We argue that infants perform an abstraction over their auditory representations of the vowels of individual speakers, by mapping them to a mediating space of speaker-independent representations, guided by vocal imitative interaction with their caretakers, as a first step in the phonological acquisition process. Furthermore, we proffer a methodology for modeling this abstraction which involves the alignment of the cognitive structures, or manifolds, that the infant builds from the auditory representations of the vowels of individual speakers. As a demonstration of the methodology, we show that higher-dimensional "excitation pattern" representations facilitate modeling of the influence of the imitative process on perception and abstraction more so than lower-dimensional formant representations.

Index Terms: vowel normalization, manifold alignment, vocal imitation, phonological abstraction