| dc.description.abstract | In the canonical analysis of the acoustic and perceptual structure of phonetic systems, experiments and modeling procedures are based on a balanced inventory of phones whose characteristics and relational structures are largely treated as independent of the wider system of higher-order distinctions in which they occur. The present study presents an alternative, lexical approach wherein the fundamental unit of phonetic analysis is rather a relation between contrastive words, thus shifting the state space of the system from a small, largely symmetric inventory of phones, to a large, heterogeneous ensemble of real-word contrasts in the lexicon. The ultimate goal in developing such an approach is twofold. First, by directly linking the study of phonetic systems to the higher-order units they encode, phonological and acoustic asymmetries, as well as interactions with top-down information, are implicit in the model, thus reducing the dependence of phonetic analyses on phonological inventory definitions, and providing for more scalable estimates of phonetic characteristics to real-word production. Second, the comparison of a lexical approach to the canonical inventory approach provides an explicit test of the consequences of the inventory assumption for models of cue integration and the acoustic/perceptual structure of the phonetic system. Using the English obstruent system as a test case, several discrepancies in the acoustics and perception of obstruent contrasts in the two systems, as well as the link between the two in the form of models of cue integration, were uncovered. The data used as the basis for acoustic analyses and cue-integration models in the lexicon and inventory was based on a large sample of real words and controlled syllables, respectively, produced by a single native English speaker. Perception data came from two experiments measuring listener word recognition of minimal-pair contrasts in noise. For example, the presence of voicing in the signal is a more robust indicator of word-final voicing contrasts in controlled syllable data than in real words, whereas coarticulatory cues to obstruent place of articulation, such as in F2 at vowel onset/offset, are more robust among lexical contrasts. Perceptually, due to asymmetries in phonological distributions in the lexicon, the overall accuracy of listeners in word recognition depends primarily on their recognition of the plosives [p, t, k, b, d] and the fricatives [f, v, s, z], where the sibilants are generally robust in perception, while the plosives and labial fricatives are a common source of errors. When directly linked in statistical models of cue integration in the inventory and lexicon—both of ideal discrimination and recognition patterns by human listeners—points of disagreement in cue weights can be classified into distributional, acoustic, and composite sources. Examples of distributional disagreement include the downweighting of relative F3 amplitude in the inventory model due to the reduced prevalence of sibilance contrasts, while acoustic disagreements include cases such as the poor scaling of vowel-offset F2 due to coarticulatory differences between controlled syllables and real words. Composite disagreement arises from a combination of these two sources. Finally, the structure of the system of obstruent contrasts in the lexicon was studied by measuring the response of several measures, including minimal pair count, functional load, edge disjoint strength, and average path length, to noise- and cue-perturbation. Overall, the impact of background noise on plosive and sibilance contrasts has the greatest impact on the system, meaning cues such as noise duration, spectral tilt, F1, and relative F3 amplitude are the most influential in maintaining lexical distinctions. The combination of these results challenges the viability of the independence assumption implied by the study of phonetic inventories, and provides empirical motivation for the study of phonetic systems as linked to the higher-order structures they encode. | |
