A technique for the automatic location and description of pitch contours

A technique is presented for automatically locating and describing the shape of fundamental frequency (F0) patterns throughout a sentence. The function y(t) = at^be^ctis used to model the rise-fall shapes which define F0phrases. The function parameters are derived from a stepwise regression with the F0data, where new values are added to the region of analysis until the variance of the fit gets too large or the next F0point falls outside a prediction interval. The modeling procedure was applied to sentences taken from simulated man-machine interactions; the locations and shapes of the resulting F0phrases were evaluated against phonological phrase structures determined by perceptual experiments. 80% of the phrase boundaries perceived by listeners were located by the F0model. However, 21% of the boundaries found by the F0model were not found perceptually and are classed as False Alarms. In a comparison of perceived phrase types and contour shapes defined by the model, sentence-final phrases marked as complete clause boundaries by listeners were generally found as falling or level patterns, while those sentence-final phrases called incomplete were all found as rising contours by the model.