Are phonetic units discretely different from each other? Can careful listeners identify speech sounds reliably? These are the fundamental issues raised by Prof. Manaster-Ramer (AMR) in his letter to the editor [Manaster-Ramer, 1996]. The author is concerned about a phenomenon reported a number of times in the phonetics literature over the past decade. He is alarmed apparently that these results may undermine the methodological and theoretical basis of phonology. In this essay, I will first review the phenomenon itself and agree with AMR that `incomplete neutralization' is just a slight variant of many other familiar phenomena from experimental phonetics and phonology, and that it is troubling. Many widely attested phenomena offer strong evidence that the set of sounds of human speech, that is, the universal inventory of phonetic categories, cannot be assumed to fall into discretely distinct types that can be reliably differentiated by a first language learner and by a trained linguist or phonetician. Of course, often speech sounds are obviously different (eg, the initial stops in bad vs. pad or the vowels in bad vs. bed). But can they always be assumed to be discretely different from each other and sufficiently distinct that a careful listener could hear the difference? Traditional linguistic phonetics says yes, but I think the evidence makes this very unlikely. That is, if you ask regarding two short sound bites ``Are these two instances of the same sound or different ones?,'' one simply cannot assume that there will always be a correct or consistent answer. But if this is so, then one cannot use phonetics to explain the discreteness of bad and pad either. The evident discreteness of phonology will need to be accounted for in some other way.
AMR is quite correct to be concerned about the issues of the Failure of Phonetic Discreteness and the Failure of Auditory Identifiability - about the fact that neither a linguist nor a native speaker can necessarily identify discretely and confidently which phonetic elements they observe in a sample of speech. This failure raises distressing questions about how the discipline of linguistics should obtain data regarding the phonological systems of human language. The phenomenon of incomplete neutralization is a worrisome chink in the dam that supports all of current symbol-based phonological theory - just as AMR seems to fear.
Many phonologists prefer to act as though one could assume confidently that the process of careful listening followed by phonetic transcription (at least when performed by a professional) `makes available' to the phonologist a representation of all possible linguistically relevant aspects of a sample of speech. From this discrete symbolic description, the phonologist constructs various data structures, from phonological segments to syllabic trees to metrical grids. Without this discreteness and the presumed positive identifiability of these units, phonology (as well as the language-learning child) would appear to have no place to begin work. How could phonology account for symbolic structures like language-specific consonant and vowel types, allophones, stress markings, constraint hierarchies, etc.? Where could these complex cognitive objects possibly come from?
Chomsky and Halle (1968) offered a simple answer. They proposed that the universal phonetic space is a discrete alphabet. There is some list of features that is the maximum set of linguistically relevant sound types in languages of the world. All members of our species are born with the ability to reliably and almost effortlessly identify these perceptual objects. These static features serve as units for language learning children and constrain the possible speech sounds of language such that they take the form of a two dimensional matrix of feature values versus discrete time. (Although C-H asserted that the features might have continuous values rather than discrete ones, linguists invariably normally treat them as though they had discrete, quantal values.) Thus, surprisingly, the data structures of phonological descriptions inherit their discreteness from the phonetic atoms of which they are constructed. A fundamental discreteness is built into linguistic ontology by Chomsky and Halle. Whatever the physical world might be like, it is assumed that as far as language is concerned, nothing other than discrete, abstract phonetic objects need to be the concern of linguists. This premise seems to be closely related to the Cartesian view that the human mind is quite a different sort of entity from the physical world.
For phoneticians, on the other hand, the failure of phonetic discreteness is in fact rather familiar. Phoneticians have long observed that phonetic objects are discretely different only in ideal circumstances [Lisker and Abramson, 1971, Klatt, 1976, Keating, 1985, Lindblom, 1983]. However, although AMR is concerned about the partial neutralization problem and suggests that great amounts of experimental work might have to precede any linguistic analysis, he does not seem to put his finger on the magnitude of the theoretical threat. In this essay, I will try to clarify what is at issue and why it is important. Finally I will suggest how a theoretical solution to the problem of phonetic discreteness may be found.