Schwa in Phonological Theory

Marc van Oostendorp

This article appeared in GLOT International, 1999.

1. Introduction

If a language has schwa in its vowel inventory, this segment usually has a special role to play in the phonology of the language. It can only occur in a simple type of syllable; or it is invisible for the stress system; or it is epenthetic; or it is the result of reduction; etc. Linguistic theory has to explain this special behaviour of schwa: why is it exactly this segment which behaves in exactly this way in so many languages? A lot of subtheories of phonology are at stake -- syllable structure theory, stress theory, the theory of (sub)segmental representation, and of interaction between segments -- and the questions surrounding schwa therefore will probably not be resolved until the perfect theory of phonology has been discovered.

On the other hand, schwa offers an excellent testcase for many theories of phonology. It therefore should come as no surprise that the vowel has been studied intensively within various branches of generative phonology. This article offers an overview of some of the ideas that have been proposed and some of the facts that have been discovered over the past thirty years. None of these analyses seems perfect in the sense that it can handle all the relevant facts; but together the different propsals give us a rather precise picture of the behaviour of schwa.

It is my aim in this paper to present in a more or less systematic way those facts about schwa which I consider to be most characteristic. I also briefly present some of the more influential theories surrounding this vowel within the generative paradigm.

I find it useful to distinguish pretheoretically between three types of schwa:

I do not think the taxonomy just presented should be awarded any theoretical status: a schwa may change its behaviour over the years, and indeed we also find cases where a vowel is e-schwa and r-schwa at the same time (as in the French paradigm appeler [Aple~Ap@le] `to call' j'appelle [ZApEl] `I call').

The classification is presented here because it helps to distinguish the various roles that schwa can play in the phonology of a language. Before we can study the behaviour of schwa in natural language, it may be useful to go into the question of how the segment should be represented phonologically.

2. The representation of schwa

Schwa is special not only from the point of view of phonological theory. We find that it is a special vowel also if we look at it from an articulatory point of view: it could be described as a `targetless vowel' for which no inherent articulatory target has been specified, or as a vowel which targets a neutral vocalic position, `the mean tongue-tract variable position for all the full vowels.' (Cf. Browman and Goldstein 1992).

From a standard autosegmental/metrical point of view, we have several options to represent the emptiness of schwa. We can divide these options roughly into two sets. On the one hand, we can award schwa the status of an empty prosodic position without (auto)segmental material (an empty mora, an empty nucleus, an empty V-slot in a CV-model). On the other hand, we can assign some minimal segmental specification to this vowel. In the feature geometric model of McCarthy (1988), for instance, major class features constitute the root node. Schwa can then be seen as a bare [-cons] root, or as a root node with an empty (vocalic) place node. It is hard to find good empirical tests that could distinguish between these options. Many things depend on the other theoretical choices we make vis à vis the structure of syllables and segments. In the remainder of this article, I will arbitrarily assume that schwa could be pictured phonologically as a bare root node, in (1):

(1) [-cons]

Schwa can be literally seen as the maximally unmarked vowel: it is marked for being a vowel ([-cons]) but for nothing else. I believe that, in the ideal case, as many phonological properties as possible should be made to follow from the interaction of this fact and general principles of phonology. Ideally, no linguistic rule or constraint should specifically refer to schwa.

I will take the fact that many languages do not have a schwa in their inventory as an example. Within constraint-based phonology, we could analyse this fact as a consequence of a specific constraint against schwa:

(2) *Schwa: Don't have schwa

Yet it seems more reasonable to see this as a specific consequence of something more general. For instance the set of constraints that have been called Fill within Optimality Theory (Prince and Smolensky 1993):

(3) Fill: Empty segments are disallowed.

In Van Oostendorp (1995) it is proposed that the behaviour of schwa can be understood in terms of a relation between segmental material and prosodic structure. If a segment has certain features, it should project to certain positions (say, the head of a syllable or a foot), and if a segment appears in the head position of a certain syllable or a foot, it should have certain (vocalic) features. This is regulated by the following general constraint scheme:

(4)Where F=a (vocalic) feature, P=a prosodic position

Project (F, P): If a vowel is [+F], it should appears in position P.
Project (P, F): If a vowel appears in position P, it should be [+F].

Van Oostendorp (1995) shows that particular settings for the arguments F and P can explain the behaviour of many vowels in a number of languages. Schwa is a special case of this. Since schwa does not have any vocalic feature at all, it is restricted to a very small number of positions. In certain languages this number may be so limited, that it is null.

3. E-schwa

One of the most prominent characteristics of schwa is that it alternates with zero. As a matter of fact the term shvarabakti vowel was initially meant to refer to the epenthetic vowel in Tiberian Hebrew. Epenthesis -- insertion of non-underlying material -- is of course one possible source of the schwa-zero alternation that defines e-schwa. Another source is deletion of schwa.

If schwa is a very simple, there are reasons to assume that this vowel is particularly eligible for both epenthesis and deletion. In the case of epenthesis, it is the most economical choice to make if we have to insert a vowel: no material needs to be epenthesized beyond the feature [-consonantal]. In the case of deletion, it is simpler to delete only the specification [-consonantal] than to delete this feature alongside others (which is needed in case of the deletion of a full vowel). It is thus predicted that if a language (i) allows schwa on the surface (of a given subphonology) and (ii) has a vowel-zero alternation (within that same subphonology), schwa will be among the alternating vowels. I have no evidence that would falsify this prediction; I know of no study which tries to systematically verify it either.

Assuming for the sake of simplicity that there can be no constraints on underlying forms (this is of course the standard assumption within Optimality Theory), we have evidence for deletion if there are certain phonologically well-defined contexts in which schwa does not surface. We have evidence for epenthesis on the other hand if there are phonologically well-defined contexts in which we always find a schwa. Interestingly, there are languages in which we find evidence both for schwa epenthesis and for schwa deletion: there are certain contexts in which schwa always occurs, there are other contexts in which schwa never occurs and there may even be contexts in which it is unpredictable whether we find a schwa or not.

In languages of this type, epenthesis and deletion often seem to be defined in complementary contexts. These languages are of particular interest to constraint-based theories of phonology, because they often involve some form of `rule conspiracy': epenthesis and deletion seem to conspire to attain some absolutely well-formed (syllable) structure. A rather well-known instance of such a language is French. The distribution of schwa in French has been a topic of debate during virtually any period of time within the history of phonology (Anderson 1982, Basbøll 1981, 1988, Charette 1991, Dell 1973/1985, Durand 1976, 1986, 1990, Morin 1982, 1988, Noske 1982, 1988, 1992, Van Oostendorp 1995, Schane 1965, 1968, 1974, Selkirk 1978, 1982, Tranel 1981, 1987, 1994). The facts of French are rather complicated, and a lot of dialectal and probably even idiolectal variation is involved. In any case, the French facts are quite typical of the behaviour of e-schwa in natural language. Noske (1992) presents a useful classification of the facts of French e-schwa, which I will take as my guideline here.

In the following discussion I distinguish betwee `static' and `dynamic' pieces of evidence. `Static' evidence for a constraint (or rule) is distributional: we never find schwa in a certain position, therefore there should be a constraint or a mprheme structure condition against schwa in that position (or a rule deleting it there); we always find a schwa in a certain position, therefore there should be a constraint forcing it to be present (or a rule epenthesizing it). In order to be able to accept `static' evidence for an output constraint, we have to assume some version of OT's `Richness of the Base' (Prince and Smolensky 1993): there are no constraints on underlying representations, and everything can be input to the grammar.

`Dynamic' evidence for a rule or a constraint is based on alternations. A schwa in form A corresponds to a zero in exactly the same environment in the morphologically related form B. I take this to mean that there is a constraint against zero in A, or against schwa in B (or that there is an epenthesis rule applying to A or a deletion rule applying to B).

Type A: Prevocalic schwa deletion. The evidence for this type of process is that schwa cannot be followed by a full vowel. It is actually hard to find evidence for the rule in French, because we need a context where /X@#V/ surfaces as [X#V], and schwa is deleted at the end of a word or phrase in many cases anyway. The alternation of the masculine singular determiner le is sometimes cited as evidence for the rule. This clitic surfaces as /l@/ before a consonant-initial stem (le camerade `the comrade' [l@kamrad]) and as /l/ before a vowel-initial stem (l'ami `the friend' [lami]). The problem is, however, that a similar alternation applies to the feminine singular determiner la (la camerade - l'amie), but we have no reason to postulate a general rule of a-deletion before a full vowel.

The other dynamic pieces of evidence usually given are equally dependent on other assumptions. For instance, it is often assumed that the feminine singular suffix is a schwa (autre `other, FEM.' /otR/+/@/ [otR]). The problem is that this schwa never surfaces at the end of a phrase, and in those environments in which it does surface (autre femme `other woman' [otR@fam]), this may well be due to schwa epenthesis (alternation type D discussed below). Although the indirect phonological evidence, which is based on the process of liaison but which I cannot review here (cf. Encrevé 1988 and references cited there), is favourable to the hypothesis that the feminine suffix is schwa, there is no really strong phonetic evidence that the schwa is present at all, or that it is ever deleted specifically in the context before a vowel.

On the other hand, we do have evidence that the constraint against schwa in front of a vowel is operative in French; although schwa can be rather freely distributed over all positions in the French word, there are no words in which it immediately precedes a vowel (either a full vowel or schwa). French shares this property with many other languages, including Dutch, English, German and Indonesian. In Dutch we actually have some pieces of dynamic evidence: if the schwa-final word (elite [elit@]) is followed by a vowel-initial suffix (-air [Er]), the schwa gets deleted (elitair [elitEr] `snobbish').

I propose that this behaviour of schwa can be understood within Optimality Theory in the following way. As we have seen above, we need to posit a constraint against the occurrence of schwa. This constraint outranks most faithfulness constraints on schwa, but it is itself dominated by several well-formedness constraints on syllable structure. Schematically:

() Well-formedness >> *Schwa >> Faithfulness.

This ranking implies that schwa never surfaces unless it is required to do so by the syllable structure constraints. This in turn means that it is almost impossible to say whether we see deletion or epenthesis at work in a given case: schwa is deleted wherever it is not necessary because of *Schwa and inserted wherever it is necessary because of high-ranking Well-formedness.

If an underlying schwa is both preceded and followed by a consonant, it needs to surface in order for the word to satisfy general constraints on syllable structure, in particular constraints against complex onsets (*[lkamrad]-[l@kamrad]). Yet if schwa is followed by a full vowel, there is no necessity for it to surface ([lami] is perfectly well-formed as far as syllable structure is concerned). This may be the reason behind the deletion of schwa.

A famous exception to the deletion of schwa before full vowels is formed by a class of words with a so-called h aspiré: a consonant position that was probably filled by a /h/ in earlier stages of the language but that it is presently phonetically empty, although phonologically it still behaves as a full-fledged consonant:


mets le dessus [mEl@dsy] 'put it on!'
retourne la [rturnla] 'return it!'
mets le dehors [mEl@d@Or] 'put it outside!'
rehausse la [r@osla] 'raise it again'

It is quite standardly assumed that this behaviour can be understood by assuming some more or less abstract consonant in this position. The schwa is therefore not adjacent to the full vowel in cases like these; hence, it is not subject to the deletion process.

Type B: Postvocalic schwa deletion. A similar line of reasoning may help to explain why schwa does not surface on the righthand side of a full vowel. Also in this context, there is no necessity for schwa to surface. The syllable structure of [ami] is actually better than the structure of hypothetical [ami@], because the latter has one extra onsetless syllable.

Again it is rather hard to find dynamic evidence for a deletion rule -- the evidence that is cited by Noske 1993 for instance involves once again the feminine suffix which reputedly gets deleted after a full vowel: risée `laughed at, FEM' /rise/+/@/->[rise], but as I have already pointed out above, word-final schwa gets deleted in many other cases as well -- but there is some static evidence in that the relevant constraint is at work in the languages just mentioned: we don't find any (monomorphemic) words in which a schwa immediately follows a full vowel.

The question may now arise why not all vowels get deleted when they occur next to another vowel. I think that the reason for this is that deletion of schwa is less costly than deletion of other vowels, due to the fact that the schwa literally is a substructure of all the other vowels. If it is allowed to delete a full vowel in a certain context, this implies that it is allowed to delete a root node in that context; hence, also schwa will be deletable in that context.

Type C: schwa deletion in a 'two sided open syllable'. A schwa which occurs in an open syllable that is preceded by another open syllable may get deleted. In practice this means that an underlying schwa occurring in the context ..V$C__$C(C)V... can disappear on the surface. In contradistinction to the two previous cases, this type of alternation is either optional or otherwise subject to stylistic conditions.

The following relatively famous examples serve to illustrate this point.


Henri devrait partir

'Henri would have to leave'

Jacques devrait partir

'Jacques would have to leave'


In this case it looks as if underlying schwa's should be deleted in principle, but deletion is blocked if it would result in an unwellformed cluster of consonants, in particular one which cannot form a reasonable syllable structure. Certain additional stipulations are necessary, however, since [dvr] is not otherwise a wellformed onset cluster of French, yet the syllabification that most speakers would assign to Henri devrait partir seems to be [ãri.dvRE.paR.tiR]. This can be understood in several ways. It is possible for instance that there is still a `trace' of schwa between [d] and [v]; this is the solution adopted in Charette (1991) modulo technical details. It is also possible that there is a distinction between two rounds of syllabification (e.g. lexical and postlexical); again modulo technical details, this seems the type of solution favoured by Noske (1993).

Type D: Schwa epenthesis in the environment CC] __ [C. A schwa is (optionally) inserted between two words, if the first word ends in a cluster, and the second word starts with a consonant. In this case syllable structure seems the driving force (rather than a force blocking deletion); clusters of consonants are avoided also in this case:


un contact pénible [õekõtakt(@)penibl] 'a painful contact'
un index formidable [õeEdEks(@)fORmidabl] 'a terrific index'

Maybe we can understand these examples in the following way: complex codas are dispreferred in French. A schwa is needed to support the final consonant in a cluster. However, schwa can never surface at the end of a phrase in French, as we will see below. It also cannot appear (is deleted) before a vowel. The only case in which we therefore see the supporting schwa appear on the surface is phrase-internal before a consonant. The fact that we have epenthesis in these cases is therefore readily understood; the fact that the epenthetic vowel is schwa also is readily understood. Schwa is the simplest vowel of all; therefore it is more economical to insert this vowel than it is to insert any other.

Charette (1991) observes that within compounds, stress also seems to play a role in the behaviour of schwa. A schwa has to surface if the first component ends in a consonant cluster and the second is monosyllabic; if the second is polysyllabic, the schwa cannot surface (according to Charette, a similar pattern is found in Québec French in phrases as well).

porte-clefs 'keyring'
garde-fou 'railing'

porte-manteau 'coat rack'

garde-manger 'meat safe'

Since stress is on the final syllable of the compound, it does not seem unreasonable to suppose that it plays a role in these examples. The question is of course why schwa should surface when it is immediately followed by a stressed syllable. Charette's answer to this is that French feet are iambic, and that the unstressed syllable in the foot should be a schwa. These assumptions are not uncontroversial, but as far as I know alternatives to Charette's proposal are yet to be worked out.

Type E: Schwa deletion in phrase-initial syllables. The existence ot this property, mentioned by Noske (1993), seems subject to debate: not all speakers of French agree on the data. In any case, some speakers allow alternations such as the following:

revenez (demain) [r@v@ne]~[rv@ne] 'demain'
te fais (pas de bil) [t@fE]~[tfE] 'don't worry'

Syllable onsets [rv] or [tf] are not normally allowed in French. It is possibke that the initial syllable of the word allows a more complex structure (as it does in Polish, cf. Rubach and Booij 1990).

According to Charette (1991) this type of deletion is restricted to bisyllabic words: cheval `horse' may be pronounced as [Sfal] , but chevalier `knight' is [S@valje]. In this case, it looks as if stressed syllables allow more complex onsets. On the other hand, the explanation given here seems to run counter to the explanation of porte-clefs vs. porte-manteau just given. In this case, French speakers do not try to build a `perfect' iamb [S@val].

Type F: Schwa deletion in phrase-final syllables. This final type of `alternation' is one of the most important ones: schwa never surfaces phrase-finally (except in Midi French, Durand 1986).


je vois l'autre [Z@vwalotR] 'I see the other'
voilà mon oncle [vwalamõnõkl] 'there is my uncle'
la terre est plate [latErEplat] 'the earth is flat'
la route est longue [laRutElõg] 'the road is long'

It is not immediately clear why schwa should be deleted in final position. One possibility, suggested in Van Oostendorp (1995), is to extend the scope of the constraint FinalC to the phrase. The constraint would thus say that all phrases have to end in a consonant. Deletion of full vowels under this view would be blocked by the fact that this would involve deletion of too many features. Deletion of schwa, on the other hand, is less costly, and therefore, it is allowed in this particular configuration.

The solution of Government Phonology is technically a little bit different. According to Charette (1991), a domain-final empty nucleus is exceptionally licensed to remain empty in French. In any case, it seems as if something special needs to be said about the end of the phrase in French. No special theory is needed, however, in order to describe why it is exactly schwa that behaves as special in this particular spot.

4. R-schwa

Schwa not only alternates with zero in languages of the world: it also often alternates with full vowels. There is an instance of this in French: schwa alternates with [E] in the following contexts in this language (Dell 1973/1985):

It seems reasonable to assume that underlying schwa turns to [E] in certain environments: namely when it is stressed or when it is in a closed syllable. This is not surprising; schwa disfavours stressed positions or closed syllables also in other languages, as we will see below. This thus seems to be an instance of fortition rather than reduction. We may of course wonder why schwa turns to [E] rather than to some other vowel; a good theory of segmental structure should provide us with an answer to this question.

Reduction to schwa is well-known from the study of Germanic languages. A very simple theory of vowel reduction in English is presented in Chomsky and Halle (1968). The reduction rule in this work basically amounts to (9):

(9) Unstressed vowels turn to schwa.

Syllable structure plays an important role in the analysis of E-schwa; in the case of R-schwa it is often stress that seems to trigger the alternation.

Within Optimality Theory, there are two possible ways to explain (9). In Alderete (1995) it is proposed that in principle all vowels turn to schwa. Schwa is the unmarked vowel and Alderete assumes that there is a constraint which requires all vowels in the word to be unmarked (I will call this constraint TurnSchwa for the sake of simplicity, but it can of course be seen as an instance of the general constraint *Structure). On the other hand there are position-specific faithfulness constraints that specifically require stressed syllables to be faithful to the underlying structure. These position-specific constraints I will shorthand here as Faithful-Stress.

An analysis for the English word apron would now run along the following lines.



Alderete (1996) claims that FaithfulStress constraints can also be put to use in analyses of phenomena other than reduction. For instance, he shows that the fact that epenthetic vowels usually avoid stressed positions (even if they are not schwa) can be understood in terms of this constraint. It is worse to epenthesize into a stressed position and violate both FaithfulStress and Faithful, than to epenthesize into an unstressed position in which case only Faithful will be violated.

Yet there are also problems connected to an approach based on position-specific faithfulness. We have already seen evidence for the fact that we need to posit a constraint against schwa (e.g. *Schwa, (2)) within the universal inventory of constraints, if only because there are languages which do not have schwa. We then have a ranking FaithfulStress >> *Schwa >> Faithful within our factorial typology. This however gives us a result which is probably absurd: a language in which underlying schwa can only surface in a stressed position:



More in general, a FaithfulStress account implies that there are more contrasts possible in stressed positions than in unstressed positions. While this is true in general, it is not true in the case of schwa. The prediction that there are languages in which we can only have schwa in a stressed position does not seem to be borne out.

An alternative is to make TurnSchwa rather than faithfulness into a position-specific constraint, or a set of such constraints. We then get some form of projection constraints, as proposed in Van Oostendorp (1995):

(12) Project-Foot

a. Unstressed vowels turn into schwa

b. Schwa wants to be unstressed

These constraints fit into the general theory of Projection outlined in (4): (12b) says that a vowel in the head position of a foot needs to have some minimal feature structure; (12a) says that vowels with a certain feature structure desire to be in the head of a foot. If they are not, they turn into schwa. An analysis of English apron would run along the following lines.


Project-Foot (a)
Project-Foot (b)

A language that only allows schwa in stressed positions is not part of the typology in this system: if we rank *Schwa very highly, we get a language in which schwa does not appear at all; if we rank Faithful very highly, we get a language in which schwa can in principle occur in any position.

Another advantage of this approach is it that it explains why there are languages in which underlying schwa's are just as stressless as the epenthetic ones or the ones which are the result of reduction. (Underlying schwas will be the topic of the next section.) There are also problems with this approach, however: most importantly, the fact that generally stressed positions do license more feature contrasts than unlicensed ones, and that epenthetic vowels are unstressable requires an explanation.

Regardless of which of the two lines of research will turn out to be most fruitful, the analysis of English vowel reduction should be more sophisticated. Burzio (1994) argues, quite convincingly in my view, that the traditional view on reduction (every unstressed vowel gets reduced to schwa) is too simple. It causes all sorts of problems, because it forces us to assign stress in an extremely complicated way and furthermore to introduce the formal device of destressing rules into the theory in order to undo some of the results of the stressing rules. Burzio (1994) proposes to simplify the stress rules. As a result of this, the conditions on reduction get more complicated. Burzio's main point is that reduction is generally blocked in unstressed closed syllables ending in an obstruent, but not in unstressed syllables ending in a sonorant; in this way, he can get rid of the so-called `sonorant destressing' rules.

I find Burzio's arguments convincing because the difference between closed and open syllables, and the sonority of a following consonant also play a role in the behaviour of other types of schwa: we have already seen that schwa cannot occur in closed syllables in French, and I will give an example of S-schwa in Dutch below. It therefore seems more natural to say that unstressed vowels resist reduction because they occur in a closed syllable, than to say that they are stressed, and possibly later destressed. There is no independent evidence for the latter assumption.


I will now turn to a language in which reduction is subject to even more factors than it is in English. This language is Dutch (the following discussion is based on Booij 1981, 1995, Kager 1989, Martin 1969, Van Oostendorp 1995). Factors that presumably play a role in Dutch vowel reduction are:

i. Frequency:

Vowels in more frequent words tend to get reduced more easily than those in non-frequent ones.

ii. Differences between sociolects and dialects, and differences betwee style levels

Certain dialects/sociolects/style levens show more reduction than others.

It is hard to see how factors (i) and (ii) could possibly be taken into account within generative grammar; presumably, this is the reason why they are usually ignored in the literature (but see Hinskens et al. (1997) for several recent attempts to integrate some of these concepts into phonological theory). We will not go into these factors either. Other factors can be analysed more easily:

iii. Syllable type:

Vowels in open syllables are easier to reduce than those in closed syllables; in some idiolects (e.g. the one of Booij 1981, and the one of the author of this article, but not in the one of Kager 1989) reduction in closed syllables is only possible if the following consonant is deleted.


/bEnzin@/ [b@zín@] (*[b@nzín@]) `petrol'
/kAntor/ [k@tór] (*[k@ntor]) `office'

This is the same restriction as the one that Burzio (1994) claimed to exist for English. Yet in the literature Dutch there is less controversy that the purported restriction on open syllables belongs to the theory of reduction rather than to the theory of stress assignment. The reason for this is that reduction is optional in Dutch; apparently it is more attractive to say that reduction can optionally apply to invariably assigned metrical structures than it is to say that the stress rules themselves are optional.

I think that the restriction can be understood in the following way. A closed syllable is more complex than an open syllable. A minimal vowel is not sufficient to license such a complex structure. A projection constraint relating the quality of the vowel to the rhyme structure of the syllable in which that vowel occurs is therefore to be preferred.

iv. Stress position:

In Dutch -- as in English and apparently any other language in which reduction is dependent on the position in the metrical structure -- only unstressed vowels get reduced. As a matter of fact, we can distinguish between two types of unstressed vowel (Van Zonneveld 1985). Vowels in so-called `semi-weak' positions -- immediately following a stressed position -- are easier to reduce than those in `weak' positions.


/fonoloGi/ formal style [fònoloGí]
  less formal style [fòn@loGí]
  informal style [fòn@l@Gí]
  impossible *[fònol@Gí]

Similar facts can be found in other languages as well (Burzio 1994:113). In English, tat[@]magouchi seems preferable to tatam@gouchi. Jacobs (1989) observes a similar patterning for syncope in the history of French:

(16) Latin: similitudinem -> simlitudinem -> Old French: sembletume `resemblance'

The precise way to describe the difference between the two types of unstressed position is dependent on our precise assumptions regarding metrical structure. If we assume, with Burzio (1994) that feet can be ternary, we should distinguish between `foot-internal' and `peripheral' unstressed syllables. If we work in a theory in which only binary feet are allowed, we should distinguish between weak syllables in a foot and syllables which are somehow left unfooted by the parsing algorithm.

Vowel reduction to low vowels in Russian (/stol-á/ - stal-á) only occurs in the syllable directly preceding the stressed syllable (henceforth: the pretonic syllable henceforth). All vowels reduce to [@] in unstressed, non pretonic syllables (the following facts have been taken from Alderete 1995).


Nom. Sg. stól slóv-o
Gen. stal-á slóv-a
Dat. stal-ú slóv-u
Instr. stal-óm slóv-om
Loc. stal-é slóv-e
v[ó]d@j nom. pl.
v[a]d-á nom. sing.
v[@]davóz 'water carrier'

Here it seems that the position next to (on the lefthand side of) a stress is less likely to undergo full reduction than vowels in other positions.

v. Position in the word:

Vowels in absolute word-initial or absolute word-final position do not reduce. This observation is absolutely surface-true in Dutch. No matter what the style of speech is, we do not find reduction of the vowels that are absolutely word-initial or absolutely word-final. The /e/ is reduced quite easily in Dutch, for instance, but not if it is the first or the last segment of the word, as the following segments are intended to show:


/plezir/ [pl@zír] `fun'
/eGal/ [egál] (*[@Gal]) `even'
/tOfe/ [tÓfe] (*[tÓf@]) `toffee'

We may think of this as some kind of alignment effect of phonological to morphological structure: the edges of output words should correspond exactly to their morphological specifications.

An alternative explanation for the word-initial vowels is that syllables consisting of only schwa is disallowed; a syllable needs some minimal specification (Cohen et al. 1959). An argument for this alternative may be that vowels next to /h/ also cannot be reduced:


/plezir/ [pl@zír] `fun'
/helas/ [helas] (*[h@las]) `alas'

We will see in the next section that also underlying schwas cannnot occur in absolutely word-initial position; this is true for many other languages as well, including German, Indonesian and French.

There is also an alternative explanation for the non-reduction at the right edge of the word. Some authors (e.g. Brink 1970) have claimed that Dutch schwa is always the result of reduction: words such as mode `fashion' are underlyingly specified as [mode]; in that case reduction is obligatory in these environments and the word toffee [tOfe] should somehow be marked as an exception (it is true that there is only a handful of words in Dutch ending in unstressed [e]).

vi. Vowel quality:

The quality of the underlying vowel plays a role in the choice between reduction and non-reduction in many languages. In English, for instance, high vowels are not as likely to undergo reduction as non-high vowels (Chomsky and Halle 1968). Mid vowels in Dutch are easier to reduce than high and low vowels; front vowels are easier to reduce than back vowels; rounded vowels are easier to reduce than unrounded ones. This is illustrated in the following table (from Kager 1989):


Reduces in Weak position Semi-weak position
/e/ formal formal
/a/ semi-formal semi-formal
/o/, /i/ semi-formal informal
/u/, /y/ informal excluded

Interestingly, Dressler (1973) found more or less the same hierarchy of segments for Breton vowel reduction to schwa. As far as I am aware, the structure of this hierarchy is still in need of an explanation.

5. S-schwa

R-schwa can be a non-underlying schwa which is the result of a reduction operation; e-schwa can be a non-underlying schwa which is the result of minimal vowel epenthesis. Both types of schwa can also be underlying; the alternation in the case of r-schwa can then be seen as a case of fortition. The alternation in the case of e-schwa can be analysed as a result of deletion. As a matter of logics, we also expect underlying schwas which do not alternate; these vowels I call s-schwa.

Even though s-schwas are stable in the sense that they occur both in the underlying and in the surface structure of a word, they still often display some special behaviour. The factors involved may be familiar by now: s-schwa does not occur in an unstressed position or in a syllable that is too complex.

We have seen that schwa in French cannot occur in a closed syllable. In Dutch, there are also several restrictions on the syllable shape in which a schwa can occur. For one thing, we do not find schwa in an onsetless syllable. The static evidence for this is that the vowel does not occur as the first segment of a word (*[@Gal]), and that it does not occur immediately after a vowel (*[hi@t]). This restriction seems quite common in languages of the world: it can be found also in Indonesian, Slavic languages, and French. On the other hand, Dutch schwa also cannot occur in a syllable with a complex onset (*[papavr@]). I have found no other languages yet in which this restriction seems to hold (except maybe related Germanic languages such as German and certain dialects of Norwegian, Kristoffersen 1995).

The restrictions on schwa are not just restrictions on the onset; there are also restrictions on the rhyme of a schwa syllable. Other than other (short) vowels (harp `id.' [hArp], Belg [bElx] `Belgian', ramp [rAmp] `disaster'), schwa cannot be followed by two tautosyllabic consonants, except if the second consonant is coronal (*[ad@rp], *[kat@lx], *[ad@mp], anders [And@rs] `different'). Marked coda's, just like marked onsets, are not allowed.

Even within the class of schwa syllables with a simple coda, we find a clear asymmetry. Schwa has a preference for syllables closed by a sonorant: words such as hennep, tinnef, etc. are very rare. It is not uncommon for languages to require that codas can only contain a sonorant. There is no evidence, however, that syllables headed by a full vowel are subject to this requirement in Dutch.

In general it seems that the requirements on schwa syllables are much stronger than those on syllables headed by other vowels. In Van Oostendorp (1995) I proposed that this is due to the theory of Projection in (4): schwa as an (almost) empty vowel does not have sufficient material to license all of the syllabic nodes that are necessary in order to get complex onsets or codas. Simple vowels are not allowed to project complex syllables.

Schwa in stress

In most languages that have (underlying) schwa, the vowel is left unstressed. Roughly speaking, this can happen in two ways: schwa can preferably end up in the weak position in a foot; this seems to happen in a language like Dutch, in whcih stress always falls on the penultimate syllable if the final syllable contains a schwa (ballade [bAlád@], *[bÁlad@], *[bAlad'@]); if the final syllable contains a full vowel, more (lexical) variation is possible (pánama, pyáma, chocolá). Schwa thus is not completely invisible for stress; under the standard assumption that Dutch feet are trochaic, we should say that schwa can only occur in the weak syllable of a foot.

Alternatively, schwa can behave as if it is completely invisible for the stress algorithm; an instance of this is Indonesian. In this language (Cohn 1989, Cohn and McCarthy 19xx) primary stress is usually on the penultimate syllable of the word (a-b) secondary syllable is on the first syllable and on every odd syllable counting from the right-hand side of the word. An exception to this are words which contain a schwa. These can be best understood as if schwa were not present at all:


a. cát `print'
b. cári `search for'
c. bicára `speak'
d. bìjaksána `wise'
e. kòntinuási `continuation'
f. òtobìográfi `autobiography'
g. àmerikànisási `Americanization'
a'. b@rí `give'
b'. s@t@láh `after'
c'. gám@lan `Indonesian orchestra'
d'. apárt@men `appartment'
e'. c@rìt@ra `story'
f'. p@r@mpúan `woman'
g'. dìf@rensiási `differentiation'
h'. dìv@rsifikási `diversification'

It is unclear, at least to me, how exactly these facts should be understood. We may understand why schwa cannot occur in the stressed position of a foot, but this does not clarify why it cannot occur in the unstressed position of a foot. Particularly intriguiing are the cases in (g') and (h'). There is no secondary stress on a full vowel in these cases: we do not find [dìf@rénsiási]. According to Cohn (1989) and McCarthy and Cohn (19xx) the reason for this is that stress only sees full vowels: a trochaic foot is therefore built over the first two full vowels of the word, i.e. the foot structure of this word is [(dìf@ren)si(ási)]. It thus looks as if in this case schwa is not strong enough to project a grid mark at all, not even one which appears in the weak position of a foot.

Schwa remains unstressed also in most dialects of Yupik. Different dialects of Yupik choose different means to escape an unstressed schwa. Stress is generally on the second syllable of the word in Yupik; the stressed vowel is lengthened (/qayani/ - [qaya:ni]). If this syllable contains schwa, the vowel is deleted in the General Central dialect of Yupik, as if it wants to escape a stress position. Stress is then assigned to the first syllable of the word.


/qayapigkani/ [qayáápixkani] `his own future kayak'
/qan@qa/ [qánqa] `my mouth'

GCY schwa therefore should be classified as an e-schwa; it is however different from the cases of e-schwa we have seen above because the triggering factor for the alternation is not syllable structure, but rather stress. A pattern that is fairly similar to the one of GCY can be found in Alutiiq Yupik.

In other dialects of Yupik, schwa does get stressed, but this stress does not trigger lengthening, as it does on other vowels. Different dialects of Yupik may solve this problem in different ways. In the Norton Sound dialect, for instance, the consonant following the schwa gets lengthened: /at@pik/ - [at'@ppik]. In this way, the stressed syllable is still heavy, even though schwa needs not be lengthened in order to achieve this result. In the Central Siberian dialect of Yupik, on the other hand, nothing happens to the stressed schwa-headed syllable at all: /at@pik/ - [at'@pik].

6. Conclusion

Schwa is special in the sense that it has a more limited distribution than other vowels: for instance it can often only occur in a rather simple syllable type, and only in an unstressed position. Schwa is also special in the sense that it is the target of reduction and deletion; furthermore it is easier to delete schwa than it is to delete a full vowel. Phonology should be able to describe all of these facts.

In order to fully understand the behaviour of schwa, we need a fully developed theory of syllable structure, of metrical structure, of segmental structure, and of the way in which these different dimensions of phonological structure can interact. Inversely, while developing these subtheories, we sharpen our view of schwa. Most of the facts mentioned above could probably not have been raised without sufficient understanding of phonological structure, phonological derivation and phonological constraint interaction. Some of the facts mentioned here are still rather problematic for any phonological theory known to the author of this overview; chances are that new facts may be discovered while phonologists are developing their theories. I suspect that we will not have a satisfying theory of schwa untill we have a satisfying Theory of Everything.