Analytic models in phonology

0 Introduction The goal of this paper is to describe a phonological framework in which the study of discrete 'phonological' units appears not as a separate enterprise, opposed to the study of continuous 'phonetic' phenomena, but rather as part of the larger enterprise of the scientific study of speech. Why is such a framework necessary? Given that the sound structure of language is a structure composed of discrete units defined in terms of contrast, would it not be the task of the phonetician to deal with all aspects of speech that are neither discrete nor contrastive? Unfortunately , phoneticians are not ready to pick up where the phonologists leave off and specify in detail how the discrete units used in phonology can be realized in, and recovered from, the undifferentiated continuous data provided by acoustic waveforms or articulatory records. In fact phoneticians have good reasons to be skeptical about the feasibility of the tasks imposed on them by the internal logic of phonology. For example, phonologists usually work with idealized data that preserves dialectally and grammatically conditioned variation but suppresses variation within the speech of a single individual and across individuals sharing the same dialect/sociolect. Perhaps the human apparatus for speech perception can somehow magically filter out variability along certain dimensions but not along others, but phoneticians have not succeeded in creating models that are capable of the kind of selective filtering presupposed in phonology, nor do they necessarily believe that humans actually perform such feats. The net result of the discrepancy between what phonologists assume phoneticians can do and what phoneticians actually do is that phonology depends on an unspecified and perhaps unspecifiable black box as its primary data-gathering instrument. The result of this situation is that phonological theory is of surprisingly little use to anybody outside the self-imposed boundaries of the discipline. Phoneti-cians, acousticians, and other scientists and engineers interested in speech for theoretical or practical purposes generally find the rule or constraint systems devised by phonologists to be brittle in the sense that slight changes in the data, such as the discovery of a related set of forms, bring about radical revisions of the grammar, fragmentary in the sense that what constitutes relevant data for