Cluster analysis of Korean IP-final intonation

Abstract

In the Korean Tones and Break Indices (K-ToBI) system, Intonational Phrase (IP) final f0 contours are represented using tones at two levels, Low and High. However, functional analyses have shown the need for two additional levels, Mid and Top. These four levels are also used in some speech synthesis applications. This paper uses hierarchical clustering to investigate the f0 contours on IP final syllables in spontaneous dialogues. Our aim is to develop a classification scheme for analysing surface-level intonational variation in Korean. The clustering results show differences in register for the IP final f0 contours, supporting the usefulness of a phonetic analysis with more than two levels. The IP’s position in a sentence seems to affect the IP boundary tone; the low boundary tone for the sentence-medial IP may not be as low as that for the sentence-final IP. The mapping between meaning and boundary tones is not straightforward, and the effect of post-positional particles must be considered when investigating intonational meanings in Korean. Given the variability related to morphosyntactic and sociolinguistic factors, implementing a fine-grained classification scheme that goes beyond two levels is desirable for further exploration of large corpora.