Phonetically gradient allomorph: the case of the Dutch past tense

A number of recent studies have shown that aspects of what was previously thought to be categorical phonology can be gradient, probabilistic and non-deterministic, as well as influenced by extra-grammatical factors (e.g. Pierrehumbert 2001, Davidson 2006, Kuzla et al. 2007). This argues for a more complex view of the relationship between speakers’ lexical knowledge and phonological representations, as well as of the phonology-phonetics interface, than is assumed in the modular feed-forward models of many standard phonological accounts. Our study of past tense formation in Dutch provides further evidence for this complexity, showing how both allomorph selection and the fine phonetic detail involved are partly influenced by extra- grammatical factors such as phonological neighbourhood density and lexical frequency.
In the standard account, the choice between the two allomorphs of the Dutch regular past tense morpheme, -te and -de, is based on the underlying fortis/lenis specification of the stem-final consonant of the verb it attaches to, as in (1).
(1) verb stem sg past tense verb stem sg past tense
dans /d!ns/ danste [d!nst"] “dance” bons /b#nz/ bonsde [b#nzd"] “bang” surf /s$rf/ surfte [s$rft"] “surf” durf /d$rv/ durfde [d$rvd"] “dare”
However, Ernestus and Baayen (2001,2003,2004) report that Dutch speakers often select the “wrong” allomorph, leading to misspellings such as bonste,surfde. They further show that frequency and analogy are significant predictors of the occurrence of such misspellings, and argue that these forces partly drive allomorph selection, instead of it being solely and deterministically based on underlying specifications.
Our study examines the extent to which the incongruous allomorph selection is observed in speech (as opposed to spelling), and whether or not it bears on the phonetic detail of past tense formation. We elicited 864 past tense verb forms with fricative-final stems from 8 native speakers of Standard Dutch. We analysed the obstruent clusters in these past tense forms with respect to the potential parameters of the fortis/lenis contrast (vocal fold vibration, fricative duration, stop closure duration, burst duration and intensity, f0 and f1 of the following vowel). The measurements were then used to classify all past-tense forms produced by the participants as containing a –te or –de suffix, based on a by-speaker linear discriminant analysis. 27.2% cases of mismatch were found between the classification results and the prescriptive target suffix.
A linear mixed-effects regression model predicting the occurrence of mismatches showed significant main effects of token frequency and an index of analogical strength based on neighbourhood density. In addition, a series of models showed significant interactions between the prescriptive target and whether or not it agreed with the classification result. Fricative duration, closure duration, burst duration and f0 all showed significantly greater differences between pronunciations classified as –te and –de within the prescriptively accurate cases, compared to when the realisations deviated from the prescriptive norm. In other words, the fricative+te and fricative+de clusters in verbs with a mismatched allomorph are more similar to each other with respect to these voicing cues, i.e. they show more category overlap, than those with a matching allomorph.
Based on these results we argue that the morphophonological rule as well as neighbourhood density and token frequency have an impact on allomorph selection as well as on the phonetic detail of past tense verb forms, as any conflicting information provided by these sources is reflected in more phonetically ambiguous realisations.