Improved representation of variance in measures of vowel merger.

Abstract

Current measures of vowel merger, such as the Euclidean distance between averages, have only been able to capture some of the variability between two given vowel clusters. Reliance on averages obscures the amount of variability within a given vowel class, while other techniques, such as calculating distance between minimal pairs, rely on few tokens per speaker. Hay et al. (2006) introduced an alternative approach that accounts for the variability between two vowel clusters, taking formant values as input, rather than averages. The measure is the Pillai-Bartlett statistic [Baayen (2008)], an output of a multivariate analysis of variance (MANOVA), which represents the proportion of one variance that can be predicted by another variance. A higher Pillai value indicates a lower degree of overlap between two vowel clusters in F1/F2 space. Since the value is derived from a MANOVA, Pillais can account for known internal factors influencing the production of merger, such as phonological environment, thereby reducing the need to obtain minimal pair lists. This talk argues for using Pillais as measures of merger by comparing results from low back merger in California English [Hall-Lew (2009)] with the analysis of front vowel merger by Hay et al. (2006).