Assessing merged status with Pillai scores based on dynamic formant contours

Abstract

When Pillai scores are used to study vowel mergers, formants are typically sampled from the midpoint. This study compares alternative methods for calculating Pillai scores: methods that incorporate dynamic spectral information. Eighteen speakers produced 20 tokens of Hodd and hawed. Formants were sampled at 20–35–50–65–80% duration. Seven Pillai scores were calculated, each based on a different subset of those samples with temporal pooling: (i) onsets, (ii) heads, (iii) midpoints, (iv) onsets + offsets, (v) heads + tails, (vi) onsets + midpoints + offsets, and (vii) all five. Subjects also completed a vowel identification task, and the rate of identifying one low-back vowel as the other was calculated. The results of the identification task were regressed on each Pillai score separately to identify the one with the highest correlation, through model selection. Dynamic formant contours performed better than static formant values, with midpoint sampling performing worst of all. Directions are discussed for basic research on Pillai scores in phonetics.