The analysis of phonetic systems largely operates independently of other linguistic systems. This is despite the fact that the critical components of the system—phones and the contrastive relations between them—are definitionally linked to higher-order systems such as the mental lexicon. In this paper we outline an alternative approach where the phonetic system is studied as embedded within these higher-order systems. We show that system embedding radically changes the assumed role of different contrasts in the system, and warps estimates of the relative weight of the corresponding acoustic dimensions that delineate that system. Specifically, we compare three models of obstruent discrimination in English: (1) the canonical, inventory model, where contrasts are studied between phones in controlled syllables balanced in weight; (2) the lexicon model, where all contrasts are between real words; and (3) an intermediate, weighted inventory model, where the acoustics are derived from controlled syllable data but the items are sampled to match the distribution of contrasts in the lexicon. By comparing these three models we are able to identify discrepancies in the role of each acoustic dimension under different system assumptions, and model their impact on phonetic generalizations.

This study used visual-world eye tracking to examine the effect—first observed in Chan and Vitevitch (2009)—of the phonological neighborhood clustering coefficient on the time course of lexical access in spoken word recognition. Target words from neighborhoods with relatively high clustering (i.e., neighbors of the target word are also neighbors of each other) showed a significant lag in eye fixations relative to words from less clustered neighborhoods after controlling for neighborhood density, target frequency, neighborhood frequency, and multiple phonotactic probability measures. This effect was also influenced by lexical frequency, neighborhood density, and neighborhood frequency, suggesting that neighborhood clustering influences spoken word recognition, and should be included in models of this process.