A standing body of work (e.g. [1,2,3,4]) has compared different vowel-normalization methods for sociophonetic research. Except for [2], most of this prior research focused on steady-state vowels, taking e.g. the vowel midpoint as a single representative sample. In light of concerns about the validity of this approach ([2,5,6]) and related observations that even monophthongal phonemes can be phonetically time-dynamic ([2,6,7,8]), we revisit the issue of vowel-normalization techniques taking temporal dynamics into explicit consideration. Our comparison is based on [2]’s hand-corrected measurements of the Dutch ‘teacher corpus’ [9,10]. This provides word-list data of 160 speakers from 8 regions of Standard Dutch (four in The Netherlands and four in Flanders), stratified by sex (male/female) and age (old/young), making five individuals per cell. On the basis of these data, we compared sixteen contemporary normalization methods, as implemented in Visible Vowels ([11]), and an unnormalized baseline. We compared performance along four domains:

(1) Normalizing anatomical variation, operationalized by the ability to recover the speaker sexes (cfr. [1,2]) and the individual speakers on the basis of the normalized F0—F3;
(2) Retaining vowel distinctions and the participants’ regional origins in the normalized F0—F3;
(3) The amount of variation in the normalized F0—F3 that could be explained by the separate factors of sex, vowel, region, and their interactions;
(4) The total explained variance in the vowel trajectories per region, possibly including by-speaker random effects.

(1) and (2) were operationalized using multinomial GAMs ([12]), modeling the vowels’ temporal trajectories using nonlinear smoothing splines through five timepoints (from 25% to 75% in equidistant steps; [2]), taking the classification accuracy as our metric of interest. (3) was operationalized using multivariate normal GAMs, where we computed partial-eta-squared values (cfr. [1,2]) for the eight factors of interest. (4) was operationalized using univariate GAMs, with R² as our metric of interest. For (4), we fitted models both with and without by-speaker random effects, expecting a major contribution of random effects for the unnormalized dataset, but a more modest contribution for the normalized data. We expected this because the normalized data should have eliminated some (anatomical) between-speaker variation, but the unnormalized data should leave this variation to the random effects.

Our results partly reproduce the observations by [1,2] that Lobanov, together with Gerstman and Nearey I, perform excellently on the classification tasks (domains (1) and (2)). However, other methods, particularly Heeringa & Van de Velde II, achieve higher effect sizes (domain (3)) at only marginally worse classification performance. In domain (4) we observe that all methods but one (Thomas & Kendall) accrue a similar benefit to the inclusion of random effects. Importantly, this includes the baseline, showing that rather than random effects replacing normalization, these two methodological decisions complement each other. Finally, we note a general advantage of log-mean and centroid normalization methods, which we interpret with reference to their consideration of temporal dynamics.