Finding words in a sea of speech

Abstract

Discriminating the words from normal connected speech is a notoriously difficult task; there are no known cues that act, as the white spaces on a page do, to separate individual words from each other. Nevertheless, infants must find the words to ensure the process of language acquisition and, in fact, they do so well before their first birthday. How do they do it? Two recent papers offer two very different insights into early word segmentation and identification. The first, by Saffran 1xWords in a sea of sounds: the output of infant statistical learning. Saffran, J.R. Cognition. 2001; 81: 149–169Crossref | PubMed | Scopus (134)See all References1, extends her previous work that demonstrated infants’ impressive abilities to extract nonsense tri-syllabic ‘words’ from a stream of continuous speech. The present study asks whether infants are treating these nonsense items as potential words or simply as identifiable but irrelevant patterns. In other words, do Saffran and colleagues’ results address the linguistic problem of word segmentation? Infants in this study were habituated, as in previous studies, to a continuous stream of nonsense syllables. In this stream, some tri-syllables appeared consistently as a unit whereas others did not appear as a coherent chunk; this formed the difference between the so-called ‘words’ and non-words. In one of the crucial test conditions, the nonsense words and non-words were embedded either in an English carrier phrase (e.g. ‘I like my pabiku’) or in a matched nonsense carrier phrase (e.g. ‘Zy fike ny pabiku’). The results showed that 8-month-old infants could still discriminate between the nonsense words and non-words, but only with the English carrier phrases and not the nonsense ones. Saffran argues that the statistical segmentation processes that are tapped in this task give rise to words that are indeed integrated into the larger acquisition process.In the second article, Brent and Siskind 2xThe role of exposure to isolated words in early vocabulary development. Brent, M.R. and Siskind, J.M. Cognition. 2001; 81: B33–B44Crossref | PubMed | Scopus (164)See all References2 demonstrate the power and availability of unconnected speech in the form of isolated words. Isolated words were defined as words occurring with at least 300 ms of silence on either side – in effect, words with a reasonable equivalent of acoustic white space around them. They collected speech samples in a very naturalistic setting from parents with children aged between 9 and 21 months (mothers were given recording equipment to use at home). Analysis of the parents’ speech showed that they reliably used a variety of distinct words in isolation, according to the specified criterion. More importantly, the children's own early vocabularies (as well as later word knowledge) reflected the words they had heard most frequently in isolation. Conventional wisdom has long held that isolated words would be a helpful entry point to the word and language acquisition process. Brent and Siskind have supported this conventional view (children do acquire isolated words early, and well) and moreover, they have shown that in at least some cultures (middle class American in this case), isolated words are indeed present in sufficient quantity and variety in the input to make the conventional story plausible.Taken together, these studies show us that young infants are very adept at finding real words in an unforgiving stream of continuous speech, and that in some cases, the input speech might not be so unforgiving as it first appears.