in Hsu, Chai Shune (ed.) (1996) UCLA Working Papers in Phonology 1.
Evidence for Constraints on Contrast: The Dispersion Theory of Contrast
Edward Flemming
Stanford University
1. Introduction
The thesis of this paper is that there are constraints on the well-formedness of phonological contrasts.
Specifically, the selection of phonological contrasts is argued to be subject to three functional goals:
(1)
i. Maximize the number of contrasts
ii. Maximize the distinctiveness of contrasts
iii. Minimize articulatory effort
These goals derive from language’s function as a means for the transmission of information. The
number of phonological contrasts should be maximized in order to enable us to differentiate a substantial
vocabulary of words. The auditory distinctiveness of the contrasts should be maximized so that the
differences between words can easily be perceived by a listener. The third requirement, that effort should
be minimized appears to be a general principle of human motor behaviour, and is not specific to
language.
This theory of contrast is dubbed the ‘Dispersion Theory’ after Lindblom’s (1986, 1990) ‘Theory of
Adaptive Dispersion’, which it resembles in many respects. Other important antecedents are Zipf (1949)
and Martinet (1952, 1955).
Dispersion theory has a number of important consequences for phonological theory. Firstly, the
constraints in (1i-ii) are constraints on the relationships between contrasting forms. The existence of such
constraints implies that the well-formedness of a word cannot be evaluated in isolation, it must be
evaluated with reference to a set of forms that it contrasts with. This point will be developed in detail in
the course of the paper. Secondly, distinctiveness of contrasts is an auditory-acoustic notion, so
formulating the requirement that the distinctiveness of contrasts be maximized requires reference to
auditory representations of forms. Arguments for this position are developed at length in Flemming
(1995). In this paper, we will assume this position, but note that many of the phenomena analyzed would
be difficult to account for in terms of purely articulatory representations.
2. The Dispersion Theory of Constrast
An important property of the goals formulated in (1) is that they conflict. This point can be illustrated
by considering the selection of contrasting sounds from a schematic two dimensional auditory space,
shown in figure 1. Figure 1a shows an inventory which includes only one contrast, but the contrast is
maximally distinct, i.e. the two sounds are well separated in the auditory space. If we try to fit more
sounds into the same auditory space, the sounds will necessarily be closer together, i.e. the contrasts will
be less distinct (figure 1b). Thus the goals of maximizing the number of contrasts and maximizing the
distinctiveness of contrasts inherently conflict. Minimization of effort also conflicts with maximizing
distinctiveness. Assuming that not all sounds are equally easy to produce, attempting to minimize effort
reduces the area of the auditory space available for selection of contrasts. For example, if we assume that
sounds in the periphery of the space involve greater effort than those in the interior, then, to avoid
effortful sounds it is necessary to restrict sounds to a reduced area of the space, thus the contrasts will be
less distinct, as illustrated in figure 1c. Note that while minimization of effort and maximization of the
number of contrasts both conflict with maximization of distinctiveness, they do not directly conflict with
each other.
1
E. Flemming
(a)
Two segments
Most separation
More effort
(b)
(c)
Four segments
Four segments
Less separation Least separation
More effort
Less effort
Fig 1. Selection of contrasts from a schematic auditory space.
Since the requirements on contrasts conflict, the selection of an inventory of contrasts involves
achieving a balance between them. A source of cross-linguistic variation is variation in the compromise
that different languages adopt.
Section 3 presents a formalization of dispersion theory in terms of Optimality Theory (Prince and
Smolensky 1993). Optimality theory is suitable for this purpose, because it provides a system for
specifying the resolution of conflict between constraints. Section 4 provides evidence for the model.
3. A Formal Model of Dispersion
According to dispersion theory, determining the inventory of contrasts in some context involves
selecting sounds in accordance with the constraints on contrasts outlined above. To formalize this model,
we need to provide auditory representations, and formulations of the constraints on contrast. We will
consider these components of the model in turn.
3.1. Auditory Representations
The auditory representation must include all auditory properties of a sound relevant to phonological
patterning. What details are relevant is a matter for empirical investigation, but the required level of
detail is probably quite considerable (Flemming 1995, Steriade 1995). So the features proposed here are
not intended to be a complete set, they are simply sufficient for the analyses in this paper. Further study
will undoubtedly reveal the need for enrichment of the representations.
We will represent sounds as located in a multi-dimensional auditory space. Examples of dimensions
postulated here are given in (2). Note that, given the definitions offered here, these dimensions could be
regarded as acoustic rather than auditory. The representations are labelled as auditory to emphasize the
fact that it is distinctiveness to the human ear that is relevant to language, and that these properties are
established to be perceptually significant. Also, as the theory is developed it is expected that the
particulars of the processing of sound by the peripheral auditory system will be found to be relevant.
(2)
Formant frequency (F1, F2, F3) Frequencies of formants
Noise frequency
Frequency of amplitude peak in noise spectrum
Diffuseness
Spectral diffuseness
Noise Intensity
Intensity of noise in the spectrum
Intensity
Overall intensity
VOT
Voice onset time
These dimensions are scalar, but we will decompose them into binary features in a manner familiar
from the standard treatment of vowel height, in which two binary features are used to define three
degrees of height. For example, (3) shows the decomposition of the first formant frequency (F1)
dimension into four features, distinguishing five levels of F1. This dimension essentially corresponds to
vowel height. An example vowel is given for each level of F1.
2
Evidence for Constraints on Contrast: The Dispersion Theory of Contrast
(3)
lowest F1
low F1
high F1
highest F1
i
+
+
-
I
+
-
e
+
-
E
+
-
a
+
3.1.1. Auditory Distinctiveness
Representations of this type, with scalar dimensions decomposed into binary features, allow us to
refer both to classes of sounds and to auditory distance between sounds. Binary features
straightforwardly represent class membership: a sound either is or is not [+low F1]. However, since the
features define a scale, they also represent auditory separation between sounds on a dimension: the
distance between two sounds is the number of features in which they differ. For example, [i] and [a]
differ in their specifications for all four F1 features, thus they are separated by a distance of four on the
F1 dimension. The sounds [i] and [I], on the other hand, differ in only one feature, [lowest F1], so the
distance between them on the F1 dimension is only one.
A representation of auditory distance is required to evaluate the distinctiveness of contrasts.
3.1.2 The Dimensions
Formally, auditory representations consist of two types of elements: dimensions and features.
Dimensions take as their values specifications of the set of features that constitute that dimension (4). In
practice, as long as the features are given mnemonic names, it is redundant to specify the dimension
name in addition to the features constituting that dimension, so (5a) may be written as in (5b).
(4)

+D1 
Dimension:
+D2 

 : 

(5)
a.
b.
+lowest F1
+low F1  
 F1: -high
 -highestF1F1 
F2
  -lowest

-low F2 
F2: +high F2 
 +highest F2


:
+lowest F1
+low F1 
 -high
 -highestF1F1 
F2
 -lowest

-low F2
+high F2 
+highest
 : F2
Not all dimensions are relevant to all sounds. Only sounds with a relatively open vocal tract (vowels,
approximants, and nasals) have well-defined formants, so these dimensions will be specified for these
sounds, and not for fricatives and stop closures. Similarly, not all sounds involve a significant noise
component, so not all sounds will be specified for Noise frequency.
Individual dimensions will be discussed further as they are required for analyses.
3.2. Constraints on Contrasts
We need to formalize constraints favouring the goals outlined in (1) above (repeated in (6)).
(6)
i. Maximize the number of contrasts
ii. Maximize the distinctiveness of contrasts
iii. Minimize articulatory effort
Optimality Theoretic models achieve optimization without numerical calculation by adhering to a
requirement of strict constraint dominance, i.e. where two constraints conflict, the higher-ranked
3
E. Flemming
constraint prevails (Prince and Smolensky 1993:78). In the dispersion theory, assigning complete
dominance to any one of the proposed fundamental constraints yields inappropriate results. For example,
if maximization of the number of contrasts dominates, the result will be a huge number of very fine
contrasts. The essence of the dispersion theory is that the conflicting goals are balanced against each
other.
The balancing of conflicting scalar constraints can be modelled in terms of strict dominance by
decomposing the scalar constraints into a ranked set of sub-cases. This technique is adopted by Prince
and Smolensky (1993) in the analysis of syllable structure, where a general constraint requiring a syllable
nucleus to be maximally sonorous is decomposed into a set of constraints against particular segments
being in the nucleus, with the sub-constraints being ranked according to the sonority of the segments. The
sub-constraints corresponding to the scalar constraints can then be interleaved, resulting in a balance
between them.
i. Maximize the number of contrasts
We can implement the goal of maximizing the number of contrasts in terms of a ranked set of
constraints, each requiring the maintenance of a particular number of contrasts (7). The fixed ranking
ensures that failure to maintain n contrasts is penalized more severely than failure to maintain n+1
contrasts. I.e. the smaller the number of contrasts maintained, the greater the violation. This encodes the
fact that the number of contrasts should be maximized.
(7)
Maintain 1 >> Maintain 2
contrast contrasts
>> ... >> Maintain n
contrasts
ii. Maximize the auditory distinctiveness of contrasts
The requirement that the auditory distinctiveness of contrasts should be maximized can be
decomposed into a ranked set of constraints requiring a given minimal auditory distance between
contrasting forms (8). Minimum distance requirements are specified independently for each auditory
dimension, thus the constraint ‘MindistF1 = 2’ states that sounds that contrast in F1 should differ in at
least two F1 features.
(8)
MindistF1 = 1 >> MindistF1 = 2 >>... >> MindistF1 = n
To encode the fact that auditory distinctiveness should be maximized, Mindist F1 = n is ranked above
MindistF1 = n+1, i.e. the less distinct the contrast, the greater the violation.
Balancing the requirements on contrasts
The language-specific balance between the constraints on contrasts is modelled by interleaving them.
For simplicity we will consider contrasts on independent auditory dimensions separately. That is, in
addition to specifying minimal distance requirements by dimension, we will adopt separate sets of
‘Maintain contrast’ constraints for each dimension. This allows us to consider the contrasts on a given
dimension more or less independently of contrasts on other dimensions. Adopting this approach has clear
practical benefits: The alternative would be to consider the contrasts maintained on all dimensions
simultaneously, and this is not generally feasible. On the other hand, there is some evidence for
interactions between dimensions that cannot be straightforwardly analyzed given the assumption of
dimensional independence (Flemming 1995, ch.3).
This system can be exemplified with respect to the selection of F1 (‘vowel height’) contrasts. The
table in (9) shows inventories of contrasting vowel heights, and which constraints in the ‘Maintain F1
contrasts’ and ‘MindistF1’ hierarchies they violate. We are considering constraints on contrasts, so the
candidates evaluated here are sets of contrasting forms rather than outputs for a given input. The conflict
between these two constraint hierarchies is readily apparent in (9): Sets of vowel-height contrasts which
violate fewer ‘Maintain contrast’ constraints incur worse violations of the ‘Mindist’ constraints.
4
Evidence for Constraints on Contrast: The Dispersion Theory of Contrast
(9)
Maintain 1
F1 contrast
Maintain 2
F1
contrasts
*
Maintain 3
F1
contrasts
*
*
Maintain 4
F1
contrasts
*
*
*
MindistF1
=1
MindistF1
=2
MindistF1
=3
MindistF1
=4
**
**
***
**
***
i-a
i-e-a
i-e-E-a
i-a
i-e-a
i-e-E-a
These constraints can be interleaved as illustrated in (10).
(10)
Maintain 1 >> MindistF1 = 2 >> Maintain 2 >> MindistF1 = 3, Maintain 3
F1 contrast
F1 contrasts
F1 contrasts
This ranking yields two F1 contrasts (i.e. three distinct vowel heights), as shown in the tableau in
(11). The winning candidate, with three vowel heights, maintains two contrasts, but violates the
requirement that sounds contrasting in F1 should differ in three F1 features. However if we attempt to
satisfy this constraint, maximizing distinctiveness, as in the second candidate, we violate the higherranked requirement that we maintain two F1 contrasts. It is not possible to fit two contrasts each
separated by three features on the F1 dimension, as specified in (3) above. The winning candidate also
violates ‘Maintain 3 F1 contrasts’. If we try to satisfy this constraint, maximizing the number of contrasts
as in the 3rd candidate, we violate the high-ranked constraint ‘MindistF1 = 2’ since e-E and E-a differ in
only one F1 feature.
(11)
Maintain 1
F1 contrast
> i-e-a
i-a
i-e-E-a
MindistF1
=2
Maintain 2
F1
contrasts
MindistF1
=3
**
*!
*!*
Maintain 3
F1
contrasts
*
*
***
Thus the particular balance achieved here between maximizing the number of contrasts and
maximizing the distinctiveness of the contrasts yields two moderately distinct contrasts. Altering the
ranking results in a different balance. For example, if greater weight is given to maximizing
distinctiveness, ranking ‘MindistF1 = 3’ above ‘Maintain 2 F1 contrasts’, the winning candidate has just
one maximally distinct F1 contrast (12).
(12)
Maintain
F1 contrast
i-e-a
>
i-a
i-e-E-a
MindistF1
=2
MindistF1
=3
Maintain 2
F1
contrasts
*!*
*
*!*
Maintain 3
F1
contrasts
*
*
***
It is apparent that not all combinatorially possible rankings of these constraints on contrast
correspond to possible languages. The balance between maximization of the number of contrasts and
maximization of the distinctiveness of contrasts is determined by the relative ranking of ‘Maintain
5
E. Flemming
contrast’ and ‘Mindist’ constraints. If all definable rankings were possible, we would expect to find
languages which value the number of contrasts very highly, resulting in a huge number of very fine
contrasts, and languages which value distinctiveness very highly, resulting in a handful of maximally
distinct contrasts. Neither of these extremes is attested. It seems that there is a lower bound on the
distinctiveness required for a contrast to be functional, and that there is an upper bound beyond which
additional distinctiveness provides a poor return on the effort expended1.
iii. Minimization of effort
So far we have not considered minimization of effort. No general account of the effort involved in
speech production will be proposed here, we will simply adopt specific constraints which are plausibly
motivated by this consideration, such as ‘Don’t voice obstruents’ and ‘Don’t have short peripheral
vowels’.
3.3. Comparative Constraints and the Evaluation of Individual Forms
The tableaux in (11) and (12) illustrate how the proposed constraints govern the selection of
permissible contrasting surface forms, but we have not yet considered the evaluation of individual words.
In standard optimality-theoretic analyses, the constraints evaluate candidate outputs for a given input.
The proposed constraints on contrasts are obviously applicable to this mode of analysis: a given output
form is acceptable only if it consists of acceptable contrasts. This is illustrated in the tableau in (13). This
tableau shows the evaluation of candidate outputs for the underlying form /pit/, considering only the F1
of the vowel. The first point to note is that even though we are concerned with the evaluation of the
output of a single form, in order to evaluate its well-formedness with regard to the constraints on contrast
it must be evaluated with reference to a set of contrasting forms. This is because the ‘Maintain contrast’
and ‘Mindist’ constraints are inherently comparative: i.e. they refer to properties of sets of forms, rather
than to properties of individual forms. ‘Mindist = d’ is a constraint on the auditory distance between
forms contrasting on some dimension, and ‘Maintain N contrasts’ is a constraint on the size of the set of
contrasting forms.
Thus the candidates in (13) consist of sets of contrasting forms, as in the tableaux above, but one of
these forms (underlined in the tableau) is specified to be the output corresponding to the input underlying
representation. In effect, candidates consist of an output form, and a comparison set in relation to which
the form is evaluated.
(13) /pit/
> pit pet, pat
pit pat
pit pIt, pet, pat
Maintain
F1 contrast
MindistF1
=2
Maintain 2
F1
contrasts
MindistF1
=3
**
*!
*!*
Maintain 3
F1
contrasts
*
*
***
The constraint-ranking developed so far is insufficient for the analysis of input-output relations. The
constraints considered so far only determine acceptable surface contrasts, they place no constraints on the
relationship between underlying representation and the output form. This point is illustrated in the
tableau in (14). The candidates here all consist of the same set of contrasting forms, but in each case a
1A similar
issue arise with respect to faithfulness constraints of the type posited by Prince and Smolensky (1993). If
all faithfulness constraints are at the top of the ranking then all inputs will surface as well-formed outputs - i.e. this
ranking would yield an unattested language with no restrictions on the form of words. Conversely, if all faithfulness
constraints were at the bottom of the ranking, then all inputs would be mapped to a single, maximally well-formed
output (presumably the null output).
6
Evidence for Constraints on Contrast: The Dispersion Theory of Contrast
different form is designated as the output corresponding to input /pit/. Since the sets of contrasts are the
same in each case, all the candidates are evaluated as equal by the constraints on contrast.
7
E. Flemming
(14) /pit/
Maintn
F1 contrast
MndstF1
=2
Maintn 2
F1
contrasts
MndstF1
=3
> pit pet, pat
pet pit, pat
pat pit, pet
**
Maintn 3
F1
contrasts
*
**
*
**
*
This is not the desired result - obviously, the optimal candidate should be the one in which input /pit/
is realized as [pit] in the output. To achieve this result we need constraints on the relation between input
and output. In standard OT, input-output relations are governed by the faithfulness or correspondence
constraints. These constraints can be formulated as follows:
(15)
‘don’t delete features from the input’ (cf. P&S’s ‘Parse’)
‘don’t insert features that are not present in the input’ (cf. P&S’s ‘Fill’)
The sum effect of these two types of constraints is to require the output to be as similar as possible to
the input. This is essentially the additional constraint that we need to analyze outputs for individual
forms. As shown in (16), faithfulness constraints will favour the candidate in which the output is most
similar to the input, i.e. where the output is [pit]. However there are problems with adopting the standard
formulation of faithfulness. Specifically, as we shall see, the ‘Parse’ constraints partially duplicate the
function of ‘Maintain contrast’ constraints.
(16) /pit/
Maintn
F1
contrast
MndstF1 Maintn 2
F1
=2
contrasts
> pit pet, pat
pet pit, pat
pat pit, pet
{ ‘Faithfulness’ }
MndstF1 Maintn 3 Parse
Parse
F1
[low
F1]
[hi F1]
=3
contrasts
**
*
**
*
*!
**
*
*!
*
The faithfulness constraints as formulated in (15) play two conceptually distinct roles in phonology:
(i) ensuring that allomorphs of a morpheme are similar, and (ii) specifying contrasts.
The standard analysis of similarities between allomorphs involves proposing a unique underlying
form of the morpheme from which all surface allomorphs are derived, as exemplified in (17). A second
component of the analysis must be some requirement that outputs be similar to inputs, otherwise an
output need bear no resemblance to the input, and derivation from a common underlying form would in
no way guarantee allomorphic similarity. Faithfulness fulfills this function.
(17)
[œ!|´m] [´tÓA!m-Ik]
 

/œtAm/
‘atom’

[-s][-z][-´z]

/-z/
‘(pl.)’
The ‘Parse’ constraints also serve to specify contrasts in a language, as demonstrated by Kirchner
(1995). Essentially, a constraint ‘Parse F’ favours preserving underlying differences - i.e. if the input
contains [+F], the output should contain [+F], if the input contains [-F], the output should contain [-F].
8
Evidence for Constraints on Contrast: The Dispersion Theory of Contrast
So, if ‘Parse F’ is satisfied, an underlying difference between [+F] and [-F] is preserved on the surface
and we have a contrast in F.
In the dispersion theory, this second function of ‘Parse’ constraints is taken over by the ‘Maintain
contrast’ constraints. What remains to be provided is an account of allomorphic similarity. That is, we
need to posit some constraints favouring similarity between input and output in order to derive the
correct output in a case like that shown in (16) above. However, if we adopt a formulation of input-output
relations in terms of ‘Parse’ constraints, we will have an inconsistent model in which there are two
sources of contrast maintenance: ‘Maintain contrast’ and ‘Parse’.
In Flemming (1995) I propose to resolve this problem by adopting an analysis of allomorphic
similarity in terms of a direct requirement of similarity between surface forms of a morpheme (cf.
Burzio’s (1995) ‘Anti-allomorphy’ constraint). However, the details of this proposal introduce many
issues which are extraneous to the current discussion, and in the cases considered here the issue of inputoutput correspondences is not crucial, so we will not discuss this issue further in this paper.
4. Evidence for the Dispersion Theory
In this section we will show how dispersion theory accounts for phonological phenomena relating to
contrast, and provide evidence favouring these analyses over previous accounts.
4.1. Inventory Structure
Inventory structure is usually described in terms of cooccurrence constraints. For example the
absence of front rounded and back unrounded vowels from the inventory in (18) would be accounted for
in terms of the pair of cooccurrence constraints in (19).
(18)
(19)
i
e
a
u
o
*[-back, +round], *[+back, -round]
According to the Dispersion theory, inventory structure is the result of a compromise between
maximization of the number of contrasts, maximization of the distinctiveness of contrasts, and
minimization of effort. Thus the analysis of the vowel inventory in (18) is that this language prefers a
large minimal distance for F2 contrasts over multiple F2 contrasts (21). Front and back vowels differ
primarily in the frequency of the second formant, with front vowels having a high F2, and back vowels
having a low F2. Lip-rounding tends to lower F2, so the maximally distinct F2 contrast is between front
unrounded and back rounded vowels, as indicated in the F2 feature specifications in (20). Inclusion of
front rounded vowels or back unrounded vowels would result in sub-optimal F2 contrasts such as [y-u]
and [i-¨]. Note that this relationship between backness and rounding receives a straightforward analysis
in auditory terms, but would hard to account for in terms of articulatory representations.
(20)
(21)
F2:
highest F2
high F2
+
low F2
lowest F2
i
+
+
-
y
-
È
-
¨
+
-
u
+
+
Maintain F2 contrast >> MindistF2 = 4 >> Maintain 2 F2 contrasts
These two accounts of inventory structure make very different predictions. Cooccurrence constraints
imply that markedness is a property of segments, not contrasts. They forbid cooccurrence of a particular
combination of features within a single segment. No restrictions are imposed on contrasts per se, so the
markedness of a contrast should depend simply on the markedness of the contrasting terms. On the other
hand, dispersion theory claims that markedness is a property of contrasts as well as of segments. Given
9
E. Flemming
the goal of maximizing the distinctiveness of contrasts, a contrast is marked to the extent that it is not
auditorily distinct. Individual segments are marked to the extent that they are articulatorily difficult to
produce.
The analysis of (18) in terms of cooccurrence constraints claims that front rounded vowels are
inherently marked, whereas the dispersion theory claims that it is the contrast between front rounded and
back vowels that is marked because it is less distinct than a contrast between a front unrounded vowel
and a back vowel.
This prediction of the dispersion theory is confirmed: the markedness of a sound depends on the
contrasts it enters into. The evidence involve cases in a which a sound is common in the absence of
contrast but uncommon in contrasts. A straightforward example involves dental and alveolar stops. Both
dental and alveolar stops are cross-linguistically common, and thus would appear to be unmarked
segment-types, but alveolars rarely contrast with dentals. We cannot account for these facts in terms of
cooccurrence constraints: both sounds involve unmarked feature combinations, so the contrast between
them should be unmarked. In terms of dispersion theory a contrast between an alveolar and a dental stop
is marked because these sounds are auditorily similar, but each individually can form distinct contrasts
with other stops such as labials and velars.
A second case of this type involves central vowels. Central vowels are relatively uncommon in frontback contrasts: most languages contrast front and back vowels, but comparatively few contrast front,
central and back vowels. But in the absence of front-back contrasts, central vowels are the unmarked
case, other things being equal. So-called ‘vertical’ vowel systems without backness contrasts are found in
the Caucasian languages (Catford ms, Choi 1992a), Margi (Maddieson 1987), Marshallese (Bender 1968,
Choi 1992b). In all cases such vowel systems consist of high and low or high, mid and low central
vowels, with contextual variation. The nature of this contextual variation can be illustrated from
Kabardian, a Caucasian language studied acoustically by Choi (1992a). The realizations of the vowels in
different environments are summarized in (22), with approximate vowel transcriptions.
Crucially, there are no vertical vowel inventories containing invariant [i] or [u], which are otherwise
ubiquitous (23).
(22) palatal
glide [j]
i
front
(23)
palatal C,
alveo-pal,
pal-alveolars
È5, ´5, a5
partially
fronted
unattested:
*i
e
a
labials,
alveolars,
laryngeals
È, ´, a
central
velars
uvulars
labialized Cs
È∞, ´, a
partially
backed
¨, Ø, A
back
u, o, A
back (and
rounded)
*u
o
a
Again, in terms of cooccurrence constraints, there is no account of the fact that the markedness of
vowels like [i] and [u] depends on the contrasts that they are involved in. Dispersion theory predicts this
pattern. Central vowels yield sub-optimal F2 contrasts, thus are not frequent in front-back contrasts.
However, in the absence of contrast, effort minimization dictates adopting the tongue position of
neighbouring consonants, hence the contextual variation. In the absence of contextual demands, extreme
articulations are avoided, so central vowels are preferred.
In summary, we have seen that we need to impose restrictions on contrasts to account for crosslinguistic generalizations about inventory structure; it is not adequate to simply place restrictions on
individual segments, as cooccurrence constraints do. Specifically, we have seen that contrasts are subject
to a requirement that they be maximally distinct, formalized here in terms of ‘Mindist’ constraints.
4.2. Enhancement
10
Evidence for Constraints on Contrast: The Dispersion Theory of Contrast
Another contrast-related phenomenon closely related to the issue of inventory structure is
enhancement (Stevens, Keyser and Kawasaki 1986, Stevens and Keyser 1989). We shall see that
dispersion theory offers a very straightforward account of enhancement, whereas theories that do not
refer to contrast cannot provide any adequate account.
Stevens et al observe that a basic contrast is often enhanced by ancillary features which are not
independently contrastive. For example, [round] frequently enhances [back] contrasts. Vowels
contrasting in backness differ primarily in F2, with front vowels having higher F2 than back vowels. Liprounding tends to lower F2, so rounding back vowels and unrounding front vowels maximizes the F2
difference between them, enhancing the distinctiveness of the contrast.
In the dispersion model the occurrence of enhancement is a direct consequence of the maximization
of the auditory distinctiveness of contrasts, as exemplified in (25). The constraint ranking in (25) results
in the selection of a single F2 contrast. Given the ‘Mindist F2’ constraints, the contrast selected must be
maximally distinct, i.e. rounding is used to enhance backness.
(24)
highest F2
high F2
low F2
lowest F2
(25)
i
+
+
-
y
+
-
Maintain
F2 contrast
>
i-u
i-¨
y-u
i-È-u
È
-
¨
+
-
u
+
+
MindistF2
=3
MindistF2
=4
*!*
*!
*!
**
Maintain 2
F2
contrasts
*
*
*
In theories which do not explicitly represent contrasts, enhancement relations have been formulated
in terms of redundancy rules (e.g. Stevens, Keyser and Kawasaki 1986:462f.) or implicational statements
(e.g. Archangeli and Pulleyblank 1994) of the form shown in (26)
(26)
[+back] [+round]
These approaches to enhancement avoid reference to contrast by making enhancement independent
of contrast. The rule in (25) states that [+back] vowels must be rounded regardless of whether they
contrast with [-back] vowels. The dispersion theory, on the other hand, predicts that enhancement should
only apply to contrasts since it is motivated by maximization of the distinctiveness of contrasts.
The facts are consistent with the dispersion hypothesis: in all cases of enhancement discussed in the
articles by Stevens et al, it is contrasts that are enhanced, a point noted in Stevens, Keyser and Kawasaki
1986 (p.443). However it is difficult to find cases in which the redundancy rule formulation would
predict the possibility of enhancement, but where there is no contrast. For example, it is not possible to
test the validity of (26) because, as noted in 4.1, we do not find back vowels in the absence of backness
contrasts, except as a result of assimilation. The enhancement of stop voicing by nasalization provides
one test case.
4.2.1. Enhancement of Stop Voicing by Nasalization
Voicing is difficult to maintain in stops because pressure builds up behind the closure until there is
no longer a pressure drop across the glottis. Without a sufficient pressure drop there is no airflow across
the glottis, so voicing ceases (Westbury and Keating 1986). Lowering the velum allows air to be vented
from the vocal tract, mitigating the pressure build-up, and thus facilitating voicing. In addition, greater
amplitude at low frequencies can be transmitted due to the nasal airflow, and low frequency energy is a
11
E. Flemming
primary correlate of voicing (Stevens et al 1986:439). Thus nasalization enhances voicing in stop
consonants.
Pre-nasalization is employed as an enhancement of contrasts between voiced and voiceless stops in
Guaraní, Barasano and Rotokas, among other languages (cf. Piggott 1991, Steriade 1993). However,
‘enhancement’ of voiced stops by nasalization is not observed in the absence of a contrast. We do not
find pre-nasalization of inter-vocalically voiced stops, for example2.
A parallel argument can be made with respect to implosion. The larynx lowering associated with
implosion prevents the build-up of pressure in the oral cavity, and thus facilitates voicing. Like prenasalization, implosion is employed as an enhancement of stop voicing contrasts in Nyangi and Maasai
(Maddieson 1984), and is not found in the absence of contrast, e.g. accompanying intervocalic voicing.
So to determine the applicability of enhancement, we must consider contrasts - i.e. pairs of forms not individual sounds.
It might be thought that contrastive underspecification would allow us to account for the fact that
only contrasts are enhanced, while retaining an analysis of enhancement relations in terms of redundancy
rules. With contrastive underspecification only contrastive features are present in underlying
representation, so if an enhancement rule such as (26) applies before redundant features are filled in, then
only contrasts will be enhanced.
Certainly contrastive underspecification allows us to describe this situation, but it provides no
explanation why enhancement only applies to contrasts. If pre-nasalization can apply prior to the
insertion of redundant voicing, then there is no reason why it should not apply after voicing-specification
in some other language. An account in terms of contrastive underspecification does not forge any
connection between contrast and enhancement, and thus cannot adequately account for the facts.
4.2.2. Enhancement of [-anterior] Sibilants by Rounding
The pattern of enhancement of [-anterior] sibilants, such as palato-alveolars and retroflexes, by liprounding shows that enhancement is sensitive to the contrasts that appear in a given language.
Non-anterior coronal fricatives are often produced with lip-rounding, for example palato-alveolar
fricatives in English and French and retroflex fricatives in Polish are both produced in this way
(Ladefoged and Maddieson forthcoming, Dogil 1990 (cited in Hume 1992)). There is no articulatory
basis for this relationship, but from an acoustic point of view, we can see that rounding serves to enhance
the distinctiveness of contrasts among stridents. This point can be made most simply for languages like
English and French which contrast two sets of stridents, [+anterior] [s, z] and [-anterior] [S, Z]. These
sounds are differentiated primarily by the frequencies at which fricative noise is concentrated. This
depends on the size of the resonating cavity in front of the noise source: the larger the cavity, the lower
its resonant frequency, and thus the lower the frequency of the intensity peak in the fricative. [-Anterior]
fricatives have a larger front cavity, and thus a lower intensity peak, than [+anterior] fricatives. Liprounding increases this difference by lowering the resonant frequency of the front cavity.
This situation is represented in the feature specifications for Noise Frequency in (27). Rounding the
palato-alveolar fricative and unrounding the alveolar yields the maximal distinction in NF.
(27)
NF:
lowest NF
low NF
high NF
highest NF
ßW ß SW/ÇW S/Ç
+ - - + + +
- - - - +
- - - -
2It
sW s
+
+
is always problematic to establish a negative claim of this kind. A systematic survey of stop voicing allophony by
Keating, Linker and Huffman (1984) did not find such a pattern, but they also do not describe many languages with
intervocalic voicing of stops.
12
Evidence for Constraints on Contrast: The Dispersion Theory of Contrast
A comparsion with Polish shows that an analysis in terms of any redundancy rule such as [-anterior]
 [+round], is incorrect. In fact, the correct generalization is that the sibilant with the lowest NF sibilant
is rounded, i.e. the application of enhancement by rounding is dependent on the sibilant contrasts in the
given language.
In French and English, the strident with the lowest NF is the palato-alveolar. In Polish, on the other
hand, there are stridents at three contrasting places: dental [s1], alveolopalatal [Ç] and retroflex (apical
postalveolar) [ß]. Only the strident with the lowest NF, the retroflex, is produced with lip-rounding. The
alveopalatal is [-anterior], but does not undergo rounding. This situation is predicted by considerations of
maximization of distinctiveness. Further lowering the NF of the strident with the lowest NF increases the
distinctiveness of its contrasts with higher NF stridents. As the tableau in (28) shows, rounding any other
strident, including [-anterior] [Ç] would result in sub-optimal contrasts.
(28)
Maintain 2
NF contrasts
s-Ç-ß
> s-ÇßW
s-ÇW-ßW
s-ÇW-ß
MindistNF
=1
*!
MindistNF
=2
*!
MindistNF
=3
**
**
*!
*
*
*
The generalization concerning the cross-linguistic application of enhancement of [-anterior] stridents
by rounding can only be formulated relative to the contrasts found in a given language. A redundancy
rule formulation cannot capture this sensitivity of enhancement to contrast, but it is an automatic
consequence of the dispersion theory.
4.2.3. Peripheralization vs. Dispersion
At this point we should consider the possibility of an alternative analysis of the facts of inventory
structure and enhancement which adopts the auditory representations proposed here, but rejects the
notion that contrast is directly implicated in these processes. Instead it might be proposed that the
tendencies identified as enhancement involve a preference for sounds that are peripheral in the auditory
space. According to this analysis, contrasts such as [i-u] are preferred purely because front unrounded
and back rounded vowels have extreme values of F2, not because the resulting contrast is maximal.
This attempt to eliminate direct constraints on contrasts runs into a number of problems. Firstly, it is
not clear how a preference for peripherality can account for ternary contrasts between front central and
back vowels (i.e. high, mid and low F2). A simple-minded application of a preference for peripheral
sounds would predict a preference for vowel systems with front, retracted front , and back rounded
vowels, [i, i2, u], over a better dispersed inventory with front, central and back vowels, [i, È, u], because
the vowels in the former inventory are more peripheral than in the latter. In fact front-central-back
contrasts are well attested, whereas front-retracted vs. front contrasts (without an accompanying length
difference) are unattested.
One might try to avoid this consequence by formulating peripheralization in terms of redundancy
rules such as [+high F2]  [+highest F2], i.e. a vowel with high F2 should have as high an F2 as
possible. A central vowel ([-high F2, -low F2]) would not be subject to such a peripheralization rule.
However, a rule like this is not general - it specifically sub-divides the F2 dimension into front, central
and back. A language like Turkish which has four contrasting vowels on the F2 dimension, [i, y, ¨, u]
would require different peripheralization rules, since it does not have central vowels.
Furthermore, if a rule like [+high F2]  [+highest F2] can apply independently of contrast, then we
should expect to find languages which are like Kabardian (4.1) in that they do not contrast front and back
vowels, but rank peripheralization highly. This would lead to a situation in which vowels varied
contextually between extreme front and extreme back realizations, rather than the attested situation in
which a range of intermediate variants are found, depending on the conditioning environment.
13
E. Flemming
In the same vein, this analysis is subject to the problem faced by the redundancy rule analysis of
enhancement. I.e. it predicts that enhancement effects should be independent of contrast, which is not the
case, as was argued above.
4.3. Neutralization
Neutralization, as the name suggests, involves loss of a contrast in some environment, but in theories
that do not directly represent contrast, it is not actually possible to analyze loss of contrast as essential to
the phenomenon. On the other hand, dispersion theory allows a clear representation of loss of contrast.
We shall see, following Steriade 1993, that there are neutralization phenomena that must be characterized
as loss of a contrast per se, and thus provide support for the dispersion theory.
In standard analyses of neutralization, loss of contrast is a by-product of a dsitributional restriction.
For example, coda devoicing neutralizes voicing contrasts, and can be accounted for in terms of a
restriction on the appearance of [+voice] in coda:
(29)
*Coda
|
[+voice]
In this analysis no reference is made to the fact that a contrast is neutralized, we simply observe that if
[+voice] cannot appear in coda, then no contrast between [+/- voice] can occur in this position.
In dispersion theory, neutralization of a contrast results when constraints prevent it from achieving
sufficient distinctiveness in some environment (cf. Steriade 1993).
Evidence for this account of neutralization over the standard analysis in terms of distributional
restrictions derives from cases of neutralization in which the result is free variation. An example of this
type comes from Gooniyandi, an Australian language (McGregor 1990, Hamilton 1993). In this language
the contrast between retroflex and apical alveolar stops is permitted only post-vocalically; it is
neutralized word-initially and post-consonantally. In positions of neutralization there is free variation
between retroflex and alveolar, as well as intermediate articulations.
This neutralization cannot be analyzed in terms of a restriction on the distribution of retroflexes or
alveolars, since both can appear in the neutralizing environment. It must be analyzed as the suspension of
the contrast per se (Steriade 1993). A similar situation obtains in Walmatjari (Hudson and Richards
1969), although it is not clear whether neutralization results in free variation or a specific intermediate
articulation.
Another case where neutralization yields free variation is found in Danish. In Danish, the contrast
between aspirated and unaspirated stops is neutralized word-finally, but there is free variation between
the two sounds in this position (Fischer-Jorgenson 1954). A final case is Mong Njua (Lyman 1974),
where velars and labials are in free variation preceding laterals, although they contrast elsewhere.
To show how these phenomena can be analyzed in terms of dispersion theory, we will first consider
the case of coda voicing neutralization, which involves a determinate output, then show how the analysis
can be naturally extended to cases like Gooniyandi.
As summarized above, neutralization is a consequence of constraints which prevent a contrast from
achieving sufficient distinctiveness. In the case of coda devoicing, we will simply assume that the
relevant constraint is *Voiced coda: ‘Avoid voiced obstruents in coda’. The dimension of contrast for
consonant voicing is assumed to be voice onset time (VOT). The ranking that yields neutralization is
given in (30). (The constraints are listed in a vertical column, with highest ranked constraints at the top.
Constraints whose relative ranking is indeterminate are separated by commas.)
(30)
*Voiced coda,
‘Avoid voiced obstruents in coda’
Mindistvot=1 >>
‘Sounds that contrast in VOT should differ in at least one VOT feature’
Maintain VOT contrast ‘Maintain a contrast on the VOT dimension’
As shown in (31) neutralization of a contrast occurs because ‘*Voiced coda’ motivates violation of
lower-ranked ‘Maintain VOT contrast’. The resulting form is selected to satisfy ‘*Voiced coda’, so the
14
Evidence for Constraints on Contrast: The Dispersion Theory of Contrast
output will always be voiceless. Free variation will occur where no constraints favour a particular output
after neutralization, as in Gooniyandi.
(31)
*Voiced coda
Mindistvot
=1
Maintain
VOT contrast
rad - rat
*!
>
rat
*
rad
*!
*
The fact that the distinction between apical alveolar and retroflex stops is neutralized in non-postvocalic positions can be explained as follows (Steriade 1993): Retroflexes are differentiated from apical
alveolars by low F3 transitions at closure (Stevens and Blumstein 1975). However, the tongue tip moves
forward during the closure of a retroflex and is released from the alveolar ridge, so these sounds are
articulatorily and acoustically very similar at release (Anderson and Maddieson 1994, Spajic, Ladefoged
and Bhaskararao 1994).Thus neutralization occurs in initial and post-consonantal positions because the
closure transitions which distinguish the sounds are not present.
We can formulate this analysis in terms of the following constraint ranking:
(32)
MindistF3 = 1 >>
‘Sounds that contrast in F3 transition should differ in [low F3]’
Maintain F3 transition contrast
We will assume that retroflex and apical alveolar stops are distinguished only by their closure
transitions, which are [+low F3] for the retroflex, and [-low F3] for the apical alveolar. At release they
are both [-low F3]. Hence in post-consonantal position, it is not possible to mantain the contrast, and
neutralization results (33). This situation differs from voicing neutralization because in that case the
constraint that favoured neutralization also favoured a particular outcome, namely voicelessness. In
Gooniyandi, neutralization results simply because the two sounds are too similar to form an adequate
contrast. Which of the sounds actually surfaces thus depends on other constraints active in the language.
The simplest account of the free variation is to assume that there is no constraint that differentiates the
two possibilities, so both are acceptable (33). The distinction is neutralized precisely because the sounds
are similar, so neither is to be preferred for purposes of distinctiveness from other consonants. So if
neither articulation is significantly more effortful than the other, there is no basis for a preference.
(33)
Ct ¡V-CÊV
>
Ct ¡V
>
CÊV
MindistF3
=1
*!
Maintain F3
transition contrast
*
*
Dispersion theory enables us to distinguish neutralization of a contrast, i.e. violation of a ‘Maintain
contrast’ constraint, from the actual outcome of neutralization, whcih may be determined by a variety of
constraints. Analyses in terms of restrictions on the distribution of features can only characterize
neutralization as the consequence of such a restriction, and thus are not adequate for the analysis of
neutralizations which do not yield a determinate output.
4.3.1. Further Remarks on the Analysis of Gooniyandi and Danish
As shown above, free variation in Gooniyandi results from a situation where no constraints
differentiate two candidates. In the analysis presented there, it is assumed that no constraint milittates in
favour of either candidate. This depends on the assumption that all the attested articulations from
retroflex to apical alveolar involve the same degree of effor, which is perhaps questionable. It would
seem likely that extreme retroflexion, at least, is a more effortful articulation than an apical alveolar. In
any case, it is far from obvious that the same analysis can be extended to the Danish case. In this section
we will consider a refinement of the analysis that answers these objections.
15
E. Flemming
In Danish, aspirated and unaspirated stops are in free variation in final position. It seems unlikely that
these stops are equivalent either articulatorily or auditorily. A more plausible analysis is that the
aspirated stop is both more effortful, and more distinct, since the release burst and formant transitions
present in the aspirated release provide place cues. The free variation can then be seen as resulting from
variation in the ranking of minimization of effort and maximization of distinctiveness. We can represent
this in terms of a pair of unranked constraints, informally specified in (34). Favouring effort
minimization will result in the unaspirated output, whereas favouring distinctiveness of place cues will
result in the aspirated output.
(34)
*Aspiration, ‘Maximize distinctiveness of place cues’
A parallel account of the free variation in Gooniyandi may be possible. It was suggested above that
considerations of effort minimization would disfavour the retroflex alternant, so to achieve free variation,
there must be a conflicting constraint that favours the retroflex. This could be maximization of
distinctiveness from the other coronals of Gooniyandi, namely laminal dentals and laminal palatals. For
purposes of the simple analysis presented above we assumed that retroflexes are identical to apical
alveolars at release, but they are actually distinct, although similar. While F3 lowering is much greater at
the closure of a retroflex, F3 is generally still lower than in other coronals at release (e.g. Spajic et al
1994), so this property would help to make a retroflex mores distinct from the laminal coronals than an
apical alveolar. Some support for this suggestion can be derived from the fact that Anderson and
Maddieson (1994:158) found it difficult to reliably distinguish apical alveolars from laminal coronals in
a study of similar contrasts in Tiwi. The post-alveolar, or retroflex, was distinguished from other coronals
by duration, burst amplitude, burst spectrum and F3 onset. We can represent this latter difference with a
finer differentiation of the F3 dimension:
(35)
high F3
low F3
Retroflex
approach
+
Retroflex
release
-
Dental/Alveolar
+
-
It is not clear what the primary cues differentiating non-retroflex coronals are (Anderson and
Maddieson 1994), so we will formulate the distinctiveness constraint as ‘Mindist C = n & 1 F3’ to
indicate that the neutralized stop is distinguished from the other coronals by an F3 difference in addition
to other unspecified cues. This constraint is then unranked with respect to the effort minimization
constraint ‘*Retroflex’, yielding two equally optimal outputs (36)
(36)
/Ct ¡V-CÊV/
Ct ¡V-CÊV
>
Ct ¡V
>
CÊV
MindistF3
=2
*!
*Retroflex
MindistC
= n & 1 F3
*
*
Maintain F3
trans. contrast
*
*
Note that, in spite of the greater complexity of the modified analysis, the account of free variation
after neutralization is still the same: Neutralization essentially involves violation of a ‘Maintain contrast’
constraint, the specific output then depends on the constraint ranking. A determinate ouput results when a
constraint favours one term of the contrast, and free variation results if no constraints differentiate the
terms of the contrast, or if there are conflicting constraints which are unranked with respect to each other.
The considerations discussed here simply suggest that the latter situation is the more likely source of free
variation.
4.3.2. Parallels between Neutralization and Inventory Selection
The account of neutralization developed here is precisely parallel to the account of inventory
structure given above. The connection essentially lies in the fact that the constraints on contrast operate
to select an inventory of contrasts in a particular context. For example, syllabification places constraints
16
Evidence for Constraints on Contrast: The Dispersion Theory of Contrast
on possible sequences of segments, so, in a language with only CV syllables, the constraints on contrast
can only select an inventory of vowel contrasts following a consonant. From this perspective,
neutralization occurs when a contrast that is selected in one context is not selected in another. The
difference between a contrast which does not occur in a language and one that is contextually neutralized
is that the former is not selected in any contexts whereas the latter is selected only in certain contexts.
This association between inventory selection and neutralization is supported by the observation that
vowel reduction generally produces reduced inventories which are similar to typologically common
stressed vowel inventories. E.g. The standard Italian seven vowel system /i, e, E, a, O, o, u/ reduces to the
canonical five vowel system, /i, e, a, o, u/ in unstressed syllables, and a five vowel system frequently
reduces to a three vowel system, loosely transcribed as /i, a, u/ (e.g. Sicilian, Russian, etc). According to
the present account this follows from the fact that essentially the same selectional forces apply in stressed
and unstressed syllables, but effort minimization has stronger consequences in short unstressed syllables.
We will exemplify this point with an analysis of Sicilian vowel reduction, in which a five vowel
system, /i, e, a, o, u/, reduces to three vowels /i, a, u/ ([I, ø, U]) in unstressed syllables (Mazzola
1976:41):
(37)
vínni
véni
ávi
móri
úggyi
‘he sells’
‘he comes’
‘he has’
‘he dies’
‘he boils’
vinnímu
vinímu
avíti
murímu
uggyímu
‘we sell’
‘we come’
‘you have’
‘we die’
‘we boil’
This is a reduction in height contrasts, so we need only consider the F1 dimension, which is repeated in
(38).
(38)
lowest F1
low F1
high F1
highest F1
i
+
+
-
I
+
-
e
-
ø
+
-
a
+
+
The reduction can be analyzed in terms of the constraint ranking shown in (39). This ranking is
basically the same as that in (10) above, which yields three vowel heights. The only addition is a
constraint against short peripheral vowels. This constraint is assumed to be motivated by effort
minimization: Achieving a peripheral vowel position in a short duration requires rapid movement, and
avoidance of the effort involved has been hypothesized to play an important role in the centralization of
short vowels (Lindblom 1963).
(39)
Maintain F1 contrast,
*Peripheral short vowel >>
MindistF1 = 2 >>
Maintain 2 F1 contrasts >>
MindistF1 = 3
‘Maintain a contrast on the F1 dimension’
‘Avoid peripheral short vowels’
‘Sounds that contrast in F1 should differ in at least two F1 features’
‘Maintain two contrasts on the F1 dimension’
In stressed syllables, ‘*Peripheral short vowel’ is irrelevant, so the ranking yields three vowel
heights, as shown in (40). However, in short syllables, this effort minmization constraint penalizes
peripheral vowels, so the candidate [i-e-a] is now ruled out because it contains two peripheral vowels
(41). Attempting to maintain two contrasts while avoiding peripheral vowels, as in the second candidate,
results in violations of the minimal distance requirement because [I-e] and [e-ø] differ in only one F1
feature. The winning candidate has only two vowel heights, violating ‘Maintain 2 F1 contrasts’, but
satisfying the higher-ranked minimal distance requirement.
17
E. Flemming
(40)
Maintain
F1 contrast
*Short
peripheral
vowel
MindistF1
=2
> i!-e!-a!
i!-a!
(41)
Maintain 2
F1 contrasts
MindistF1
=3
**
*!
Maintain
F1 contrast
*Short
peripheral
vowel
*!*
MindistF1
=2
Maintain 2
F1 contrasts
MindistF1
=3
i*-e*-a*
**
I*-e*-ø*
*!*
**
>
I*-ø*
*
*
This analysis illustrates the fact that contrasts can only be selected in a given context, because the
effect of constraints like ‘*Peripheral short vowel’ depends on context. We also see that differences in
context, in this case involving stress, can result in neutralization of contrasts through the selection of
different inventories.
4.4. Contrast Preservation
Direct evidence for constraints favouring the maintenance of contrasts is provided by phenomena in
which languages take measures to preserve contrasts. One case is the realization of the voicing contrast in
English stops.
4.4.1. Realization of the Voicing Contrast in English Stops
As the table in (42) shows, a voicing contrast is maintained in all positions, but with varying
realizations:
(42)
Initial
Medial
(unstressed)
[-voice]
voiceless
aspirated
voiceless
unaspirated
[+voice]
voiceless
unaspirated
voiced
We can explain this distribution as follows (cf. Kingston and Diehl 1994:428f.): As noted above,
voicing is marked in initial position because it is difficult to start vocal cord vibration during an
obstruent, so it is preferable to devoice stops in this position, as English does. But if the ‘voiced’ stop is
devoiced, then, to maintain a contrast, the voiceless stop must be aspirated. In medial position, voicing a
stop is not as difficult because it is only necessary to maintain vocal cord vibration (Westbury and
Keating 1986), so the contrast can be maintained without aspirating the voiceless stop.
According to this analysis, the realization of the stops depends on the requirement that a distinct
VOT contrast be maintained. This point can be made explicit by formalizing the analysis in terms of the
constraint ranking in (43). The top-ranked constraints require the maintenance of a VOT contrast with
distinctiveness of 1. These are ranked above the constraints ‘*Initial voiced stop’ and ‘*Aspiration’.
‘*Initial voiced stop’ is motivated by minimization of effort. Aspiration might be disfavoured because of
the effort involved, or because of the devoicng effect on a following vowel.
(43)
MindistVOT = 1,
Maintain VOT contrast >>
*Initial voiced stop >>
*Aspiration
‘Avoid initial voiced stops’
‘Avoid aspirated stops’
18
Evidence for Constraints on Contrast: The Dispersion Theory of Contrast
(44)
VOT: aspirated
voiced
tÓ t
+ - -
d
+
The tableau in (45) illustrates the selection of VOT contrasts in initial position. A contrast between
voiced and voiceless unaspirated stops is rejected because it violates ‘*Initial voiced stop’ (candidate 1).
Satisfying this constraint by simply devoicing the voiced stop results in neutralization, violating
undominated ‘Maintain VOT contrast’. The optimal candidate devoices the voiced stop, but maintains a
contrast by aspirating the voiceless stop.
In medial position ‘*Initial voiced stop’ is not relevant, so a contrast can be maintained without
aspirating the voiceless stop (46).
(45)
MindistVOT
=1
>
#t-d
#tÓ-t
#t
(46)
Maintain VOT
contrast
*Initial
voiced stop
*Aspiration
*!
*
*!
MindistVOT
=1
Maintain VOT
contrast
*Initial
voiced stop
*Aspiration
VtV-VdV
VtÓV-VdV
*!
The constraint ‘Maintain VOT contrast’ plays an essential role in this analysis. The devoicing of
initial stops is explained in terms of minimization of effort, but the aspiration of initial voiceless stops is
explained in terms of the need to maintain a contrast. Without this constraint, the result would be
neutralization.
Generally, this analysis demonstrates the necessity for comparative constraints. The fact that the
voiceless stop needs to be aspirated in initial position depends on the devoicing of the voiced stop, but
the voiced stop also cannot devoice unless the voiceless stop is aspirated. The realizations of the stops
thus cannot be determined independently, they must be fixed simultaneously through comparison of
contrasting forms.
Note that we have not accounted for aspiration of voiceless stops in stressed medial syllables. In this
case, aspiration applies even though it is not necessary for maintenance of contrast. There are a couple of
possible explanations for this distribution. Possibly stressed syllables permit the greater effort involved in
achieving the maximal contrast between fully voiced and aspirated stops. Alternatively, it was suggested
above that aspiration might be dispreferred because it tends to devoice following vowels. This would be
more problematic before short vowels where total devoicing could result (as in English [pÓ´9teI|oU]
‘potato’), but would not be problematic before long, stressed vowels.
4.4.2. Contextual Enhancement in Moroccan Arabic
The enhancement of labialization by pharyngealization in Moroccan Arabic provides a similar case
of contrast preservation. Pharyngealization is an enhancement of labialization, but this enhancement only
applies in environments where the labialization contrast would otherwise be indistinct. Thus the
enhancement is motivated by the need to maintain a contrast.
In Moroccan Arabic underlying sequences of a labial followed by a round glide are usually realized
as lengthened pharyngealized labials, often with some rounding also (Heath 1987:225ff., Harrell 1962:9):
(47)
/bw/ [bb≥W] bwaX

bb≥WaX
/mw/  [mm≥W]
mwag´n 
mm≥Wag´n
19
‘steam’
‘clocks’
E. Flemming
Labio-velar glides are characterized by a low F2, close to F1, and pharyngealization raises F1
increasing this proximity, and thus enhancing labialization. But this enhancement only applies following
a labial.
Dispersion theory provides a straightforward account of the environmental restriction on
enhancement of round glides in Moroccan Arabic: The enhancement is particularly needed following a
labial because a plain labial has a low F2 at release, so the contrast between [b] and [bw] is relatively
indistinct, an observation that explains why this contrast is not permitted in many languages, including
English. Thus the enhancement is motivated by the desire to maintain a sufficiently distinct contrast.
The basic form of the analysis is illustrated in (48) and (49) 3. The higher-ranked constraints require
the maintenance of a labialization contrast with a minimum distance of 2. A constraint disfavouring
pharyngealization is ranked below these, but above more stringent requirements on distinctiveness of the
contrast (MindistF2-F1 = 3). Given the latter ranking, enhancement will not apply in contexts where a
sufficiently distinct contrast can be achieved without it (48). The former ranking results in enhancement
where the contrast would not be sufficiently distinct without it (49).
(48)
Maintain
F2-F1
contrast
>
dWd
d≥W-d
d
(49)
>
*Pharyngealization
MindistF2-F1
=3
*
*!
*!
Maintain
F2-F1
contrast
bW-b
bb≥Wb
b
MindistF2-F1
=2
MindistF2-F1
=2
*Pharyngealization
*!
*
MindistF2-F1
=3
*
*
*!
Again, the ‘Maintain contrast’ constraint plays a crucial role: the enhancement is motivated by the
need to satisfy this constraint and its concomitant ‘Mindist’ requirement. This analysis also crucially
depends on the representation of contrasts. A round glide can contrast adequately with its absence
following, e.g., an alveolar, but the same glide does not produce an adequate contrast following a plain
labial. We cannot determine that the round glide is ill-formed without reference to the contrasting plain
labial.
5. Summary
We have seen evidence from a variety of phenomena indicating the existence of constraints on the
well-formedness of contrasts. Specifically, we have argued for constraints favouring the following goals
(repeated from (1)):
(50)
i. Maximize the number of contrasts
ii. Maximize the distinctiveness of contrasts
3The
dimension F1-F2 is used here to capture the fact that the labialization contrast is enhanced by raising F1 or
lowering F2, since both result in a smaller difference between F1 and F2. A small difference between F1 and F2 is
often regarded as the defining characteristic of back rounded vowels (Stevens, Keyser and Kawasaki 1986). It is
possible to reformulate the analysis to treat F1 and F2 as independent dimensions (Flemming 1995).
20
Evidence for Constraints on Contrast: The Dispersion Theory of Contrast
iii. Minimize articulatory effort
Constraints of this type imply that the well-formedness of a word cannot be evaluated in isolation
because it depends on the well-formedness of the contrasts that it enters into. They also give a central
role to auditory representations in phonology, since these representations form the basis for the
evaluation of the distinctiveness of contrasts.
21
E. Flemming
References
Anderson, Victoria B. and Ian M. Maddieson (1994). Acoustic Characteristics of Tiwi Coronal
Stops.UCLA Working Papers in Phonetics 87, 131-162.
Archangeli, Diane, and Douglas Pulleyblank (1994). Grounded Phonology. MIT Press, Cambridge.
Bender, Byron (1968). Marshallese phonology. Oceanic Linguistics 7.
Catford, John C. (ms.). Vowel systems of Caucasian languages. Ms., University of Michigan.
Choi, John D. (1992a). An acoustic study of Kabardian vowels. JIPA 21, 1-12.
(1992b). Phonetic Underspecification and Target Interpolation: An Acoustic Study of Marshallese
Vowel Allophony. UCLA Working Papers in Phonetics 82.
Dave, Radhekant (1977). Retroflex and dental consonants in Gujarati. A palatographic and acoustic
study. Annual Report of the Institute of Phonetics, University of Copenhagen 11: 27-156.
Dogil, G. (1990). Hissing and hushing fricatives: A comment on non-anterior spirants in Polish. Ms.
Universität Bielefeld.
Fischer-Jørgenson, Eli (1952). Acoustic analysis of stop consonants. Miscellanea Phonetica, vol II, 4259.
Flemming, Edward (1995). Auditory Representations in Phonology. Ph.D. dissertation, UCLA.
Hamilton, Philip (1993). Markedness and coronal structure. Ms., University of Toronto.
Harrell, Richard S. (1962). A Short Reference Grammar of Moroccan Arabic. Georgetown University
Press, Washington, D.C.
Heath, Jeffrey (1987). Ablaut and Ambiguity: Phonology of a Moroccan Arabic Dialect. State University
of New York Press, Albany.
Hudson, Joyce and E. Richards (1969). The phonology of Walmatjari. Oceanic Linguistics 8, 171-189.
Keating, Patricia A., Wendy Linker, and Marie Huffman (1984). Patterns in allophone distribution for
voiced and voiceless stops. Journal of Phonetics 12.
Kingston, John, and Randy L. Diehl (1994). Phonetic knowledge. Language 70, 419-55.
Kirchner, Robert (1995). Contrastiveness is an epiphenomenon of constraint ranking. Forthcoming in
BLS.
Ladefoged, Peter, and Ian M. Maddieson (1996). The Sounds of the World’s Languages. Blackwell,
Oxford.
Lindblom, Björn (1967). Vowel duration and a model of lip-mandible coordination.STL-QPSR 4. 1-29.
22
Evidence for Constraints on Contrast: The Dispersion Theory of Contrast
(1986). Phonetic universals in vowel systems. J.J. Ohala and J.J. Jaeger (eds) Experimental
Phonology. Academic Press, Orlando.
(1990). Phonetic content in phonology. PERILUS 11.
Lyman, Thomas A. (1974). Dictionary of Mong Njua, a Miao (Meo) language of Southeast Asia.
Mouton, The Hague.
Maddieson, Ian M. (1984). Patterns of Sounds. Cambridge University Press, Cambridge.
(1987). The Margi vowel system and labiocoronals. Studies in African Linguistics 18.
Martinet, Andre (1952). Function, structure, and sound change.Word 8:1.
(1955). Economie des Changements Phonétiques. Francke, Berne.
Mazzola, M. L. (1976). Proto-Romance and Sicilian. Peter de Ridder, Lisse.
McGregor, William (1990). A Functional Grammar of Gooniyandi, John Benjamins, Amsterdam.
Piggott, Glyne (1992). Variability in feature dependency: The case of nasality. Natural Language and
Linguistic Theory 10, 33-78.
Prince, Alan, and Paul Smolensky (1993) Optimality TheoryI. Ms., Rutgers and University of Colorado.
Spajic!, Sinis #a, Peter Ladefoged, and Peri Bhaskararao (1994). The rhotics of Toda. UCLA Working
Papers in Phonetics 87, 35-66.
Steriade, Donca (1993) Positional Neutralization. Talk given at NELS 24, University of Massachusetts,
Amherst.
(1995). Neutralization and the expression of contrast. Ms., UCLA.
Stevens, Kenneth N., and Sheila E. Blumstein (1975). Quantal aspects of consonant production and
perception: A study of retroflex stop consonants. Journal of Phonetics 3, 215-33.
, and S. Jay Keyser (1989). Primary features and their enhancement in consonants. Language 65, 81106.
, , and Haruko Kawasaki (1986). Toward a phonetic and phonologicaltheory of redundant
features. Joseph S. Perkell and Dennis H. Klatt (eds) Invariance and Variability in Speech Processes.
Lawrence Erlbaum, Hillsdale.
Westbury, John, and Patricia A. Keating (1986). On the naturalness of stop consonant voicing. Journal of
Linguistics 22, 145-66.
Zipf, George K. (1949) Human Behavior and the Principle of Least Effort. Addison-Wesley, Cambridge.
23