Abstract
This study compares three major quantitative methods of contour analysis with the aim to establish their merit for intonation research. Utterances of 24 speakers (7–9 syllables long) taken from short dialogues were used to see whether general prosodic patterning determined by the intonational grammar of the language and the individual production habits of the speakers can be captured by computational means.
The three methods exploited were: k-means clustering (KMC), polynomial regression analysis (PRA), and functional principal component analysis (FPCA). The numerical outputs of the methods are confronted with human perception of the contour in both auditory and visual domains.
The results suggest that the observed contour properties are reflected by all three methods reasonably well: phonetically interpretable outcomes can be achieved by each of them. As to speakers’ individual features, KMC seems to be least vulnerable to spurious effects.