Clusters and Classes in the Rhythm Metrics

San Diego Linguistics Papers 4 (2013) 28-52 Clusters and Classes in the Rhythm Metrics First reader: Amalia Arvaniti Russell Readers: Roger Levy, Andrew Kehler Russell Horton Horton and Amalia Arvaniti This co-authored paper incorporates work from my final paper for LIGN201, co-authored with Cody Brimhall and Emily Morgan, which was presented as a poster at the Acoustic Society of America in 2010, and from an associated paper draft co-authored with Brimhall, Morgan and Amalia Arvaniti. Abstract. The classic formulation of the rhythm class hypothesis holds that human languages are either stress-timed, with roughly equal duration of periods in between stresses, or syllable-timed, with all sylla- bles tending toward equal duration. This strict isochrony failed to be confirmed empirically, and various metrics have been proposed to quantify the typology in terms of durational variation of segments instead. We show that a Naive Bayes classifier and a graph-based clusterer can use certain metrics to discriminate between stress-timed English and German and syllable-timed Italian and Spanish with greater than 80% accuracy. They cannot, however, consistently and strongly predict rhythm types for sentences from the unclassified languages Greek and Korean, indicating that the rhythm typology may not generalize beyond the prototypical exemplars of each class. Although the purported success of the metrics in modeling the rhythm typology has been used as a proof of the existence of the typology itself, this is unwarranted, as the main argument that the metrics are valid measures of rhythm is the fact that they partially accord with the predictions of the rhythm typology. Further, the success of the metrics is in proportion to the degree that they correlate with tempo, suggesting that they may merely reflect the reduced durational variability of segments at the high speaking rates that characterize syllable-timed languages, rather than reflecting inherent differences in rhythm class timing. Although the clusters generated on the basis of the metrics appear to replicate the typology fairly well, this is due to partially overlapping ranges between languages and not strict separation. We show that the actual distributions of the metric values are incompatible with the strong claims made for the categorical, universal and fundamental nature of rhythm in speech. Introduction and Background Origins of the Rhythm Typology In his 1940 primer on telephony for a war-time audience, Arthur Lloyd James stresses the importance of rhythm for effective communication. Speech rhythms, he asserts, are of one of two kinds: the “morse-code” rhythm of languages like English, and the “machine-gun” rhythm of languages like French (Lloyd James, 1940). Five years later, Kenneth Pike formalizes this distinction, devoting three pages in “The Intonation of American English” to a topic he calls “Simple Rhythm Units (Stress-Time and Syllable-Time)”(Pike, 1945). In those pages, he sets out a theory of rhythm whose broad outlines remain influential today, even if many of its core precepts have had to be abandoned, “simplicity” first among them. Pike describes English as being structured around rhythm units, each containing a single strongly stressed segment, and each of roughly equal duration. Due to this principle of isochrony, the main stresses determine *This paper builds on work from my final paper for LIGN201, co-authored with Cody Brimhall and Emily Morgan, which was presented as a poster at the Acoustic Society of America in 2010, and from an associated paper draft co- authored with Brimhall, Morgan and Amalia Arvaniti.