Learning phonology from surface distributions, considering Dutch and English vowel duration

Abstract

In learning language, children must discover how to interpret the linguistic significance of phonetic variation. On some accounts, receptive phonology is grounded in perceptual learning of phonetic categories from phonetic distributions drawn over the infant's sample of speech. On other accounts, receptive phonology is instead based on phonetic generalizations over the words in the lexicon. Tests of these hypotheses have been rare and indirect, usually making use of idealized estimates of phonetic variation. Here we evaluated these hypotheses, using as our test case English and Dutch toddlers' different interpretation of the lexical significance of vowel duration. Analysis of thousands of vowels of one Dutch and three English mothers' speech suggests that children's language-specific differences in interpretation of vowel duration are likely due to detection of lexically specific patterns, rather than bimodality in raw phonetic distributions.

Keywords: language acquisition; lexical development; phonetic categorization; speech perception; word learning.

Figure 1.

Density plots of vowel durations. Each plot shows one {long,short} pair, with a different pair in each row, and a different mother in each column. The Dutch data are shown on the left, English on the right. The plot for phonologically long vowels is filled in a dark color, short vowels a light color. Duration is given on the x-axis in logarithmically scaled space. The y-axis shows the counts in the histograms. The smoothed density plots are arbitrarily scaled on the y-axis, though each is scaled the same way. The number in the upper right corner or each plot is the rank-biserial correlation between duration and phonological length; higher numbers indicate greater separation.

Figure 2.

Dutch vowels’ duration distributions, by pair, split according to word category (content word, demonstrative, function word), shown as histograms and density plots. Curves representing fewer than 10 tokens are bounded with dotted lines. The degree of separation between phonologically long and short vowels is operationalized using rank biserial correlation, indicated in the upper right of each panel, where larger numbers mean greater separation. Note that in Dutch, [ɪ-i] is not expected to show a duration distinction.

Figure 3.

English mother D1’s duration distributions, by pair, split according to word type (content word, demonstrative, function word). Plotting conventions are as in Figure 2

Figure 4.

Dutch duration distributions for individual word types containing the vowels [ɑ] or [aː]. Each column presents the mean duration (black point) and boxplot (IQR and range) for each word type, with the relevant syllable number indicated after the word along the x axis. The phonologically short vowel is shown with the darker symbols (red online), the long vowel with lighter symbols (yellow online). Content words, demonstratives, and function words are shown in separate panels. Numbers at lower left in the first and third panels are rank-biserial correlation coefficients.

Figure 5.

English duration distributions for individual word types containing the vowels [ε] or [æ] for mother W1. Conventions are as in Figure 4. The right panel omits one token of and that was 990 ms long.

Figure 6.

Duration distributions for the phonologically long and short vowels in individual words, for the 3 pairs in each language where phonetic differences would be expected. Columns of panels present data from different corpora, with Dutch on the left. Rows of panels correspond to vowel pairs. Within each panel, words are divided by category: content words, demonstratives, and function words. Darker points, on the left, are phonologically long vowels; lighter points are phonologically short vowels. Words occurring twice in each corpus are plotted with plus signs; words occurring more than twice are plotted with circles. Numbers above each pair of sets of points are rank-biserial correlation coefficients, which were computed when the number of word types exceeded ten. The data in the upper left plot correspond to the data in Figure 4.