Bergmann, C. 1, 2 , ten Bosch, L. 1 & Boves, L. 1
1 Centre for Language Studies, Radboud University Nijmegen, The Netherlands
2 International Max Planck Research School for Language Sciences
Early language perception and processing abilities are usually investigated using a variety of procedures that rely on a neural or a behavioural response as indirect measure of infants' abilities. Computational models offer a different perspective on language acquisition, since modelling requires precise statements about all assumptions and hypotheses. By the use of computational modelling, we can closely examine (1) which underlying factors need minimally be present to give rise to neural responses or behavioural patterns usually observed during experiments and (2) which factors, be it linguistic or extralinguistic, these methods are susceptible to. We present a model of the headturn preference procedure (HPP) that simulates infant behaviour based on minimal assumptions, such as the lack of language-specific phoneme representations, no explicit segmentation of continuous speech and a module directly converting recognition into overt behaviour, the eponymous headturns. Headturns are only modulated by recognition and infant attention span. Individual attention span accounts for some of the frequently observed between participant variability in HPP studies. According to our model, too short or too long attention span can even mask underlying abilities. As a case study, we replicated HPP experiments on cross-speaker generalisation. Infant data show a mixed pattern. Our model shows that cross-speaker generalisation depends on acoustic properties of the test speaker. This result can help reunite seemingly conflicting infant data on cross-speaker generalisation. Thanks to our modelling result, we can make concrete predictions on the influence of stimulus material on infant performance during a typical HPP experiment on segmentation. Based on our model, we predict that infant performance depends on which speaker infants are being tested with in the lab. Future studies are necessary to investigate this issue by providing detailed acoustic comparisons of test speakers and linking those to infant performance.