Vowel production in infant-directed speech: an assessment of hyperarticulation and distributional learning
This thesis considers two related research questions that are relevant to vowel production in infant-directed speech (IDS). The ﬁrst of these research questions determines the extent to which caregivers intentionally clarify vowel distinctions which correspond to diﬀerences in word meaning when they address infants in order to promote language learning. The second research question determines the extent to which infants can learn these distinctions by observing statistical regularities in the acoustic input that they receive. In order to observe the modiﬁcations that speakers make when they address infants and the eﬀect that these modiﬁcations have on infants’ learning of the distinctions between vowel sounds, I carried out an acoustic analysis of vowel production data sampled from a large naturalistic corpus of American English caregivers speaking to both infants and adults. This comparative analysis of infant- and adult-directed speech (IDS and ADS) extends previous research in that it applies multiple measures of discriminability to high dimensional acoustic data which details the properties of all of the categories in the system. Though speakers produced a greater dispersion between vowels when they addressed infant learners than adults, their vowel production in IDS was also more variable than ADS. Because of this, the modiﬁcations that characterise IDS are not consistent with pedagogical accounts of this register. In order to assess whether infants can identify distinctions from the statistical properties of the input, I applied a series of learning models to the acoustic data that was sampled from each register. Models applied to both IDS and ADS failed to recover the identity of vowels in American English. These results suggest that vowel distinctions cannot solely be learnt by observing the statistical regularities of the input. This research indicates that phonetic data must be examined closely in order to determine the intentions behind the modiﬁcations that caregivers make in IDS. The learning models presented here inform current theories of perceptual development by indicating that perceptual learning additionally requires the consideration of infants’ early knowledge of multiple levels of linguistic structure. Further to this, this thesis demonstrates that the successful cases of vowel learning in laboratory contexts cannot be trivially generalised to allow learners to discover distinctions in the input that they are exposed to.