Human newborns are able to discriminate between certain languages but not others. This ability has long been attributed to sensitivity to rhythm-the temporal regularities in speech of different languages. Here, we demonstrate through a series of computational simulations that this discrimination behavior can be achieved using no temporal information at all. Our results raise the possibility that early language discrimination may be independent of speech rhythm, and call for theoretical reconsideration of how infants learn suprasegmental information in the first few months of their lives. SUMMARY: We simulate newborn language discrimination with machine learning models on naturalistic speech stimuli. Models can discriminate between languages just like newborns, even when rhythmic information is removed. This means that infants' behavior could rely on their ability to perceive global properties of speech, rather than rhythm. Our results challenge theories of how newborns perceive speech, and how that shapes later language.
Famularo et al. (Fri,) studied this question.