Comparing Methods for Locating Pitch "Elbows"

The labeling of “elbows” in an F0 contour is considered an enterprise beset with difficulty due to the inability of humans to locate pitch elbows with accuracy, consistency and in a manner devoid of theoretical bias. This paper investigates the extent to which human labelers agree with one another in locating elbows and how four algorithms compare to their results. Humans are found to be more consistent than has been suggested and a least-squares fitting algorithm best approaches their intuition. The success of algorithmic elbow location depends on the selection of the contour stretch in which the elbow is to be located. This elbow location is most consistent if performed by a theoretically-informed annotator, suggesting that an atheoretical annotation of F0 contours may be impossible to achieve, and ultimately undesirable.