A model of prenatal acquisition of vowels

Humans learn much about their language while still in the womb. Prenatal exposure has been repeatedly shown to affect newborn infants' processing of the prosodic characteristics of native language speech. Little is known about whether and how prenatal exposure affects infants' perception of speech sound segments. Here we simulated prenatal learning of vowels in two virtual fetuses whose mothers spoke (slightly) different languages. The learners were two-layer neural networks and were each exposed to vowel tokens sampled from an existent five-vowel language (Spanish and Czech, respectively). The input acoustic properties approximated the speech signal that could possibly be heard in the intrauterine environment, and the learners' auditory system was relatively immature. Without supervision, the virtual fetuses came to warp the continuous acoustic signal into “proto-categories” that were specific to their linguistic environment. Both learners came to create two categorization patterns and did so in language-specific ways, primarily on the basis of the vowels' first-formant characteristics. Such prenatally formed proto-categories were not adult-like in that they entirely collapsed some of the native-language contrasts. At the same time, the categories reflected features of the adult language in that they were language-specific. These results can inspire future work on speech and language acquisition in real young humans.