Segmenting artificial languages in infancy: a meta-analytic view

Black, A. 1 & Bergmann, C. 2

1 University of British Columbia, Department of Linguistics
2 Ecole Normale Supérieure, PSL Research University, Département d'Etudes Cognitives, Laboratoire de Sciences Cognitives et Psycholinguistique (ENS, EHESS, CNRS)

Infants' ability to extract statistical patterns from a continuous speech stream for word segmentation has become a foundation of theories of language acquisition. However, it is as yet unclear how robust and large the underlying effect is. For example, studies have shown that infants fail when presented with words of mixed syllable length - calling into question how much statistical learning for speech segmentation scales to natural experiences. For theory building as well as experiment planning, it is important to (1) establish the strength of the effect, and (2) understand how it might be modulated. We present a meta-analysis that reveals a small effect size in direct replications of the seminal report that is significantly different from zero (Hedge's g = 0.2, p = .03), but no effect when considering all broadly-defined conceptual replications (g = 0.09, p = .1). Follow-up analyses reveal no effect of infant age (contra models of information processing and maturation), but that synthetically produced stimuli elicit a different behavioral pattern compared to natural stimuli, despite the presence of statistical information in both types of studies. In sum, this meta-analysis invites future examination of infants' statistical learning capabilities in the lab and in daily life.