Abstract
Deciding how much two vowel categories overlap is a core problem in laboratory phonology and sociophonetics, and the field has proposed several metrics to quantify overlap: the Pillai trace, Bhattacharyya distance, and Euclidean distance, among others. Researchers rarely compute multiple metrics on the same data, and what look like distinct overlap definitions are sometimes different estimates of the same quantity. I introduce Phontrast, an open-source R toolkit that computes multiple overlap metrics on the same tokens, and I simulate distributions with known overlap to disentangle each measure's estimator from its estimand. Two public vowel corpora show how measures behave on real data, in F1 × F2 space and thirteen-dimensional MFCC space. Mean- or Gaussian-based measures are precise but unresponsive to separation in higher moments. Kernel estimates are biased and bandwidth-dependent, but within the simulation's operating region they recover the ordering on average, whatever the mechanism. Pillai is the most precise for small samples and whenever separation is driven by means. Since this is rarely known a priori, I foreground Jensen–Shannon distance, a distribution-free distance between the two vowels' probability distributions that registers separation however it arises, for ordering speakers and contrasts—with Pillai beside it to show movement in category means.