A comparison of four vowel overlap measures

Abstract

Measures of vowel overlap to explore the acoustic similarity between proposed and existing vowel categories. They typically compare F1 and F2, and sometimes duration. In the present study, we investigate four methods of quantifying vowel overlap: the spectral overlap assessment metric (Wassink, 2006), the a posteriori probability-based metric (Morrison, 2008), the vowel overlap assessment with convex hulls method (Haynes and Taylor, 2014), and the Pillai score as used by Hay et al. (2006). Based on the data for /i/ and /ɪ/ in the dataset of Hillenbrand et al. (1995), we used Monte Carlo style simulations and repeated subsampling to assess each method. We examined both the two-dimensional (F1 and F2) and three-dimensional (F1, F2, and duration) versions of the methods. We took the methods’ outputs as accurate if they produced values close to expected target values for each type of simulation, and we took the results as precise if there was little spread among the output values. The results suggest that the a posteriori probability-based metric is the most generally applicable, while the Pillai score should be used in scenarios where sensitivity to complete overlap is needed or where data cannot be said to be normally distributed.