Proceedings of the Annual Meeting of the Cognitive Science Society

Humans learn much about their language while still in thewomb. Prenatal exposure has been repeatedly shown to affectnewborn infants’ processing of the prosodic characteristics ofnative language speech. Little is known about whether and howprenatal exposure affects infants’ perception of speech soundsegments. Here we simulated prenatal learning of vowels intwo virtual fetuses whose mothers spoke (slightly) differentlanguages. The learners were two-layer neural networks andwere each exposed to vowel tokens sampled from an existentfive-vowel language (Spanish and Czech, respectively). Theinput acoustic properties approximated the speech signal thatcould possibly be heard in the intrauterine environment, andthe learners’ auditory system was relatively immature. Withoutsupervision, the virtual fetuses came to warp the continuousacoustic signal into “proto-categories” that were specific totheir linguistic environment. Both learners came to create twocategorization patterns and did so in language-specific ways,primarily on the basis of the vowels’ first-formantcharacteristics. Such prenatally formed proto-categories werenot adult-like in that they entirely collapsed some of the native-language contrasts. At the same time, the categories reflectedfeatures of the adult language in that they were language-specific. These results can inspire future work on speech andlanguage acquisition in real young humans.