The phenomenon that concerns AMR can be illustrated with the German stems Bund- ``brotherhood''and bunt- ``colorful''. If the stems have a suffix, as in Bunde and bunte, then they are pronounced [bUnde] and [bUnte ]. The orthographic difference between the D and T is clearly appropriate, and illustrates one of the most striking and profound properties of human language: the use of discretely distinct sound types for `spelling' lexical items. These phonological types, the D and T, recur in many different vocabulary items and often may appear in a range of syllable positions. However, when the /d/ or /t/ occurs as the last segment in a syllable (that is, for these stems when there is no suffix), the contrast is often said to be `neutralized' since both words are pronounced as approximately [bUnt] with a final [t]. The D/T distinction appears to be lost since both pronunciations (for these words and for many other pairs of voiced and voiceless obstruents in Standard German) merge into a phonetic [t] (or the cognate voiceless stop, fricative or affricate). The traditional linguistic description of this phenomenon is to say that a `neutralization process' has applied that turned the [+voice] feature of the D into the [-voice] feature of the T. In the more recent optimality theory formulation of such data, German would be said to rank certain simultaneous constraints differently than, say, English. But the simultaneous constraints are still stated in terms of standard phonetic features.
The surprising experimental result is that if you do either a production experiment or a perception experiment using a reasonable number of tokens of each minimal pair, you find that the two sets of [t]-like sounds are consistently somewhat different [Mitleb, 1981, Port and O'Dell, 1986, Port and Crawford, 1989]. In fact, essentially this same observation had been independently made in other labs for over twenty years (See papers above for references). Of course, negative results are always easy to obtain: eg, perhaps not all speakers will exhibit difference in the parameters chosen for the experiment, and since the effect is on noisy variables (like vowel and consonant duration), one must clearly have plenty of tokens to study. But statistically significant differences have been demonstrated for production data many times. The perception result is admittedly less well verified. But given the well-attested production differences, it would not seem to be an implausible finding.
So then, production data show small but statistically significant differences in many measurable properties between the two underlying stops. The differences are typically in the direction one might expect from examination of obviously contrastive pairs like Bunde-bunte. Thus the [bUnt] that is related to underlying /t/ has a slightly shorter preceding vowel and slightly longer closure than the [bUnt] related to lexical /d/. Also, there is weaker pulsing during the stop closure and a somewhat stronger release burst on the stops - both of which imply that speakers tend to pull their vocal folds somewhat wider apart for the underlying voiceless stops than for the voiced ones. Of course much overlap remains between the two voicing sets for any of these measurements. But the word pair populations are clearly only slightly different, suggesting that the `underlying' voicing feature is still biasing the phonetic detail of the stops despite the fact that most of the difference between the voicing classes has been wiped away when the stop or fricative occurs in this syllablic position.
But can listeners make any perceptual use of these differences? We found in the 1989 paper that if one records a number of productions of this pair and asks native German listeners to identify them (in a forced choice between, eg,bunt and Bund), you find that - rather than 50% correct identification (indicating mere guessing due to the supposed neutralization) - roughly 60-80% correct identification of the word pair is obtained (depending, of course, on many linguistic and channel variables that affect the rate of correct identifications). But if subjects were instead identifying a typical minimal pair, such as Bunde vs. bunte, in these experimental conditions, they would probably identify them at around 99% correct. So the first thing that is puzzling about these data is that the traditional view about phonetics predicts performance close to either chance (= 50% in this 2-alternative case) or close to 100%. After all, two sounds are either exactly the same or else they must differ by at least one `phonetic quantum'. And one should expect that any difference will be at least reliably detectible. If such a difference were not reliably detectible, then how could children learn their native language? And how could linguists do phonetic transcriptions that capture ``all and only'' the linguistically controllable phonetic differences?
Adding further strangeness to this case of moderate performance is, as AMR notes, that many well-known phoneticians and phonologists (who should be expected to be at least as good at phonetic discrimination as ordinary lay speaker/hearers) have asserted that the German voicing contrast is `neutralized' to some kind of stop that is the same for both Bund and bunt. Not only did the great German phonetician, E. Sievers consider the neutralization to be a complete neutralization, but so did L. Bloomfield (1933), W. Moulton (1962) and many others. The D simply turns into a T when it occrs in final position in a syllable. So AMR is understandably puzzled at the paradox: ``How could this difference be fairly easy for native speakers but seemingly be completely impossible for (or perhaps negligently overlooked by) highly respected linguists and phoneticians?'' And the problem is compounded by claims of similar phenomena in syllable-final devoicing in Russian, Polish, Catalan, etc.
In fact, for me personally, one reason why the whole phenomenon seemed quite plausible is that soon after serendipitously discovering this phenomenon with Fares Mitleb, working on his 1981 dissertation, I noticed a very similar effect within my own speech! For me, as in most dialects of American and British English, a similar partial neutralization is found for the flapped medial /t/ and /d/ at least after certain vowels. Thus, for me, pairs like budding-butting, biddy-bitty and ladder-latter all seem to sound the same (although for some Americans the last pair may exhibit a more noticeable difference in the pronunciation of the [æ] before /d/). Indeed in teaching introductory phonetics classes I had taught students for years that these pairs were neutralized! ``Say them'', I would say, ``You can hear that they sound the same despite the spelling difference. This is neutralization,'' and the students would nod in agreement. So although bud and butt are obviously conrastive, budding and butting seem at first to be indistinguishable.
However, Fox and Terbeek (1977) showed that some Americans produce these two `flaps' with somewhat different timing (the preceding vowel tends to be shorter before the underlying /t/). In unpublished results, my students and I have found that in some dialects (eg, New York City), the contrast appears to be completely neutralized except after certain vowels (like [ay]) while in various other dialects differences may be greater and more widespread across the vowels [Port, 1976, Huff, 1980, Chin, 1986]. With isolated words read aloud from a list, I find it fairly easy to demonstrate in the classroom that American listeners can guess with much better than chance accuracy whether budding or butting was intended (using myself as speaker).
So the English flapping situation and German syllable-final devoicing are rather similar. The primary difference is that in English both the /t/ and /d/ are modified (in the post-stress, intervocalic context) to a third sound, the apical flap, whereas in German the voiced obstruents seem to merge directly into the voiceless ones. Both cases are effective enough as neutralizations that native speakers (like phoneticians) do not immediately notice the difference. And in both cases, there is morphological support for the underlying contrast since pairs like Bund-bunt and budding-butting are only homophonous in certain morphological contexts. Again, in both languages there exist cases where there is no basis beyond orthography for choosing the underlying consonant (cf. German und and English water).
In the next section I take a closer look at the traditional view of phonetics as presented by Chomsky and Halle, 1968 in order to show why this raises serious theoretical issues.