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
Past experience shapes infant language acquisition and plays a decisive role for performance during tests in the lab, most crucially tasks which investigate recognition of putatively well-known words. Infants show that they can indeed recognise highly frequent words from their daily experience, most prominently their own name, early on and distinguish them from matched foils. However, it is not entirely clear which factors influence performance and which aspects of past experience infants benefit from. Using computational modelling, we simulate a word-recognition task with explicit reliance on previous encounters. The computational model used in this study assumes that infants bootstrap their language acquisition using only general-purpose perceptual and learning procedures. Specifically, the model does not rely on symbolic representations of linguistic units such as words or phonemes. Learning is incremental, i.e., internal representations are updated after each new utterance. Input is presented as real speech. In our study, we manipulate two factors that have been suggested to crucially shape infants' recognition abilities: The amount of experience (word frequency) and whether a number of different voices or just one speaker uttered a word (occasional variability). Word frequency has been assumed to aid recognition, but studies using high-frequency words encountered large between-participant variability pointing to a high degree of individual differences. Variability of speakers has been suggested to aid word learning, as it helps infants pay attention to the crucial aspects of the input signal. However, in their daily life infants with one main caregiver only occasionally encounter different voices. Our modelling results show that frequency, but not occasional speaker variability, improves recognition. By examining the impact of these two factors on modelled infant recognition performance in isolation, we can make predictions for future lab studies that will reveal how much variability is beneficial and to what degree new voices help word learning.