We can reduce the data discussed above to a single chart if we display only the means of the stressed and unstressed distributions for each vowel. Obviously, reducing clouds of data to a few points will radically change the picture: the overlap of the distributions will not be shown at all, though as we have seen, that overlap is considerable. At this cost, we are able to summarize the effects of stress on all the vowels.
Notice that so far this presentation of the effects of stress has included clitic words with non-clitic words. Many clitic words are likely to have a phonological schwa in place of the phoneme which occurs in stressed isolated pronunciations. That is, ``that'', when pronounced [æt] should not be classed phonologically with ``that'' when pronounced [t]. Because of possible confusion over what phonemes occur in many clitics items, two charts are shown in Figure : with vowels in non-clitic words only, and with vowels in both clitic and non-clitic words.9.23
The means of stressed (primary or secondary) and unstressed allophones of each vowel class are plotted with an arrow from the stressed mean to the unstressed mean. Dashed arrows signify an insignificant difference while solid arrows are statistically significant (5%, two-tailed t-test, cf. page ). Vowels lacking multiple tokens of both stressed and unstressed allophones in non-clitic word English words are displayed at the mean of the class that does occur, with an apostrophe, `, signifying stress (as with /w, er/), and no mark signifying unstressed (as with ``L'', which stands for syllabic /l/.9.24
I will discuss just the upper, clitics-excluded chart, because it exhibits the clearest pattern. The means of the distributions of all the non-low vowels, /i:, I, , , , U, u:, o:, or, oy/, shift quite precisely in the direction of a single focal point. Although some of these individual effects are insignificant -- at a rather stringent level9.25 -- the fact that all of the arrows follow a consistent pattern suggests that a ssingle, quite significant process applies to all of them. The arrows point fairly accurately in the direction of a single location in F1-F2 space; this suggests the metaphor of a reduction ``target''. This target in LACE corresponds to a vowel quality close to . For these vowels, stress reduction should be described as a single process which applies to all of them in the same way, by shifting them (on average) in the direction of a maximally neutralized vowel, which corresponds closely to the F1-F2 location of .
Exceptions to this aggregate process of shift toward the target  are two: the class of low vowels, and /e:, ir/. /e:, ir/ may be ignored since the effect of stress on them is insignificant. /e:/, for example, is quite widely spread out in the F2 dimension, with stressed allophones as a whole forming a row just below the unstressed allophones. The great variance in the F2 dimension, plus the least amount of skew in the data introduced by other unbalanced factors could account for the fact that /e:/ does not shift in the expected direction.
The reduction process does not apply with such clarity to the class of low vowels shown in the upper chart. This may be due to the paucity of non-clitic data for these vowels (5 unstressed /æ/ tokens, 3 each of unstressed /, r/, none of /w/). They do not show the same degree of inconsistency with the general process in the lower chart, where clitics are included, and these vowels all raise directly upward.
The third exceptional low vowel in the clitics-excluded chart, /y/, has a more traditional characterization: the nucleus is shifting in the direction of the glide. This amounts to increased coarticulation between the nucleus and glide in the unstressed forms of this vowel. (Similar effects also occur with certain gliding vowels in Alabama and Chicago, pages , ).
Does reduction shrink vowel space, or does stress expand it? Some vowel nuclei overlap much more when unstressed than when stressed. A basic structuralist principle says that if two classes merge in some particular context, then the distinction between the two must be present underlyingly, and neutralized in that context, rather than being absent underlyingly, and produced unpredictably in the contexts where it does appear. The underlying, or ``true'', form is the distinct form, while the derived form is the merged one. This principle can be used with /U/ and /u:/, for example, to say that the stressed forms are the ``true'' or underlying targets, while the unstressed forms which appear to be nearly neutralized (in this representation), are derivative. Similarly, all vowels are relatively neutralized if they all shift towards a common reduction target, and thus towards each other. If this interpretation is followed, the correct characterization of phonetic vowel reduction is not the vowel space expands when stress is applied, but that vowels move in the direction of  when stress is reduced. Thus the following rule of phonetic grammar.
When phrasal stress is reduced, the phonetic realizations of vowels shift towards the reduction target.
The grammatical characterization of this process is a subtle matter. The location of the phonetic realization of // is especially significant here. The reduction target may be identical to the phonetic realization of //; certainly it is seductive to simply characterize vowel reduction as shift towards //, though there are hidden complications in such a statement. The shift is actually towards the phonetic realization of //, not towards , which is mid-central, not high-front, like this target, and not towards //, which is a phonological entity (written with / /'s) rather than a phonetic one. Statistical tests that may require even larger quantities of data would be necessary to distinguish the realization of // from the global target of vowel reduction. Such a test would have as a null hypothesis that the two are identical. Until such tests are made, I will assume that the realization of // is in fact identical to the reduction target.
Next, consider the phonological treatment of //. It is analysed as underlyingly absent, or featureless in Phonological Preliminaries. That is, it is unspecified for height, backness, or stress, and may even lack an underlying Nucleus position. Essentially, this analysis states that // is a phonological vowel that is nothing at all (that is, it is constituted by no underlying features, only relatively indistinct phonetic features). So // doesn't actually exist, only its realization does, which is entirely a creation of low-level processes of phonetic implementation, such as processes that insert a Nucleus slot where none is present, in order to satisfy syllable-structure well-formedness conditions.
If the Nucleus slot is inserted in an unstressed syllable, and is filled by no other features9.26, then without further specification, the vowel will be realized with the phonetic quality of the reduction target, which is the only source (besides segments in the adjacent environment) of phonetic vowel quality for this vowel segment. Finally, the ancient, but still hardly confirmed idea of Sievers (1876), that different languages have different rest positions for the vocal tract may be relevant here. The reduction target could be not just the realization of the unspecified vowel, it could also be the rest target. It appears that this target is different from those for other dialects, as summarized in the concluding chapter.
This section has shown a clear pattern of vowel shifting that correlates with phrasally (or lexically) destressing a vowel. The pattern seems to be best explained as a phonetic rule of vowel reduction which shifts nuclei in the direction of the ``reduction target'', a location in F1-F2 space that corresponds to impressionistic , and which we may assume is identical to the target, or mean location, of the phonologically reduced vowel //.