Learning vowels before and after birth: A simulation of how prenatal input may affect perceptual attunement throughout infancy

Abstract

How can discrete representations of speech sounds emerge from continuous phonetic experience? In an earlier study, we asked whether category emergence for speech sounds can begin before birth. Here I ask what happens right after birth, when the learner encounters a much richer speech signal than before, and whether it matters that the learner has experienced prenatal speech at all. I extend our earlier two-layer neural-network to simulate virtual infants exposed to prenatal and postnatal Czech and Spanish vowels with those that begin learning only after birth. Comparing similarities at the learners’ “phonology” layers suggest that prenatal learning may affect subsequent perceptual narrowing, with the prenatally trained network generally developing more differentiated vowel representations compared to the network in which learning does not start until birth. These effects seem to be at least to some extent contrast-specific, and some developmental trajectories suggest U-shaped patterns: some vowels sometimes become similar before differentiating again. I then consider how these simulation outcomes relate to what we know about the development of speech perception in real infants and how the virtual learners could generate predictions about vowel perception in real babies with different prenatal auditory experiences.