Languages differ in their phonotactics, that is, in the segments and segment sequences that are permissible. Therefore, sensitivity to phonotactics must be learned from the input infants receive. I will use behavioral looking time experiments to show that sensitivity to phonotactics is induced by 5-months (n=98).

There are three proposals for how infants learn phonotactics. In the prelexical hypothesis, infants learn phonotactics from unsegmented utterances in the speech stream. In the proto-lexical hypothesis, infants learn phonotactics from the word forms they initially extract from the speech stream, but do not yet associate with referents. In the lexical hypothesis, infants learn phonotactics from words they know and associate with referents. I will use logistic regression with k-fold cross validation to evaluate all three proposals. The models embodying the proposals will be benchmarked against the results from the infant experiments to get at critical components of a mechanism for phonotactic acquisition.

More generally, I will argue that an approach evaluating models against developmental benchmarks is necessary to distinguish multiple plausible proposals that achieve similar performance despite encoding differing assumptions about cognitive representations, properties of the data attended to by the learner, or learning mechanisms across domains and disciplines.