other work were inadequate for the prediction of peak alignment, so we adopted the use of rhyme length [8] (though non proportional, as supported in [11]), onset and coda classes [11], and onset length [5] (compare accent peak position correlation of 0.42 without use of these features to 0.51 when they are included).

Of these new inclusions, rhyme length was a central feature in all of the peak position predictions (accents, boundaries, with and without ToBI). Coda classification was useful in non-ToBI boundary peak alignment, as well as non-peak predictions (accent amplitude and boundary durations, without ToBI). Onset type was used in both accent peak alignments, but not in the boundary peak alignments. Onset length was among the optimised feature sets for only boundary peak position (with ToBI), but appeared in both ToBI and non-ToBI sets for accent duration and accent starting F0. Thus, while not all of the features meant to aid peak alignment appeared in prediction models as expected, they did improve the prediction of the peak position, as well as contributing to other prediction models.

Kurt Dusterhoff
Tue Jul 1 11:51:11 BST 1997