Distributional learning of vowel categories is supported by prosody in infant-directed speech

Infants ’ acquisition of phonetic categories involves a dis-tributional learning mechanism that operates on acoustic dimensions of the input. However, natural infant-directed speech shows large degrees of phonetic variability, and the resulting overlap between categories suggests that category learning based on distributional clustering may not be feasible without constraints on the learning process, or exploitation of other sources of information. The present study examines whether mothers ’ prosodic modifications within infant-directed speech help the distributional learning of vowel categories. Specifically, we hypothesize that ‘motherese’ provides the infant with a subset of high-quality learning tokens that improve category learning. In an analysis of vowel tokens taken from natural mother-infant interactions, we found that prosody can be used to distinguish high-quality tokens (with expanded formant frequencies) from low-quality tokens in the input. Moreover, in simulations of distributional learning we found that models trained on this small set of high-quality tokens provide better classification than models trained on the complete set of tokens. Taken together, these findings show that distributional learning of vowel categories can be improved by attributing importance to tokens that are prosodically prominent in the input. The prosodic properties of motherese might thus be a helpful cue for infants i