An Introduction To Vowel Harmony

Chapter 1
An introduction to vowel harmony
The first issue to be addressed in this work is, of course, identifying the
phenomenon of vowel harmony. Even though vowel harmony is one of the
standard examples of phonological feature interaction in almost every
introductory textbook on linguistics, whatever its theoretical orientation,
there is little consensus on which phenomena exactly may be labelled
'vowel harmony' and which ones do not deserve this name, or which
characteristic property sets vowel harmony apart from other types of pho-
nological assimilation. On this issue, see for instance the discussion in
Clements (1976) and Anderson (1980). For the current purpose, a rather
rough statement will be fully sufficient. I regard vowel harmony as the
phenomenon where potentially all vowels in adjacent moras or syllables
within a domain like the phonological or morphological word (or a smaller
morphological domain) systematically agree with each other with regard to
one or more articulatory features. The presence of a certain feature speci-
fication (either underlyingly or in the surface form) on a vowel triggers a
systematic alternation in vowels which are in direct neighbourhood on the
syllabic or moraic level of representation with the result that the involved
vowels look alike with respect to the active feature.
(1) Vowel harmony

a. disharmony1 b. harmony2
σ σ σ σ σ σ
V V V V V V
[+F] [-F] [+F] [+F] [+F]
This description even though very vague excludes many types of Umlaut,
where (as in German) a neutral affix vowel imposes fronting on the preced-
ing root vowel. Consequently, I will have nothing to say about umlaut in
this work.
4 An introduction to vowel harmony
The generalisation that adjacent vowels agree with respect to a certain

feature is often overshadowed by other conflicting phonological wellform-
edness conditions in a given language. For one or the other reason certain
vowels may be excluded from harmony. They behave as neutral in that they
either do not undergo harmony, or do not trigger harmony in the neighbour-
ing vowels, or reject both the role as trigger as well as that as a target of
harmony. One of the most discussed examples of an obscured harmony pat-
tern is the vowel harmony found in Yawelmani (Archangeli 1985, Archan-
geli and Suzuki 1997, Cole and Kisseberth 1995, Dell 1973, Goldsmith
1993, Hockett 1973, Kenstowicz and Kisseberth 1977, 1979, Kisseberth
1969, Kuroda 1967, Lakoff 1993, McCarthy 1999, Newman 1944, Noske
1984, Prince 1987, Steriade 1986, Wheeler and Touretzky 1993) in which
the harmony requirement is obscured by the additional restriction that only
vowels of the same height have to agree in backness. A further com-
plication results from the interaction of harmony with a pattern of vowel
lowering. Lowered vowels behave as if they were high vowels with regard
to harmony. As root vowels, they trigger agreement in high affix vowels
but not in low ones, which results in a surface violation of the height
restriction. I will go into the details of these patterns later.
The view of vowel harmony as given above still covers a large range of
phenomena attested in diverse languages scattered on all continents of the
earth. In the following I will briefly discuss which types of vowel harmony
occur with regard to the features that are affected (section 1.1.1), with
regard to the attested combinations of features (section 1.1.2), featural
interaction with consonants (section 1.2), the question within which domain
the process is found to apply (section 1.3), as well as other typological pos-
sibilities of distinguishing harmony systems, like the distinction between
systems containing opaque vowels (section 1.4.1) and those having trans-
parent vowels (section 1.4.2), as well as the difference between root control
(or morphological control) and dominance (section 1.5). I hope to give the
reader an impression of the wide range of topics which relate to vowel har-
mony. I will first discuss those aspects of harmony which I will have noth-
ing or little to say about in the remainder of this work, and afterwards I will
deal with those phenomena which are the central subjects of this study.
Harmonic features 5
1.1 Harmonic features
In this section, I will give a brief overview on which phonological features

can be active in harmony systems. The features which will be referred to in
the following are articulatorily based, as proposed by Chomsky and Halle
(1968).3 They denote the relative placement of the articulator in comparison
to its neutral position: [±back], [±high/±low] refer to the relative placement
of the back of the tongue in the oral cavity, [±ATR] specifies the position
of the tongue root (whether it is advanced or retracted), and [±round],
specifies whether the lips are rounded during the articulation of a vowel or
not. This classification differs only slightly from that made in the Interna-
tional Phonetic Alphabet (revised to 1993, updated 1996), where the height
dimension is conceived of as the degree of opening of the articulatory appa-
ratus, with the high vowels being 'close' and the low vowels being 'open' at
the ends of the continuum.4
(2) The IPA vowel chart
There are, however, other possibilities to express the differences and

similarities among vowels, such as, for instance, the Jakobsonian (1951)
feature inventory, which reduces all properties of vowels as well as conson-
ants to the three features 'gravity', 'compactness', and 'diffuseness'. These
features are based on acoustic criteria. Another possibility is the reference
to the involved articulator or place of articulation as proposed by Clements
and Hume (1995). In utilising feature labels like 'labial', 'palatal', 'coronal',
and 'dorsal', again an attempt is made to unify vocalic and consonantal fea-
tures, acknowledging the various ways of interaction among both segment
types. Yet another possibility is an analysis in terms of abstract primitives
or radicals which eventually combine to more complex vowels, as proposed
in work by Kaye, Lowenstamm and Vergnaud (e.g., 1985). They assume
the three radicals I, U, and A to be the basic abstract components to derive
all other vowels.
There are two reasons for not using the latter three choices here. The
first is mere convenience, driven by the fact that the SPE/IPA-style features
are used in most of the literature. The second reason is that I do not intend
to contribute any argument in favour of one or the other approach, even
though the feature geometric approach to unified consonantal and vocalic
features as well as privative feature theory will be touched upon, where
they are relevant for the current discussion (see in particular sections 1.2
and 1.6, respectively). Similarly, I will only briefly enter the discussion
whether features are organised hierarchically as in feature geometry
(Clements 1985, see in particular Odden 1991 on vocalic feature geometry)
or not (as in the unstructured matrices in Chomsky and Halle 1968).
Instead, I will work on the hypothesis that vocalic features may be an-
chored prosodically, rather than in any root nodes or segments, but even
this statement serves as a working hypothesis only.
After these clarifying remarks on the nature of assumed features I will
advance to the actual topic of this section, an overview of attested harmony
patterns in terms of the affected features.
1.1.1 Single feature harmonies
One type of harmony affects the dimension of backness or palatality. This

type of harmony can be found in the Finno-Ugric languages (Kiparsky
2000a), as well in Turkic languages, Caucasian languages, in the North-
American language Yawelmani (or Yowlumne)5, in Chamorro (van der
Hulst and van de Weijer 1995 and references cited there), and in many
other languages. In the example below some Finnish words are listed which
contain exclusively front vowels (3a) or exclusively back vowels (3b).
Harmonic features 7
(3) Backness (or palatal) harmony in Finnish

a. Front vowel words
väkkärä 'pinwheel' käyrä 'curve'
pöytä 'table' tyhmä 'stupid'
b. Back vowel words

makkara 'sausage' kaura 'oats'
pouta 'fine weather' tuhma 'naughty'
(Ringen 1975, Kiparsky 1981, van der Hulst and van de Weijer 1995: 498)
Roundness (or labial) harmony is attested in Southern-Payute, Khalkha-

Mongolian and many Turkic languages (Kaun 1995a,b). Below I give an
example from Turkish. The vowel in the suffix alternates in accordance
with the root vowel in backness and roundness, but not height.
(4) Roundness harmony in Turkish (combined with backness harmony)

a. Unrounded words b. Rounded words
kilim-im 'my carpet' gül-üm 'my rose'
ev-im 'my house' köy-üm 'my village'
kiz-im 'my girl' kuV-um 'my bird'
kaz-im 'my goose' koz-um 'my walnut'
(Kaun 1995: 79)
Languages displaying solely labial harmony seem to be quite rare or non-

existent (van der Hulst and van de Weijer 1995: 523). Usually labial
harmony occurs together with another type of harmony or is restricted to
vowels which accidentally agree with respect to a second feature like height
or backness.
Height harmonies are found predominantly among African languages.
The dimension of height is directly affected in Bantu languages like Shona
or Kikuyu. An example from Shona is given below. In the leftmost column
the first vowel of the affix is preceded by a nonhigh vowel (i.e., e or o) and
surfaces as e, while it surfaces as i in the other column where the affix
vowel is preceded by a high vowel (i.e., i or u).
(5) Shona height harmony

a. Nonhigh vowels
per-era 'end in' tond-esa 'make to face'
son-era 'sew for' om-esa 'cause to get dry'
vere1g-eka 'be numerable' sek-erera 'laugh on and on'
b. High vowels
ip-ira 'be evil for' bvum-isa 'make agree'
bvis-ika 'be easily removed' pind-irira 'to pass right through'
(Beckman 1997: 1)
The position of the tongue root, with an advanced tongue root (ATR), as,
e.g., in the vowel [o], in opposition to a retracted tongue root (RTR), as in
the vowel [o], is the active feature in Niger-Congo languages (e.g. Yoruba,
Wolof, Fula, Diola Fogni) and many Nilo-Saharan languages (Kalenjin,
Päkot, Maasai, Luo). Below are listed some Yoruba words, which have
only mid vowels. In such words, all vowels have to have the same ATR
specification. Forms like *CeCo or *C(Co (with 'C' standing for any con-
sonant) are not attested in Yoruba. They violate the harmony requirement
on the ATR specification of vowels in a word. With words containing high
or low vowels as well the Yoruba pattern is more complex as will be dis-
cussed in chapters 4 and 6.
(6) Yoruba ATR harmony among mid vowels

a. ATR words b. RTR words
ebè 'heap for yams' (s(¼ 'foot'
epo 'oil' (¼ko 'pap'
olè 'thief' ob(¼ 'soup'
owó 'money' oko 'vehicle'
(Archangeli and Pulleyblank 1989: 177)
Geographically isolated occurrences of tongue root harmony – henceforth

'ATR harmony' – can be found for example in Nez Perce, a native North
American language, as well as in the two Afro-Asiatic languages Somali
(Cushitic) and Tangale (Chadic) (see Hall et al 1973 for an overview of
ATR harmony systems in African languages, and Hall and Hall 1980 and
Anderson and Durand 1988 on Nez Perce).
Additionally, there are several types of harmony which involve other
features than vocalic place features. In particular we find nasal harmony
Harmonic features 9
and retroflex harmony (see van der Hulst and van de Weijer 1995: 525).
For reasons of thematic and theoretical restriction, I will not go into the
details of these phenomena. For work on nasal harmony the reader may
consult Cole and Kisseberth (1994), Walker (1996, 1998), and much of the
work of Glyne Piggott (1992, 1996 and elsewhere).
1.1.2 Multiple feature harmonies
Besides the types of harmony which involve only one feature, there are
some combinations of features attested. In multiple feature harmony,
agreement between vowels is not only required for feature x in a language
but also for feature y. This pattern must be distinguished from cases where
a vowel changes two features to be opportune to a harmony requirement on
only one of the two features because the vowel system lacks the respective
allophonic vowel which differs only with regard to the active feature.
Warlpiri and Yawelmani are such cases. Warlpiri (Nash, 1979, 1986, van
der Hulst and Smith, 1985, Sagey, 1990, Cole 1991, Inkelas, 1994, Berry,
1998) has only the three vowels i, u and a (Nash, 1986: 65). In Warlpiri,
suffixes agree with their lexical host in terms of roundness and backness, as
shown in (324a,b).
(7) Warlpiri harmony

a. kurdu-kurlu-rlu-lku-ju-lu
'child-Prop-Erg-then-me-they'
'as for the children, they are with me'
b. maliki-kirli-rli-lki-ji-li
'dog-Prop-Erg-then-me-they'
'as for the dogs, they are with me'
c. minija-kurlu-rlu-lku-ju-lu
'cat-Prop-Erg-then-me-they'
'as for the cats, they are with me'
(Nash 1986: 86 cit. op. Inkelas 1994: 291, Berry 1998: 139)
However, if we consider (324c) as well it emerges that the harmonic feature

has to be backness, since with the root-final vowel a, the harmonic suffixes
turn out with a back vowel, which is disharmonic with the root vowel in
roundness. The conclusion is that roundness harmony is just a by-product

of backness harmony, triggered by the limits of the vowel inventory. Warl-
piri has neither a back unrounded high vowel nor a front rounded vowel. If
a nonlow vowel changes its backness specification in agreement with a
neighbouring vowel, it has to change its roundness specification as well,
because the system lacks a front rounded vowel as an allophone for the
back vowel and a back unrounded vowel as an allophone for the front un-
rounded vowel likewise.
In languages like Warlpiri, roundness harmony is triggered by the re-
stricted vowel inventory, rather than by a harmony rule or a harmony con-
straint.6 In contrast to this a mixed harmony system is one in which
agreement of more than one feature is not a consequence of an impover-
ished vowel inventory. Among such multiple harmonies we can further
distinguish those in which separate rules or constraints require agreement
of two or more features from systems in which one rule or constraint oper-
ates on a set of features.
The most well-known case of vowel harmony affecting two features
may be the combination of backness with roundness harmony in Turkish
and other Turkic languages (see 4, on p. 7). All vowels in a Turkish word
have to agree in backness, while agreement on the feature roundness is re-
stricted to high vowels only. Low vowels in affixes are never rounded even
if the preceding root vowel is rounded.7 This is illustrated with the plural
affix -ler/-lar below. In (8c,d,g,h), the suffix vowel in column 3 is preceded
by a round root vowel. Though it agrees with the root vowel in backness, it
does not with respect to roundness. The genitive suffix in the next column
has a high vowel which agrees with respect to both features.
(8) Turkish backness and roundness harmony

nom.sg. nom.pl. gen.sg. gen.pl.
a. ‘rope’ ip ip-ljer ip-in ip-ljer-in
b. ‘girl’ kÕz kÕz-lar kÕz-Õn kÕz-lar-Õn
c. ‘face’ yüz yüz-ljer yüz-ün yüz-ljer-in
d. ‘stamp’ pul pul-lar pul-un pul-lar-Õn
e. ‘hand’ elj elj-ljer elj-in elj-ljer-in
f. ‘stalk’ sap sap-lar sap-Õn sap-lar-Õn
g. ‘village’ kjöy kjöy-ljer kjöy-ün kjöy-ljer-in
h. ‘end’ son son-lar son-un son-lar-Õn
(Clements and Sezer 1982: 216)
Harmonic features 11
In this case, roundness harmony is not triggered by the limits of the vowel
inventory, since Turkish has the rounded and unrounded variants of both
the high front vowels as well as those of the high back vowels.
That in Turkish backness harmony can occur without roundness har-
mony is an indicator that here two rules or constraints are operative. Odden
(1991) discusses Eastern Cheremis and Tunica in which both features can
be analysed as being subject to one rule or constraint only.
I am not aware of languages where roundness harmony and height har-
mony co-occur. In Yawelmani (Kuroda 1967), only those vowels harmon-
ise which are of the same height. The restricted vowel inventory does not
allow a decision on whether roundness or backness is the harmonic feature.
If Yawelmani turns out to be a case of rounding harmony (which I doubt),
then this instance of harmony is at least tied to height harmony. Height har-
mony is no active process in this language, but if vowels agree in height,
then this agreement triggers rounding harmony.
Labial harmony may co-occur with ATR harmony, as attested in the
Niger-Congo language Dagaare (Bodomo 1997) or in Chumburung and Ig-
bo (see van der Hulst and van de Weijer 1995, p. 523 and references cited
there). However, the few instances of roundness harmony combined with
ATR harmony are quite dubious in that the second harmonising feature
could also be backness instead of roundness.
Compare in this respect the examples from Dagaare, cited from Bodomo
(1997: 24). In Dagaare, the imperfective suffix -ro/ -ro/ -re/ -r(/ -ra alter-
nates not only with respect to ATR but also with respect to roundness ac-
cording to Bodomo. In (9a) we find words with round vowels only, while
the vowels in the words in (9b) are all nonround. At the same time, the suf-
fix vowels alternate in backness. One of both alternations, i.e. either back-
ness or roundness alternation, is triggered by the asymmetry of the vowel
system, which has only front unrounded and back rounded vowels, and
lacks front rounded as well as back unrounded vowels (with the exception
of the vowel D). Since the suffix vowel is forced to change always both
feature specifications (i.e., backness and roundness), in order to harmonise
with the preceding vowel with respect to one of the two features, it cannot
be determined whether the active feature is backness or roundness. The
usual case is that roundness harmony occurs only in systems with a
phonemic rounding contrast among both front as well as back vowels. The
same holds for ATR harmony. Systems which do not distinguish pho-
nemically between ATR and RTR vowels are unlikely to display ATR
harmony. Dagaare lacks the rounding contrast, but, nevertheless, has a
contrast in the dimension of backness. So the active feature might also be

[±back].
(9) Dagaare multiple feature vowel harmony

verb root imperfective form
a. do 'to climb' duoro 'climbing'
tu 'to dig' tuuro 'digging'
ko to get dry' koro 'getting dry'
bo 'to look for' b8oro 'looking for'

mo 'to wrestle' m8oro 'wrestling'
b8 'to discuss/plan' b88ro 'discussing/planning'
b. mi 'to rain miire 'raining'

piiri 'to discover' piire 'discovering'
gbe 'to grind roughly' gbiere 'grinding roughly'
s,,r, 'to touch' s,,r( 'touching'

1m( 'to beat' 1m,(r( 'beating'
kp( 'to enter' kp,(r( 'entering'
c. kpa 'to boil' kpaara 'boiling'

la 'to laugh' laara 'laughing'
mar, 'to paste' mara 'pasting'
The data in (9c) suggest that the feature height or lowness plays an active
role here as well, since the affix vowel surfaces as nonlow with nonlow
root vowels and as low with a low root vowel.
According to Akinlabi (1997), the language Kalabari (Niger-Congo)
displays an instance of restricted backness harmony combined with ATR
harmony. This language has no phonemic rounding contrast either. Kala-
bari, however, shows evidence that the second harmonising feature (besides
ATR) is backness, not roundness. The low vowel a combines with high
vowels of either backness value, but it is not found with instances of e in
adjacent syllables within a word. Since a can hardly be described as being
rounded Akinlabi concludes that the harmonically active feature has to be
backness.
A more straightforward example of a combination of height harmony
with ATR harmony can be found in the Bantu language Kimatuumbi
(Odden 1991, 1996). Van der Hulst and van de Weijer (1995: 520, and
references cited there) mention Klao and Togo-remnant languages in this
respect as well. According to Odden, Kimatuumbi has a set of super high
vowels L and X, and a set of high vowels i and u, which are roughly equi-
valent to the sets i and u versus , and 8, which differ in their ATR speci-
fication (Odden 1991: 281, 1996: 5). The mid vowels e and o are lax, i.e.
RTR, and the low vowel D is retracted as well. Within a word, all vowels
have to agree with respect to ATR and height. The vowel D does not parti-
cipate in this pattern. It occurs freely with all sorts of vowels. The pattern
can best be observed with the causative suffix -iy which surfaces as -iy, -,y,
or -(y, depending on the quality of the preceding root vowel in (10).
(10) Height and ATR harmony in Kimatuumbi

a. Root plus low affix vowel b. Root plus nonlow affix vowel
út-a 'to pull' út-iy-a 'to make pull'
yíb-a 'to steal' yíb-iy-a 'to make steal'
y8ºy88t-a 'to whisper' y8ºy88t-,y-a 'to make whisper'
b,º,k-a 'to put' b,º,k-,y-a 'to make put'
goºonj-a 'to sleep' goºonj-(y-a 'to make sleep'
ch(º(ng-a 'to build' ch(º(ng-(y-a 'to make build'
káat-a 'to cut' káat-iy-a 'to make cut'
(Odden 1991: 281)
Hyman (2002), citing data from Paulian (1986a,b, 2001), describes a

harmony pattern involving three features. Kàlo¼1, a Bantu language spoken
in southern Cameroon, displays a combination of ATR harmony with
front/backness harmony and rounding harmony. In the examples in (11),
the last vowel of the word alternates according to the ATR, backness and
roundness specification of the preceding vowel. Underlying /a/ can be
realised as either a, e, (, o or o depending on the context. (Note that the
vowel u in the prefix in all these words has no possible more harmonic
alternants, i.e., the vowels 8, y, <, , are not part of the Kàlo¼1 vowel
system. For this reason, the language allows disharmony between this
prefix and the root vowel.)
(11) ATR, backness and roundness harmony in Kàlo¼1

a. /kù-tím-à/ → kù-tím-è 'creuser'
/kù-fùk-à/ → kù-fùk-è 'fermer'
b. /kù-fén-à/ → kù-fén-è 'dédaigner'

/kù-s(ºl-à/ → kù-s(ºl-(¼ 'éplucher'
c. /kù-pós-à/ → kù-pós-ò 'aboyer'

/kù-ko¼k-à/ → kù-ko¼k-o¼ 'tirer'
d. /kù-yàn-à/ → kù-yàn-à 'to play'

(Hyman 2002)
Complete copying of all vocalic place features from one vowel to the next
is observed in a variety of languages, such as, for instance, Yucatec Maya,
a native Central-American language (Krämer 1999, 2001) or Ainu, which is
spoken in Japan (Itô 1984). This instance of total harmony does not apply
across the board but is restricted to vowels in some morphemes only, as
illustrated by the data in (12). The imperfective and the subjunctive suffix
for intransitive verbs in (12a,b) always surface with the vowel quality of the
preceding root vowel, while the imperfective suffix for transitive verbs and
the perfective marker in (12c,d) have invariable feature specifications.
(12) Yucatec Maya complete harmony and disharmony:

a. Intransitive imperfective b. Intransitive subjunctive
ah-al wake.up-IMPF ah-ak wake.up-SUBJ
ok-ol enter-IMPF ok-ok enter-SUBJ
lub'-ul fall-IMPF lub'-uk fall-SUBJ
wen-el sleep-IMPF wen-ek sleep-SUBJ
kíim-il die-IMPF kíim-ik die-SUBJ
c. Transitive imperfective d. Perfective

yil-ik see-IMPF yil-ah see-PERF
tsol-ik explain-IMPF tsol-ah explain-PERF
put6-ik hit-IMPF put6-ah hit-PERF
(Krämer 1999: 184f.)
Since rounding is redundant in the vowel systems of languages like Ainu

and Yucatec, which have basically 5 vowels, disregarding length and tone,
(i, e, a, o, u), such instances of complete harmony can be described non-
redundantly as affecting the dimensions of backness and height only (see
Krämer 2001, on Yucatec). One might argue, however, whether these pat-
terns fall under the category of vowel harmony or should rather be treated
as reduplication, since harmony applies only between the root vowel and a
few affix vowels and the process is not iterative. Given that this type of
complete harmony affects maximally one vowel in a word it could be
regarded as separate from vowel harmony, just as umlaut.
In the following paragraphs I will introduce harmony systems in which
only a subset of the vowel inventory is subject to a harmony requirement as
well as cases in which both the trigger and the target have to meet a criter-
ion which activates the process.
1.1.3 Restricted harmonies
In many languages, vowel harmony applies only if the target and/or the
trigger of harmony meet certain criteria which can usually be defined in
terms of features which themselves are potential harmonic features. For
instance I have mentioned already that in Yawelmani vowels have to agree
in height in order for backness or roundness harmony to be applicable. The
affix -hin/ -hun is realised with the front high vowel when preceded by a
front high vowel and it has a back high vowel when preceded by a back
high vowel in (13a). In combination with a root containing a nonhigh
vowel, the high affix vowel invariably turns out as i (13b). The nonhigh af-
fix vowel in (13c,d) agrees in backness and roundness only with nonhigh
root vowels. In combination with a high vowel in the root, it surfaces as D,
regardless of the backness/roundness specification of the root vowel.
(13) Yawelmani height uniform harmony

a. xil-hin 'tangles, non-future'
dub-hun 'leads by the hand, non-future'
b. xat-hin 'eats, non-future'

bok'-hin 'finds, non-future'
c. xat-al 'might eat' d. xil-al 'might tangle'

bok'-ol 'might find' dub-al might lead by the hand'
(Cole and Kisseberth, 1995: 1f)
In Turkish, rounding harmony affects only high vowels. Only these vowels
alternate in their rounding specification in order to meet the rounding har-
mony requirement. Nonhigh affix vowels are invariably unrounded.
Kaun (1994) gives a typology of triggering factors for rounding har-

mony. She reports on languages where the undergoer of rounding harmony
must be a high vowel as in Turkish. Hixkaryana, Kachin and Tsou allow
rounding harmony only among high vowels. In Eastern Mongolian dialects,
Murut and Tungusic languages, both trigger and target must be nonhigh
vowels. Yakut allows rounding harmony only if trigger and target are of the
same height or if the target is a high vowel. In Chulym Tatar and
Karakalpak, rounding harmony applies among all front vowels, but among
back vowels the target must be a high vowel in order to show rounding
harmony. Kyzyl Khakass shows the same pattern of front vowels as the
latter two languages but is more restrictive with back vowels in that among
these trigger and target have to be high vowels. Finally, Kaun lists Kirghiz-
B and Altai where harmony applies unlimited among front vowels while
back vowels have to agree in height or the back target has to be a high
vowel.
Donnelly (2000) reports on Phuthi, a Bantu language, in which tongue
root harmony applies only among vowels of the same height. High vowels
agree in tongue root position with other high vowels, and mid vowels agree
with mid vowels, while the only low vowel is excluded from the pattern,
and neither high vowels trigger tongue root alternations in mid vowels nor
is the reverse the case. Tunica (Odden 1991) allows alternation to conform
to backness/roundness harmony only in low vowels.
What is most obvious from all these examples is that harmony patterns
are shaped by the dimension of height in most cases, and that the choice
between rounding and unrounding is more restricted in back vowels than in
front vowels.
1.1.4 Summary and discussion
So far we have seen which features can function as harmonic features and
which combinations of features are attested in the world's languages to
date. Chart (14) shows possible harmony types and possible combinations
of features in harmony. & marks attested types, while ' marks unattested
harmonies, and those harmony types of which no clear instances have been
found are indicated by a question mark.
(14) Types of vowel harmony

a. Single feature harmonies example language
& Palatal or backness harmony Finnish
? Labial or Rounding harmony Warlpiri?
& Height harmony Shona
& Tongue root (ATR/RTR) harmony Yoruba
& Nasal harmony Kikongo
& Retroflex harmony
' Length harmony
b. Feature combinations example language

i. & backness + roundness Turkish; Eastern Cheremis
ii. & backness + height Yucatec Maya (see also viii.)
iii. & backness + ATR/RTR Kalabari
iv. & height + ATR/RTR Kimatuumbi
v. ? roundness + ATR/RTR
vi. ' roundness + height --
vii. & backness + roundness + ATR Kàlo¼1
viii. ? backness + roundness +height Yucatec Maya ?
ix. ? backness + height + ATR Dagaare
The fact that roundness harmony predominantly co-occurs with backness

harmony, but not with height harmony and only in dubious cases with ATR
harmony, while height and ATR readily co-occur, has given rise to the
postulation of geometric hierarchies for vowel features in the literature.
Goad (1993) and Odden (1996) have argued for a height constituent
consisting of the two dimensions height and ATR/RTR, and Odden (1991)
has argued for a back/round, or 'colour' constituent within the theory of
feature geometry (Clements (1985).
(15) A feature geometric organisation of vocalic place features

X
height colour
ATR/RTR high low back round
The idea behind such a representation is that according to association

conventions, a node is delinked from its root node (the X in the
representation above) and instead the root node is associated with a neigh-
bouring node of the same type as the delinked node. Contrary to this
assumption it seems as if two-feature harmonies in many cases involve
separate actions, i.e. separate rules or separate instances of constraint inter-
action. The harmony pattern in a language like Warlpiri can quite elegantly
be analysed as double linking of the colour constituent, since vowels auto-
matically change their roundness specification, when backness alternation
is required by the harmony rule or constraint. However such an analysis
either ignores the fact that the alternation of two features is triggered by the
restricted vowel inventory, or it accounts for this fact redundantly. Har-
mony systems like the Turkish one cannot be analysed by the assumption
of a rule affecting the whole colour node, since backness harmony applies
even if rounding harmony is blocked. We would expect a pattern where the
inapplicability of harmony with regard to one of the two features blocks
harmony of the other feature. Furthermore, with such a representation it
seems somewhat astonishing that backness harmony is paired with ATR
harmony in some instances.
From an articulatory perspective it seems only natural that height and
ATR/RTR harmony co-occur since it is more natural for low vowels to
have a retracted tongue root, while it is more natural for high vowels to
have an advanced tongue root position in the sense that the articulatory
gestures are easier to combine than the opposite combinations (i.e., low
plus advanced and high plus retracted respectively, realised by a movement
of two parts of the tongue into the opposite direction, and which should be
particularly inconvenient for the latter feature combination). Hall et al.
(1974) argue that it is also more natural for advanced vowels to be front
rather than to be back, which can also be explained on grounds of the in-
volved articulatory movement of the different parts of the tongue. They cite
in this respect the example of Somali (as described in Tucker and Bryan
1966: 496), a language with ATR harmony in which all ATR vowels are
front, but vowels with retracted tongue root position occur as front and as
back. In this language all back vowels which are [+ATR] have become
fronted historically and retained their original rounding specification. The
result is the quirky vowel system in (16), which is asymmetric in the sense
that the ATR set lacks back vowels, while it is overcrowded with front
vowels, and the RTR set in contrast contains the basic set of front and back
vowels.
(16) The Somali vowel system

ATR set RTR set
front back front back
high i, ü , 8
mid e, ö ( o
low 4 a
The Somali harmony data in (17a) seen in isolation could mislead a linguist
to assume backness harmony to be at work here. The suffix vowel is back
with a back root vowel and front with a front root vowel. The data in (17b)
bring light into the misty scenery, which is obscured by the asymmetry of
the inventory. Here we see that it is not the backness of the root vowel
which triggers alternation in the affix vowel but rather the difference in
tongue root position in the root vowel which causes the alternation of the
low vowel from retracted to advanced tongue root, which is accompanied
by a backness alternation. In the first item in (17b), the root vowel is front,
but the affix vowel is disharmonically back. Thus, the harmonic feature
must be ATR/ RTR, not backness.
(17) Somali vowel harmony

singular plural
a. sab sab-o 'outcast'
'4d '4d-ö 'piece of meat'
b. r((r-ka 'the village'

gees-k4 'the horn' (Hall et al. 1974: 261)
From a historical perspective one might imagine that we see a system here
which is about to change from an ATR type of harmony to a mixed system,
including also backness harmony. If in a next historical step, the front re-
tracted vowels are moved backwards to minimise the articulatory tension
between tongue root position and the position of the back of the tongue, as
well as to maximise the contrast between the pairs i ~ ,, and e ~ (, respec-
tively, we end up with a system similar to that of Turkish. The backwards
movement of unrounded front vowels creates an apparent roundness
contrast among back vowels just as fronting of advanced rounded back
vowels has created an apparent rounding contrast within the set of front
vowels. Former ATR harmony can now be reinterpreted as backness
harmony. This is illustrated in the table in (18).
(18) Potential historical movement from ATR to backness

ATR set / front set RTR set / back set
(front) (back) (front) (back)
high i, ü (,) 8
mid e, ö (() ) o
low 4 a
With the observation as background that being front is articulatorily pre-

ferred among ATR vowels, the co-occurrence of ATR/RTR harmony with
backness harmony also seems to be more likely than the combination of
ATR/RTR harmony with rounding harmony.
Furthermore it can be observed in most vowel systems that front vowels
tend to be unrounded while the back vowels are usually rounded. This
interrelation of backness with roundness suggests that articulatorily it is
more likely that backness harmony combines with rounding harmony than
with height, and that roundness harmony is less prone to combine with
height harmony since the latter two features are largely independent from
each other (except the fact that the lowest vowel is rarely rounded).
In the chart below I listed the feature combinations which are articulato-
rily most convenient to be expressed on one segment, compared with the
most common feature combinations in multi-feature vowel harmony sys-
tems. The correlation is intriguing.
(19) Natural feature combinations

a. In single segments: b. In vowel harmonies:
[αATR] & [αhigh] ATR & height
[αATR] & [-αback] ATR & backness
[αback] & [αround] backness & roundness
To include the correlation of backness and roundness is not correct here if

we speak only of articulatorily natural pairs, since lip rounding and tongue
positioning are not physically related, in contrast to the physical troubles
caused by upward movement of the back of the tongue and downward
movement of the tongue root at the same time, which emerge during the
articulation of a high retracted vowel (i.e., 8 or ,). The universal correlation
between backness and roundness in five-vowel-systems may instead be the
result of contrast maximisation, either purely phonologically or also in
terms of sonority. Articulation in the backward part of the articulatory
apparatus combined with lip rounding maximally reduces the sonority of a
The role of consonants 21
vowel (Maas 1999), while the front unrounded vowels can be said to use all
feature dimensions except height to optimise sonority.
The low vowel D does not need to explore the backness dimension
because it already is maximally sonorous, the exhaled air has got the maxi-
mum space to escape. Fronting would be accompanied by raising and, thus,
reduce sonority in this case, whereas with all other vowels fronting in-
creases sonority, as does spreading of the lips.
A drawback to this assumption is of course that usually a sonority hie-
rarchy is assumed where D is more sonorous than e, which is more sonorous
than o which is more sonorous than i, which is more sonorous than u. Ac-
cording to the above argument, the sonority relation between o and i should
be the reverse.
Odden (1991) gives an argument from acoustic phonetics for his back/
round and height/ATR constituents. Backness and roundness are reflected
in the height of the second formant, while height and ATR have an impact
on the first formant. With respect to the back/round correlation also Kaun
(1993) observes that rounding lowers the second formant of a vowel and a
relatively low value of the second formant is the characteristic feature of
back vowels as well. Hence, (de)rounding enhances the backness contrast
(at least among nonlow vowels), in Kaun's view.
In short, there is a natural correlation of backness and roundness, but the
motivation might be different from that of the other featural correlations in
single segments as well as in harmony systems. I will leave this issue at this
point and move on to a brief discussion of the role of consonants in vowel
harmony.
1.2 The role of consonants in vowel harmony
Consonants may participate in various ways in vowel harmony. Van der

Hulst and van de Weijer (1995) list three basic types of consonantal inter-
ference: Consonants may alternate in their secondary place specification in
accordance with the surrounding vowels; they may influence the vowel har-
mony pattern by their secondary place feature; or they may influence the
harmony process by their primary place feature. I would like to add one
more type of intervention: consonants may also entirely block the harmony
process.
The most harmless case is consonants alternating in their secondary
articulation in agreement with the surrounding harmonic vowels. Van der
Hulst and van de Weijer cite as an example the language Bashkir, which
has backness and roundness harmony. In words with front vowels only,
dorsals are realised as velar, and in words with back vowels, the dorsal
consonants are uvular. Turkish is slightly more complicated. It has the
palatal and non-palatal variants of the consonants k, g, and l (Clements and
Sezer 1982: 233f.). There are minimal pairs among stems which only differ
in the palatality of one of these consonants. However, most of these conso-
nants agree in palatality with the vowel with which they share a syllable.
Some disharmonic stem-final consonants behave less harmlessly: they
trigger palatal assimilation to their specification in following affixes. Alter-
nation of the primary place feature is found for instance in the interaction of
consonants and vowels in Warlpiri. Labial consonants are always followed
by u in this language (van der Hulst and van de Weijer p. 529, Berry 1998).
A case where consonants shape the vowels of the whole word can be
found in Coeur d'Alene (Doak, 1992, Mithun, 1999, and references there).
Coeur d'Alene has a set of 'faucal' consonants (T, T , Tw, Tw, [. , [. w, µ, µ, µw, µw,
r. , r. ), which cause retraction or lowering in preceding vowels. The pattern is
illustrated in (20). The vowels which change to conform with the retraction
of the consonant are underlined, triggers are boldfaced.
(20) Coeur d'Alene

a. cíš-t 'it is long' i/(
c(š-alTw 'he is tall'
b. [. (c-p 'he became curious' (/$

t-[. $ºc-[. c-us 'he has curious eyes'
c. s-tpúm-lxw 'hide with fur' u/o

s-poºm-$lqs 'fur coat' (Doak 1992: 3-4)
In these examples, i is retracted or lowered to (, ( is retracted to $, and u is

retracted to o when followed by one of the faucal consonants. The fact that
( can be the output of harmony derived from underlying /i/, while ( is itself
lowered to $ suggests that this is a kind of chain shift operation and thus
slightly different from what was conceived of as vowel harmony here.
The last type of consonantal interference is blocking. In Yucatec Maya,
the harmony pattern exemplified above in (12) is blocked when more than
one consonant stands between trigger and target. In this case the affix
vowel always surfaces as D, regardless of the quality of the intervening
The role of consonants 23
consonants. Krämer (2001) proposes a prosodic analysis in which

consonants in moraic position can block harmony, because the process
applies between adjacent moras. Consonants in onset position are trans-
parent because they are not moraic and, hence, neither trigger nor target of
harmony.
In Tunica, back and round harmony targets low vowels only. Harmony
may spread through laryngeal consonants (i.e., glottal stop and h) but is
blocked by all other consonants (Odden 1991: 275). There are different
possibilities to handle this instance of blocking. One may assume that in
Tunica vowel harmony goes from place node to place node and therefore is
blocked by all consonants except those that lack a place node (i.e., laryn-
geals), while in other languages vowel harmony applies at the level of the
syllable head, and is thus not blocked by consonants (Archangeli and
Pulleyblank 1987). Odden proposes to explain the data by assuming a con-
dition on the target of harmony in Tunica. Targets must be [+low] in his
view, which is the feature that unites low vowels and laryngeal consonants.
Clements and Hume (1995) propose a feature geometry in which vowels
and consonants are characterised by the same place features. They only
differ in that the consonantal place node is closer to the root node, while for
vowels an additional vocalic place node is inserted in the hierarchy between
C-place node and the actual place features.
(21) Unified feature geometry for vowels and consonants

a. Consonants b. Vocoids
oral cavity vocalic
[continuant] aperture
C-place V-place
...
[labial] [labial]
[coronal] [coronal]
[dorsal] [dorsal]
(Clements and Hume 1995: 276)
To my knowledge there have been found no clear instances of consonants

triggering harmony in entire words, that is, for example, an instance where
a word-initial consonant determines roundness or palatality in all following
consonants and vowels in the word. From the absence of such a pattern and
from the absence of long-distance spreading of consonantal place features,
Ní Chiosáin and Padgett (1997) conclude that assimilation applies strictly

locally from segment to segment.
In their view, there are different place features for vowels and conson-
ants. Vocalic place features can be linked to consonants as secondary arti-
culations, while consonantal place features would also impose consonant-
hood on a vowel if linked to it (the 'bottleneck effect' in their terminology).
Thus, assimilation of a vowel to consonantal place features would turn the
vowel into a consonant, resulting in a vowel-less segment chain, which
cannot be syllabified. A labial consonant which triggers rounding in a fol-
lowing vowel should have a vocalic place feature labial in addition to its
consonantal place feature labial, then, in order to be able to trigger round-
ing in the following vowel. Technically, also nonlabial consonants should
be able to bear the vocalic labiality feature and trigger rounding in adjacent
vowels, which is unattested so far.
Besides the lack of evidence for harmony over an entire word triggered
by a consonant there is clear evidence of the featural interaction of vowels
and consonants, such as palatalisation for instance. Such interactions could
likewise be analysed as treating the intersegmental features as syllabically
licensed or as effects of a unified feature geometry for vowels and conson-
ants.
I will depart here from these issues and devote the next passages to the
question of in which domain the process of vowel harmony applies.
1.3 The domain of harmony
The domain of harmony is usually conceived of as that of the prosodic or

morphological word or a smaller morphological unit. This distinguishes
vowel harmony from other assimilatory processes like voicing assimilation
or tone sandhi, which both go over word boundaries and apply within
phrases. Likewise, this equates vowel harmony with consonantal place
assimilation, which also does not apply beyond word boundaries cross-
linguistically.
There are a few cases in which vowel harmony takes place in a larger
domain than that of the word. According to Hall et al. (1974: 261), the
harmony process in Somali covers entire clauses. The clauses in (22a,b)
differ in that (22a) has only retracted vowels and (22b) has advanced
vowels only. The alternation is triggered by one of the two last elements,
The domain of harmony 25
s4 meØyei. According to Hall et al., the harmony pattern of the whole clause
is shaped by the last element, the verb in both cases.
(22) Somali clause harmony

a. b(Øra '8s8b baØ loØ b(Øra,
'New gardens were cultivated for them'
b. beØr4 'üsüb b4Ø löØ s4 meØyei

'New gardens were made for them' (Hall et al. 1974: 261)
Cahill and Bodomo (2000) found that particles in Konni agree with their
syntactic host with respect to the feature ATR, and that here the syntactic
configuration is crucial rather than any morphological or prosodic category.
In Finnish, harmony is restricted to the same domain as stress assign-
ment and syllabification. Main stress is assigned to the first syllable of a
word. Since Finnish is a suffixing language, the stress-bearing unit is al-
ways the first syllable of the root. Proclitics, which do not influence the
placement of stress, only participate in vowel harmony in fast speech
(Skousen 1975). From the inactive behaviour of clitics with regard to both
processes, it can be concluded that the domain of vowel harmony coincides
with that of the prosodic word in Finnish.
In Turkish, the issue is a little more complicated. Harmony affects all
affixes, but also postclitics, which are clearly outside the domain of stress
assignment (Kabak and Vogel 2000, in press). Kabak and Vogel (2000)
base their claim that Turkish harmony applies rather within the domain of
the Clitic Group (Nespor and Vogel 1986) than within the prosodic word
on the participation of clitics in harmony. Some affixes in Turkish block the
harmony process and start their own harmonic domain. These affixes be-
have irregularly with regard to stress assignment as well. On the basis of
this observation Kabak and Vogel (in press) argue that vowel harmony in
Turkish simply goes from an underlyingly specified vowel up to the next
specified vowel. It remains an open question in this analysis why dishar-
monic affixes are also irregular with respect to stress assignment.
Problematic for a clearer definition of the domain of vowel harmony is
also the fact that in most languages compounds consist of as many har-
monic domains as compound members. If each of these stems also consti-
tutes a single domain of syllabification and stress assignment of its own one
may again posit the prosodic word as the domain of vowel harmony. In
derivational accounts, the autonomy of each compound member is seen as
evidence for the claim that vowel harmony is a lexical process while
compounding applies post-lexically.
Van der Hulst and van de Weijer (1995) argue that because disharmonic
affixes often do not constitute a prosodic unit of their own with regard to
syllabification, and because of the fact that in many languages we find
many roots which have disharmonic vowels, vowel harmony refers to some
morphological unit and happens somewhere within the morphological deri-
vation. If vowel harmony is indeed linked to the morphology, this may be
an indication on the function of vowel harmony. One possible function
could be to mark certain morphological boundaries to deliver a cue for the
listener in the interpretation of speech. Vroomen, Tuomainen and de Gelder
(1998) report that Finnish test persons find it easier to detect word boun-
daries if there is a mismatch in vowel harmony preceding them than with-
out such a mismatch. From their comparison of word detection tasks per-
formed by Finnish, Dutch and French test persons, they conclude that stress
as well as vowel(dis)harmony are language-specific cues to the retrieval of
word boundaries.
The other often claimed motivation behind vowel harmony is simply
ease of articulation. This would explain why in Finnish fast speech for ex-
ample the harmony domain is extended to the proclitics.
Now that I have discussed the basic issues of which features can be ac-
tive in harmony, whether the phenomenon is restricted to vowels or not and
in which domain it applies, it is time to move to those aspects of vowel har-
mony which will be central for this book.
1.4 Opacity and transparency
Under certain circumstances a word may consist of two harmonic stretches

with different specifications of the harmonic feature, or a sequence of har-
monic vowels may be interrupted by one or more vowels. Such patterns are
usually triggered by vowels which are referred to in the literature as opaque
vowels and transparent / neutral vowels, respectively.
There is, however a third type of vowel which behaves as phonologic-
ally opaque in that the neighbouring vowel always has to have the feature
specification antagonistic to that of this triggering vowel. I will refer to
such vowels as Trojan vowels and discuss them in section 1.4.3. In sub-
section 1.4.3 also those vowels are addressed which allow for free variation
in the neighbouring vowel. For ease of reference I will call them 'hybrids'.
Opacity and transparency 27
This section starts with an introduction to opaque vowels (1.4.1), goes on

with transparent vowels (1.4.2), and then turns to Trojan and hybrid vowels
in the remaining subsection.
1.4.1 Opaque vowels
Opaque vowels resist assimilation to the feature specification of the adjac-

ent potentially triggering vowel. Instead, opaque vowels start a new har-
monic domain with their own feature specification. This is why they are
called opaque: the harmony process stops in front of these vowels and can-
not permeate through them. As an example I will take the low vowel D in
Tangale, a language with ATR harmony (van der Hulst and van de Weijer
1995: 496f.).
(23) Tangale harmony (a) and the Tangale opaque vowel (b,c,d)
a. 18ld(d( 'dog' b. n kas-ko 'I have cut'
seb-u 'look' (imp.) a-no 'my belly'
k(n-8 'enter' (imp.) war-8 'go' (imp.)
tug-o 'pounding'
w8d-o 'farming'
c. top-a 'start' (nom.) d. ped-na-n-go 'untied me'

top-a 'answer' (nom.) peer-na-n-go 'compelled me'
peer-na 'compelled' Êob-na-g-g8 called you' (pl.)
p(d-na 'untied' Êib-na-m-g8 'cooked for us'
la-pido 'tree' (dim.) kulag-do 'her frying pan'
la-18ld(d( 'dog' (dim.)
(van der Hulst and van de Weijer 1995: 497)
In (23a), we see that affixes, such as –u/8 and –o/o, alternate in their ATR
specification in accordance with the adjacent root vowel. (23b) illustrates
the triggering capacity of the low vowel D in case it is a root vowel and is
followed by a potential undergoer of harmony. (23c) shows that root
vowels do not assimilate to the low vowel, regardless of whether the low
vowel precedes the potential target vowel in the root or follows it. Finally
(23d) shows that when the low vowel is preceded by a disharmonic root
vowel, the vowel following the low vowel agrees with the low vowel,
rather than with the root vowel, with regard to ATR. The low vowel is
opaque in the sense that the ATR spreading action or harmony rule cannot
pass through it. Instead, the opaque vowel initiates a new harmonic domain.
This is rather trivial and gives evidence for the theoretical claim that
harmony applies somehow locally in that it affects adjacent vowels only.
However, this fact can also be used as evidence for serial derivation of
output forms or the directionality of vowel harmony.
The issue becomes more interesting when another kind of harmony,
namely affix-controlled harmony (to be introduced in section 1.5) is con-
sidered in chapter 3.6. The behaviour of opaque vowels in such a system
reveals that vowel opacity is anything else but evidence for serial or cyclic
derivation. However, I will postpone this discussion to the indicated chap-
ters and go on with transparent vowels.
1.4.2 Transparent vowels
Transparent vowels, in contrast to opaque vowels, pose a problem for the

assumption that phonological feature assimilation is local per se. A trans-
parent vowel is one that is immune to assimilation as well, but instead of
initiating its own harmonic domain to one side, the vowel lets the harmonic
feature specification 'pass through' from one side and affect the vowel to its
other side.
Regard in this respect the examples from Wolof. Wolof has ATR har-
mony like Tangale, which is illustrated by (24a). High vowels trigger ad-
vancement of the tongue root in affix vowels (24b). If a high vowel inter-
venes between two nonhigh vowels, the rightward nonhigh vowel is not af-
fected by its direct neighbour but surfaces with the ATR specification of
the far away nonhigh vowel (24c).
(24) Wolof harmony and transparency

a. r(Ør-oØn 'had dinner'
reØr-oØn 'was lost' (Archangeli and Pulleyblank 1994: 227)
b. gis-e 'to see in'

suØl-e 'to bury with'
nir-oØ 'to look alike'
jiØt-le 'to help with' (Pulleyblank 1996: 320f.)
c. t(Ør-uw-oØn 'welcomed'
t(k-ki-l(Øn 'untie!' (Archangeli and Pulleyblank 1994: 231)
The high vowels are transparent in that they are skipped by harmony.
Progressive assimilation proceeds on the right side of these vowels notwith-
standing their antagonistic feature specification. This phenomenon has led
phonologists to a variety of assumptions about the nature of phonological
features and their interaction in the generation of speech. In his summary of
the main types of analysis applied to transparency, Bakoviü (2000) lists
three basic approaches.
One group of authors simply abandons the assumption of locality and
assumes that the harmony process literally skips the transparent vowels
(i.e., Anderson 1980, Archangeli and Pulleyblank 1987, Booij 1984, Cole
and Kisseberth 1994, Ka 1988, Kiparsky 1981, Ringen 1975, Smolensky
1993, Spencer 1986, Steriade 1987, Vago 1988). These types of analysis do
not acknowledge, however, that most instances of phonological feature
assimilation are indeed strictly local, like consonantal place assimilation,
and that even vowel harmony respects certain restrictions on the distance in
which the process applies.
In the account of Calabrese (1995), and Halle, Vaux and Wolfe (2000)
for instance, it is assumed that harmony is not spreading of the harmonic
feature from segment to segment, feature node to feature node, or feature
bearer to feature bearer, but rather applies only to those segments which are
contrastively specified for the harmonic feature.
(25) Locality of harmony with a contrastive feature

X Y (Halle, Vaux and Wolfe 2000: 399)8
[back] (or [ATR])

where X = a segment contrastively specified for [back] (or [ATR])
Y = a segment that can bear a [back] (or [ATR])
specification
Any transparent vowel Z could be placed between segments X and Y in the

diagram (25) without blocking the spreading of [back] or [ATR], because
of Calabrese's (1995) and Halle, Vaux and Wolfe's (2000) assumption that
the transparent vowels do not bear the harmonic feature. They do not con-
trast with respect to backness (in the case of Uyghur, which is discussed in
Halle, Vaux and Wolfe). Applied to our example of Wolof, given in (24),
this means that the two high vowels are ignored by the rule spreading
[±ATR] because they do not bear this feature contrastively.
The difference between languages with transparent vowels and lan-
guages with opaque vowels then lies in the exact formulation of the spread-
ing rule. We will see below in chapter 6 that Yoruba high vowels are
opaque to ATR harmony and that the Hungarian high front vowel is opaque
to backness harmony. The difference between these two languages and
Wolof and Finnish (which displays backness harmony with transparent i
and e, see chapter 5), respectively, then lies in the different definition of the
trigger and target in the spreading rules of each language. The spreading
rule operative in Wolof and Finnish applies to contrastively specified
vowels only, while the respective spreading rule in Yoruba and Hungarian
has to apply to all segments specified for the harmonic feature.
Though this account saves at least a relativised notion of locality, it
raises the question why languages should make such a subtle difference in
their rules at all. A further question arising in the context of this analysis is
why the low vowel which served as an example of an opaque vowel in (23)
does not behave as transparent in many languages (see also the behaviour
of the low vowel in Akan; Clements 1976).9 Either such a low vowel is re-
paired with another nonhigh vowel, which does not exactly match the fea-
ture profile of the low vowel or it behaves as opaque to the harmony pro-
cess. The last objection is tied to the role of the transparent vowel in the
harmony process. In such an analysis, the vowel is not only transparent, it
is also inactive as a trigger (i.e., it is supposed to be completely neutral).
Thus, we would expect that alternating affix vowels would surface with
their underlying feature specification when preceded by a transparent/
neutral vowel only. As will be shown in chapters 5 and 6, such 'transparent/
neutral' vowels impose severe restrictions on their environment, which runs
counter to the prediction made by such accounts.
A similar idea underlies accounts in which vowel transparency is seen
entirely as an effect of different representations for the different vowels. A
theory of this kind is advocated by van der Hulst and Smith (1986) for in-
stance. Since they attempt to explain the behaviour of Trojan vowels as
well with their theory I popstpone discussion of this account until after the
introduction of these vowels.
In the second type of approach feature copying is assumed instead of a
spreading device (Archangeli and Pulleyblank 1994, Pulleyblank 1996).
The copying mechanism assumed by Pulleyblank (1996) crucially relies on
featural alignment constraints. I will argue later that the device of featural
alignment is not appropriate to handle phonological feature assimilation,

since it is not capable to account for both, vocalic as well as consonantal
assimilation.
The third strand explores the possibility that transparency involves deri-
vational opacity in one or the other sense. In derivationalist transforma-
tional accounts the harmony rule changes the feature specification of the
transparent vowel by agreement with the neighbouring vowel to one side.
At this stage the transparent vowel has a form which is unattested in surface
forms. The transparent vowel itself then triggers assimilation of the vowel
to its other side. In a later step of derivation, the harmonic feature of the
transparent vowel is set back to its initial state by further rule application.
This is the classical case of the Duke-of-York-Gambit (Pullum 1976) or
phonological opacity (Kiparsky 1971, 1973).
(26) Wolof transparency as derivational opacity:

Input /t((r-uw-oon/
Vowel harmony 1. /u/ → 8/V[-ATR]_, 2. /oo/ → oo/V[-ATR]_
t((r-8w-oon
Repair 8→u
Output [t((r-uw-oon]
Instances of phonological opacity have always been regarded as the most

important argument in favour of serialist derivations of linguistic structures.
One approach to model phonological opacity within Optimality Theory
without reference to different steps of derivation is Sympathy Theory
(McCarthy 1999), which will be discussed in more detail below in chapters
5 and 6. However, as Bakoviü (2000) points out, the failure of such an
account is that it still crucially refers to more representations than just the
input and the output. Even though Sympathy Theory does away with the
assumption of intermediate representations between input and output, it
relies on faithfulness relations between the output and specific failed can-
didates. Furthermore, as will become clear in section 5.3, in the theoretical
boundaries set by McCarthy, the type of opacity displayed by transparency
is technically intractable in Sympathy Theory.
Bakoviü (2000) and Bakoviü and Wilson (2000) model transparency as
an effect of constraint targeting and cumulative candidate evaluation.
Under this account candidates are not compared as a whole set, as usual,
but pairwise constraint by constraint. A certain targeted constraint imposes
an ordering on selected pairs of candidates, which is compared to the

pairwise candidate orderings of other lower ranked constraints.
The more plausible nontechnical intuition behind this is that the gram-
mar, no matter on which dubious paths, selects the candidate as optimal
which resembles most the completely harmonic form. To put it simple, in a
case like Wolof, the phonologically opaque form with a transparent high
vowel t(ØruwoØn is chosen as optimal because it looks more like the com-
pletely harmonic form *t(Ør8woØn than the phonologically transparent form
with an opaque high vowel *t(ØruwoØn.10 Seen from this perspective con-
straint targeting is the same as stepwise candidate evaluation in the sense of
Kiparsky's (1999) LPM-OT, where three chronologically ordered stages of
evaluation are assumed and candidates are always in direct correspondence
with the output of the preceding stage that serves as the input for the stage
actually at work. More detailed discussions of the Sympathy approach as
well as the targeted constraints analysis will follow the analyses of trans-
parent vowels in chapter 4.
In the current work I will go into an entirely different direction in the
analysis of transparent vowels. The basic idea is that these vowels are not
deficient or excluded from vowel harmony, but rather that by their
asymmetry with regard to the vowel system, they receive a special status.
This special status enables these vowels to impose a rigid requirement on
their environment. These vowels either agree with both surrounding vowels
or they disagree with both. An imbalanced state of affairs where the 'trans-
parent', or better balanced, vowel agrees with one neighbour and disagrees
with the other with respect to the harmonic feature is excluded. This will be
formalised quite straightforwardly as an instance of local constraint con-
junction in chapter 3 and will later be applied to Finnish and Wolof in
chapter 5.
One may categorise languages now into those having opaque vowels,
those with transparent vowels, and those which have both. Turkish and
Tangale for instance have opaque vowels, while Finnish and Wolof have
transparent vowels apart from fully harmonic vowels. The latter, Wolof,
additionally has an opaque vowel. The long low vowel is immune to har-
mony and triggers alternation to RTR in the following vowel (see 27,
especially c). Wolof, thus, has both transparent vowels and an opaque
vowel.11
(27) The Wolof low long vowel

a. xaØr-( 'to wait in' c. doØr-aØt-( 'to hit usually'
jaØy-l( 'to help sell' genn-aØl( 'to go out also'
b. yab-aØt-( 'to lack respect for'

woØw-aØl( 'to call also' (Pulleyblank 1996: 316)
Apart from opaque and transparent vowels there is, however, a third and
fourth type of vowel which display a remarkable behaviour with regard to
their neighbours, viz. the one group systematically disagrees with its neigh-
bour, while the other causes free alternation in adjacent potentially har-
monic vowels. These vowels may be best illustrated with an example from
Hungarian.
1.4.3 Trojans and Hybrids
Hungarian displays backness harmony, but lacks the back alternant to the
two front nonlow vowels i and e. Usually, vowels following one or more of
these two vowels in a word have the backness specification of the vowel
preceding the neutral or transparent one.
In some mono-syllabic roots which contain one of these two vowels, the
vowel triggers the front alternant of potentially harmonic suffix vowels to
surface (28a), which is just what one would expect since the triggering
vowel is a front vowel as well. In other mono-syllabic roots containing one
of the two vowels under discussion, the suffix vowels occur as back (28b),
which is quite astonishing at first sight, since this creates a harmony mis-
match. In a third group, the same suffix vowels occur sometimes as back
and sometimes as front (28c).
(28) Hungarian strange vowels

root gloss adess.
a. kéz 'hand' kéznél
film 'film' filmnél
ablative
b. híd 'bridge' hídtól
cél 'aim; target' céltól
delative adessive
c. pozitív 'positive' pozitívról / pozitívnál /
pozitívrĘl pozitívnél
balëk 'fool, greenhorn' balëkról / balëknál /
balëkrĘl balëknél
(Olsson 1992: 79)
The first group of vowels is rather unspectacular, while the other two are
more interesting. In a serialist approach, the 'neutral' vowels in (28b) can be
analysed as underlyingly back and harmony is assumed to occur at a stage
where the vowel still has its underlying feature specification. After the
application of harmony, the triggering vowel is changed from [+back] to
[-back] by another rule. In an alternative non-serialist view, Ringen and
Vago (1998) posit a floating feature [+back] as belonging to the respective
root.
In reminiscence of the former view I will refer to these vowels as
'Trojan vowels', since superficially they look like peaceful front vowels,
which is not what they really are, and they 'invade' their neighbour by im-
posing their underlying feature specification on this target vowel. Under
this view they are of the same type as a particular class of vowels in Yawel-
mani, which apparently contradict the generalisation that harmony applies
only to height-uniform vowels. These vowels are long low vowels which
trigger harmony in adjacent high vowels but not in low ones. The pattern is
exemplified and analysed in more detail in chapter 6, section 4.
The group of Hungarian vowels in (28c) behaves as hybrid, and these
stems are labelled as 'vacillating stems' in the literature (see Olsson 1992).
Such data suggest two treatments, either they might be underspecified or
we have to store two separate items for each stem, one with a back and one
with a front vowel underlyingly. Under the first analysis one might wonder
why the backness specification of the preceding non-neutral vowels does
not simply permeate through the underspecified item. Therefore I opt for
the second choice, i.e., that two variants of the respective stems are stored
in the lexicon.
Trojan vowels occur in systems with opaque vowels (Yoruba), as well
as in systems with transparent vowels. I should add here that the existence
of transparent vowels in Hungarian becomes questionable on the grounds
of the analysis proposed in chapter 6. This, however, is rather a side effect.
The basic analysis of Trojan vowels will follow the path provided by that of
balanced (i.e., transparent) vowels in that the behaviour of Trojan vowels
Dominance, morphological control, and Umlaut 35
will be shown an effect of local constraint conjunction. This allows us to

dispense with any variety of serialist account of this phenomenon, as well
as to abandon assumptions like floating features or similar devices.
1.5 Dominance, morphological control, and Umlaut
Harmony systems can also be categorised according to the morphophono-

logical characteristics of the triggering and the target vowels.
Along these lines harmony systems are traditionally divided in root-
controlled versus dominant-recessive patterns. In this section I would like
to add the pattern of affix-controlled harmony to this typology and finally
draw attention to stress-driven harmony as well.
Most examples which have been given in this introduction so far have
been of the root controlled type. In this type of harmony, the specification
of the harmonic feature is induced on the whole word by one of the root or
stem vowels, if there is no opaque vowel. Such opaque vowels shape only
those vowels which are more marginal in the word than the opaque vowel
itself. In these systems it never happens that an affix vowel determines the
quality of a root or stem vowel, and any vowel situated between a root
vowel and an opaque affix vowel with a conflicting feature specification
always gets the specification of the root vowel.
In the affix controlled type of harmony, which is also sensitive to the
categorical status of the morpheme the triggering vowel belongs to, it is the
affix vowels which systematically trigger harmony. In affix-controlled
harmony, the vowel between triggering affix vowel and opaque vowel
always harmonises with the affix vowel. Anderson (1980) states that "there
are apparently no systems in which suffixes exclusively control harmony".
McCarthy and Prince (1995) claim that affix-controlled harmony is impos-
sible as an effect of the universal meta-ranking of root faithfulness over af-
fix faithfulness. Bakoviü (2000) excludes this pattern in his OT account of
vowel harmony. However, Noske (2001) argues that Turkana is just of that
type. Here, Fula (as reported in Paradis 1992 and Breedveld 1995) will
serve to illustrate this kind of harmony, because it is a much more straight-
forward case. Before coming to this pattern I will discuss markedness
driven dominance.
In dominant-recessive systems, in contrast to morphologically con-
trolled ones, the categorial status of the morpheme bearing the triggering
vowel plays no role at all. In such systems, one feature specification is
dominant, and if it occurs in one morpheme the whole word has to look the
same. The language is not sensitive to whether the morpheme containing
the dominant feature specification is a root or an affix.
In Kalenjin, for instance, the dominant feature is [+ATR]. If all mor-
phemes contain only retracted vowels, these surface as such (29a), but if
only one morpheme, regardless whether it is an affix or a root, contains an
advanced vowel, this vowel triggers tongue root advancement in all other
vowels (29b,c). In (29b) the root kHØr contributes the advanced tongue root
position, which alters the ATR specification in the other vowels. In (29c) it
is the suffix -e which causes the alternation from retracted to advanced
tongue root in the other vowels, including the stem vowel as well.
(29) Kalenjin ATR harmony

a. K,- $- k(r → [k,$g(r]
DIST.PAST 1SG shut 'I shut it'
b. K,- $- kHØr -m → [kingHØrin]

DIST.PAST 1SG See 2SG 'I see you(sg.)'
c. K,- $- k(r -e → [kingere]

DIST.PAST 1SG Shut NON.COMPL. 'I was shutting it'
(Bakoviü 2000: 52)
In this context, also the question arises which feature (specification) it is

that spreads. Is ATR or RTR the relevant feature in Kalenjin? From the
point of view that harmony is a kind of neutralisation and neutralisation
goes always from the marked to the unmarked value of a feature, one has to
conclude that it is the feature Retracted Tongue Root which is active in
Kalenjin, rather than Advanced Tongue Root. Under this view [+RTR]
vowels are marked and remain as such if no unmarked feature value is pre-
sent in a word. If one vowel lexically contributes [-RTR], the unmarked
feature value, this is extended onto all other vowels reducing the overall
markedness of the word in question.
A further issue that arises here with regard to the nature of phonological
features is whether some or all features are privative or binary. Privative
features do not have a two valued possibility of specification as either
minus- or plus-valued. They are either absent or salient. Under a privative
view (on ATR/RTR as privative features see the work of Pulleyblank for
instance), RTR cannot be the active feature in Kalenjin, because any RTR
which is salient on a feature bearer would be extended via spreading to all

other vowels. If RTR is absent (i.e., on advanced vowels) there is nothing
that can spread or cause the neighbouring vowels to look the same. For
instance in the privative system proposed by Kaye, Lowenstamm and Ver-
gnaud (1985) and van der Hulst and Smith (1986) the radical for the tongue
root dimension is A, representing advancement of the tongue. Van der
Hulst and Smith's analysis of dominant-recessive ATR harmony holds that
dominant vowels contribute the A feature which spreads to all the other
featureless vowels. If this feature theory were correct we would not expect
the reverse picture in any language. Dominance of RTR should not occur.
There are however languages, such as Nez Perce (Aoki 1966, 1970, Bako-
viü 2000) which seem to be of that type. In Nez Perce a single vowel with a
lexically retracted tongue root causes retraction in all other vowels in the
word. For Nez Perce one would have to introduce a parameter which
chooses either A or R as the ATR feature for each language.
There are numerous arguments for and against either position (i.e., ATR
as active, RTR as active, ATR as privative, ATR as binary, RTR as priva-
tive, RTR as binary feature) which I will not reproduce here.
In the literature there have also been numerous discussions on the priva-
tivity/binarity of other features. In particular the feature roundness has been
assumed to be privative by many linguists (Archangeli and Pulleyblank
1994, Steriade 1987, and others). Clements and Hume (1995) assume that
all features referring to specific articulators are unary (i.e., privative) while
the rest are binary (such as [±consonantal]).
The question of whether features are binary or privative is closely re-
lated to the question of underspecification. Archangeli (1988) discusses
three reasons for underspecification, which all result in different patterns of
(under)specification in underlying as well as surface forms. Some feature
bearers might be inherently underspecified for certain features due to their
nature, for instance sonorants might be underspecified for the feature
[voice], because they are almost always voiced. The voice feature is then
inserted via a redundancy rule or a co-occurrence constraint. On the other
hand, voiceless obstruents might be seen as underspecified for voicing be-
cause being voiceless seems to be the natural state for an obstruent. A se-
cond view may favour specification of contrastive features only which re-
sults in underspecification of all other features. A third option, radical un-
derspecification, is specification of all and only unpredictable features. Fea-
tures that are predictable by the nature of a feature bearer, by the occurr-
ence of other unpredictable features via co-occurrence restrictions, or by a

phonological rule or process are said to be underspecified now.
The latter aspect, namely that alternating features are underspecified, is
picked up in Inkelas' (1994) theory of Archephonemic Underspecification
and Lexicon Optimization in Optimality Theory. Optimality Theory de-
mands that even predictable features are specified underlyingly, because
every insertion of a feature incurs violations of anti-insertion constraints.
In this theory, the input/underlying form is chosen for an output/surface
form which produces the least constraint violations in the input-output
mapping. The underlyingly underspecified form violates anti-insertion con-
straints, which the underlyingly fully specified form satisfies. From this
logic it results also that alternating features are underspecified, because this
saves the candidates from violations of faithfulness constraints. Root-con-
trolled vowel harmony is then due to the circumstance that all or almost all
affixes in a language are underspecified for a given feature. Roots never
alternate in root-controlled harmony, and, hence, are fully specified for the
harmonic feature underlyingly. The latter point, full specification of all root
vowels in languages displaying root-controlled vowel harmony, has been
questioned by Harrison and Kaun (2000) on the basis of data from word
games. However, assuming that root control arises out of the underspecific-
ation of affixes with regard to the harmonic feature would be circular. The
question left open would be why affixes are prone to underspecification
while roots are reluctantly underspecified in the world's languages. I will
come back to this issue later in chapter 3 and advance to the third poss-
ibility besides root control and dominance.
This third option apart from root control and dominance is affix control.
Van der Hulst and van de Weijer mention Turkana as a language which
might be of that type. Also Noske (2001) argues for this view. However,
Bakoviü (2000) analyses Turkana as of the dominant-recessive type.
Paradis (1992) examines a dialect of Fula, which, in my view, gives clearer
evidence for affix controlled harmony. In Futankoore Pulaar (a dialect of
Fula), affixation goes rightward and ATR harmony goes strictly leftward.
This is illustrated in (30). The root vowel always agrees with the following
affix vowel with regard to ATR. This can be be inferred from the compari-
son of the same roots in combination with different affixes in (30a) and
(30b).
(30) Pulaar mid root vowels and harmony

a. ATR words Gloss b. RTR words
sof-ru 'chick' cof-on
ser-du 'rifle butt' s(r-on
peec-i 'slits' p((c-on
dog-oo-ru 'runner' dog-o-w-on
lef-el 'ribbon' l(f-on
keer-el 'boundary' k((r-on (Paradis 1992: 87ff.)
Even though affix control is extremely rare, the pattern is quite robust in
Fula. Breedveld (1995) reports the same harmony pattern also for other
Fula dialects.12 The Fulfulde pattern poses severe problems for Bakoviü's
(2000) account of vowel harmony, where root control is modelled as an
instance of base-output correspondence, since this approach systematically
excludes affix control. This issue will be subject to a more thorough dis-
cussion in chapter 4.
A more widespread pattern where root vowels depend on the quality of
affix vowels is Umlaut, which can be found in Veneto Italian for instance
(Walker 2001). However, in such languages the feature specification of the
affix has an influence only on the neighbouring root vowel, as can be seen
from the data in (31). Mid root vowels are raised before a high affix vowel
in the Veneto dialect of Italian. The harmony process affects only the
stressed vowel preceding the affix. From this fact it must be concluded that
this is an instance of parasitic licensing rather than vowel harmony.13
(31) Affix-induced feature change in Veneto Italian

a. védo te vídi 'I see/you see'
kréo kríi 'believe 1sg./2sg.pres.ind.'
córo te cúri 'I run/you run'
tóso te túsi 'boy sg./pl.'
b. tornévo tornívi 'return 1sg./2sg.imp.ind.'

benedéto benedíti 'blessed m.sg./pl.'
moróso morúsi 'lover m.sg./pl.'
(Walker 2001: 2)
In her analysis, Walker assumes that the height of the affix vowel is marked
in unstressed affix position. Thus, it occupies also the preceding root
vowel, which is in a strong, i.e., stressed position, which provides a licenser
for the feature [+high]. The feature [+high] is allowed in this position, but
not in others. If the feature is linked to the stressed syllable, it can also be
extended on the adjacent unstressed syllable, where it belongs to
underlyingly. Of course one could argue that vowel harmony is nothing
else than a case of parasitic licensing. The feature specification in all as-
similated vowels is licensed by the triggering vowel which is in a strong
position (such as the first syllable, for instance). This form of licensing is
then less marked than the licensing of each feature realisation by each
associated segment or mora or syllable. This is the view advocated for by
Beckman (1997). If this were the motivation behind harmony, we would
expect the phenomenon attested in Veneto Italian or Icelandic (where
roughly the same happens, see Grijzenhout 1990 and references cited there)
to literally go much further, that is to affect more vowels of the root than
just one or two.
For this reason, I will regard such cases not as instances of vowel
harmony, but as metaphony or umlaut, which should be treated differently
from harmony, what is obviously done in the analysis in Walker (2001).
A second argument for treating vowel harmony and affix-induced
Umlaut as different phenomena comes from the fact that Umlaut often
causes the realisation of a feature specification in the target which is not
present in the trigger. This is illustrated for the case of Icelandic here. In
Icelandic, the dative plural affix -um causes raising, rounding, and fronting
of the last low vowel in the root.14
(32) Icelandic Umlaut

pakki + um pökkum 'parcel-dat.pl'
tala + um tölum 'to speak-3.pl'
almanak + um almanökum 'almanac-dat.pl'
apparat + um apparötum 'apparatus-dat.pl'
(Grijzenhout 1990: 57f.)
However, being front is neither a property of the vowel D nor of the vowel
u. To find out about the place feature of the low vowel D in Icelandic one
may have a look at Icelandic diphthongs.
According to Küspert (1988), Icelandic has five diphthongs ei, öy, ou,
Di, and Du which all have a short/long distinction additionally. Braunmüller
(1991) also regards je as a diphthong. Except for the diphthongs containing
the low vowel the two parts of each diphthong have to share or agree in
their backness specification. If D were a front vowel, the diphthong Du

should be excluded from the Icelandic inventory.
However, i-umlaut gives evidence for D being definitely not a front
vowel phonologically, since i-umlaut consists of fronting (see Anderson
1969:67), and tense D is fronted to e, lax D becomes 4. Assuming that D is
underspecified for backness or [+back] underlyingly, its raised allophone
should be o according to general assumptions of markedness, not ö. Thus,
the feature [-back] or [+front] must be contributed by the u-umlaut
operation, even though the trigger has the opposite feature specification.
This is untypical of vowel harmony, where features which are present in the
trigger extend to the targets. Vowel harmony does not add a feature
specification it extends or copies a present feature specification, while
Umlaut can add a feature.15
Having this argument set, I would like to discuss briefly one last type of
harmony, stress dependent harmony. In Pasiego Montañes (McCarthy,
1984), height harmony is consistently triggered by the stressed vowel in the
word. All vowels within a word must be either high or mid. The low vowel
is neutral. The data in (33) illustrate this. I have added Castilian Spanish
forms to McCarthy's data to allow for a better comparison of the
distributional data in (33a). The pattern is also visible in the morphology as
shown in (33b).
(33) Pasiego Montañes stress-dependent height harmony

Pasiego gloss Castilian Spanish
a. bindi7ír 'to bless' bendecír
k8nt,ºnt8 'happy(count)' conténto
b. bebér 'to drink'

bibíØs 'drink-2pl.pr.ind.)' (McCarthy 1984: 295f.)
Pasiego Montañes displays a tense/lax or ATR harmony as well. McCarthy

observes that height harmony is exceptionless within words containing lax
vowels, but that there is a considerable number of exceptions to height
harmony in words containing tense vowels. I will neither go deeper into the
intricacies of Pasiego harmony nor discuss McCarthy's analysis here since
the case served only to exemplify one additional pattern of harmony for the
typology.
With this last pattern the following typology of vowel harmony arises.
(34) The basic types of vowel harmony

a. Morphologically driven
& root control
& affix control
b. Phonologically driven
& markedness controlled: dominance
& stress controlled
Classic cyclic rule application can easily explain morphologically con-

trolled harmony systems in which root vowels are triggers, but as we have
seen above, there are (even though rare) cases of affix control.
In a cyclic derivation, a base form (i.e., the root or stem) is derived and
all phonological rules apply including the harmony rule and feature
insertion rules (redundancy rules). After this process, the first affix is
attached, brackets are erased and the affix is subsequently subject to vowel
harmony. With the next affix the procedure is repeated, and so on until a
whole word form is generated. Since all features are inserted already in the
vowels of the stem at the time when the first affix is attached and the
harmony rule applies again, the (even inserted) features of the stem shape
the feature profile of the affix vowel(s).
This kind of derivation is mimicked in Bakoviü's (2000) Stem-Affixed-
Form-Faithfulness approach (henceforth SAF-approach), where a complex
form has to be faithful to all its simpler bases, which are the output minus
one affix, the latter form minus one affix and so forth. Dependent on how
many affixes a form has, it has one base less than affixes. I will discuss this
approach in more detail in chapter 4.6, following my analysis of morpho-
logical control.
In a derivational approach, suffix controlled harmony could be modelled
by the assumption of a post-lexical harmony rule operating from right to
left. Generally, it must be said that to date there exists no satisfactory
account of the typological observation in (34) in parallelist Optimality
Theory, since there is one group of analyses which must be rejected on
independent theoretical grounds (the alignment approach) and one (the
SAF-approach) which actually reintroduces cyclic derivation in OT. The
former will be discussed in chapter 2.5, where I will discuss the issue
whether vowel harmony is best treated as an effect of the interaction of
positional faithfulness with markedness (as proposed by Beckman 1997),
an instance of featural alignment (as proposed by Smolensky 1993,
Setting the scene 43
Kirchner 1993, and many others), or the satisfaction of intra-representa-

tional surface correspondence constraints (as proposed by Krämer 1998,
and further developed by Bakoviü 2000).
1.6 Setting the scene
We have seen in the preceding sections that languages which display vowel
harmony are characterised by the feature or feature combination which is
active, by the occurrence of opaque, transparent or otherwise quirky vowels
or a mixture of these types, by whether the harmony is morphologically
controlled or dominant, and by the domain in which harmony applies. All
of these characterising aspects of harmony raise interesting questions and
posit quirky puzzles. Some of these patterns raise questions which are of
particular interest for the generativist theoretical linguist, and theoretical
questions snowball if one looks at some aspect of vowel harmony. Besides
phonetic, psycholinguistic or socio-linguistic aspects of harmony (see for
instance Boyce 1990, Campbell 1980, Vroomen, Tuomainen and de Gelder
1998), research in vowel harmony can be divided in two larger fields, one
of which is concerned with the nature of the features as such.
In this direction one has to examine which features harmonise, which
group together and which do not, in how far do vocalic features interact
with consonants and vice versa, whether features are privative, binary or
ternary. One deals with the question which instantiation of a feature is the
marked one phonologically and phonetically (do languages explore RTR or
ATR? Does backness spread or frontness?), and which features have to be
assumed at all. For research in this direction see Halle, Vaux and Wolfe
(2000), Calabrese (1995), Clements and Hume (1995), van der Hulst and
Smith (1986), Padgett (1995), Odden (1991) and the references cited in
these works.
The other direction in which research can lead regards the interaction of
phonology and morphology. Here it becomes relevant which factors other
than featural shape the vowel harmony patterns in the languages of the
world, and whether there are connections between the grammar and the
systemic properties of vowels and their interaction. Do underlying
representations and structures shape individual harmony patterns or do
languages diverge in the grammars which have an influence on the output
forms; and if the differences are to be found in the grammars, of which
nature are these differences? Can the cross-linguistic patterns be described
by different interactions of the same set of universal constraints on output

forms? Do we have to assume derivations of differing complexity, i.e., with
different numbers of levels for different languages?
The discussion within this book aims to contribute to the latter
perspective. The issues which will be under special consideration here are
the theoretical treatment of transparency/opacity because of the tie to the
locality issue which is central to phonological theory, the modelling of
morphological control and dominance within Optimality Theory because of
the potential insights to the question whether phonological processes are
directional or not, and if not, what determines directional tendencies, and
the relation to the ever prevailing issue whether language production is a
derivational, serialist process or a parallel action, and last but not least the
interaction of vowel harmony with other phonological requirements, since
it promises answers to the latter issue of serialism versus parallelism as
well.
However, the distinction between the two latter directions of research
could be artificial. For instance the assumption of derivational opacity in
the analysis of transparent or Trojan vowels may be obviated by insights
into the nature of vocalic features. If vocalic features are privative and
neutral vowels are assumed to be characterised by the absence of the active
feature the spreading across the neutral vowel can be explained without
reference to an intermediate step of derivation. Likewise could vowel
opacity also be just a side effect of the representation of the opaque vowel.
Van der Hulst and Smith (1986) for instance propose an account of neutral
vowels which links insights from autosegmental theory and theories of
privative features and seeks to explain the distributional facts solely by
reference to phonological representations.
As noted above van der Hulst and Smith (1986) assume vocalic features
to be privative in the sense of Kaye, Lowenstamm and Vergnaud (1985).
They link Kaye, Lowenstamm and Vergnaud's abstract vowel elements
with the formalism of autosegmental theory.
The unary feature for the front/back dimension is assumed to be a
radical [i], for frontness. Roundness is represented by abstract [u], and
height is codified by an [a] feature, referring to low. An additional
dimension like tongue root position can further be added by an [A] element,
which stands for advancement of the tongue root. In van der Hulst and
Smith's approach these features can be associated to segments or morph-
emes in a range of ways. Transparent vowels are linked to their segment via
an association line. The 'usual' case, however, is that a feature is in the
scope of a segment but not associated to it or a segment has no feature at

all. In addition to different representations they invoke association con-
ventions to determine to which segments a feature can be linked. That is, if
a segment is accompanied by a floating feature this is associated with as
many segments as possible (35a). If a segment is lexically associated with a
feature no association conventions (AC's) apply and the surrounding
segments receive the default value (35b).
(35)
a. CV CV CV → CV CV CV
AC's
F F
b. CV CV CV → CV CV CV
F F
The structure in (35b) is van der Hulst and Smith's representation for trans-
parent vowels. In case the transparent vowel is preceded by another trans-
parent vowel this analysis generates a disharmonic form at first glance, as
illustrated in the lefthand representation of (36). Van der Hulst and Smith
refer to the Obligatory Contour Principle (OCP, Leben 1973, Goldsmith
1976, McCarthy 1986) to avoid such a configuration. According to the
OCP adjacent autosegments with the same specification are not allowed.
The OCP thus transforms this representation into the one on the right in
(36).
(36)
CV CV CV → CV CV CV
OCP
F F F
In this approach neither locality has to be sacrificed nor have extra deriva-
tional steps to be introduced. The approach has problems of a different na-
ture. First, we encounter technical problems with monosyllabic roots. If a
root has only one vowel and this is a neutral vowel the vowel should not
spread. The feature is lexically associated with the vowel and according to
van der Hulst and Smith these features do not spread. Data from Wolof
above (especially 24b) show that for instance in ATR harmony affix vowels
always surface with the marked value when attached to a monosyllabic root
with a neutral vowel. The only solution in van der Hulst and Smith's model
is to deny the monosyllabic root with a neutral vowel the lexical association
of the feature, as sketched in (37b). Why are neutral vowels in mono-
syllabic roots different from neutral vowels in polysyllabic roots?
(37)
a. CV -CV → CV -CV
F F
b. CV -CV → CV -CV
OCP
F F
This explains the different behaviour of neutral vowels in Hungarian (the

Trojans), but not why languages which don't explore both possibilities usu-
ally opt for choice (b) with mono-syllabic roots but representation (a) in all
other environments. Another possibility which opens up here is that the af-
fix vowels can occur with their own lexical specification in words with
mono-syllabic roots. Affixes without a feature just get the default value,
since the root vowel does not spread, while affix vowels with the lexical
feature will agree with the root vowel. This usually doesn't happen and
accordingly all affix vowels have to be regarded as underspecified. This
underspecification does not derive independently from the theory and has
to be stipulated. In this book I will explore a principled solution to the ob-
servation that in root control the affix vowels are systematically targeted
and do not surface with their lexical specification even in the environment
of a neutral vowel.
Furthermore, with respect to van der Hulst and Smith's approach the
question arises why vowels which have a harmonising counterpart do not
behave as transparent. In an ATR harmony system we could imagine a mid
vowel with the A feature lexically associated to it. If this is preceded by a
vowel without a lexical feature specification and followed by an affix
vowel (which is underspecified by convention) we should observe trans-
parent behaviour. Thus the association of the lexical value has to be re-
stricted to those vowels which lack a counterpart with the default value.
Since this threefold way of combining feature and segment cannot
account for opaque vowels van der Hulst and Smith have to introduce a
fourth possibility. Apart from being in the scope of a segment features can
also be bound to a segment or syllable. This is a different way of linking a
feature to a segment than via association. In this form of lexical
specification a segment projects its boundaries onto the plane on which the
respective autosegmental feature resides. In such a case no feature outside
this binding domain can extend an association into this domain. On the
other hand a feature residing in this domain can establish associations with
segments outside its domain. With this possibility of extending segmental
domains on autosegmental tiers the theory is thus enriched by three new
types of feature bearing units. Altogether we arrive at six different
possibilities for one autosegmental tier already where theories with binary
features usually allow two choices (maximally four if the options for under-
specification and floating features are included). The idea of feature tiers
being bound to segmental domains explains cases of disharmony and can
also be extended to the Trojan vowels. The inconvenient aspect of the
theory is that the unary feature has six different ways of pairing with its
feature bearer and that the lexical association of the feature with the trans-
parent segment has to be stipulated. Furthermore, in this approach it has to
be stipulated for every language whether all affix vowels have to be
underspecified or specified for the harmonic feature.
Another point of criticism arises with regard to the choice of the marked
and unmarked configuration of the privative features. It is debatable
whether frontness is the marked state in the front/back dimension, and in
the discussion of dominance we have seen data challenging the postulation
of a universal A element for advanced tongue root.
It is unlikely that patterns which are apparently shaped by morpho-
logical asymmetries can be explained with reference to phonological repre-
sentations alone. Furthermore, as will be shown in more detail in chapters 3
and 4, languages with suffixation show a greater variety of contrasts in the
first syllable of the root while languages with prefixation show a greater
variety and immunity in the last syllable of the root, and languages with
suffix control show relaxation of general markedness requirements in the
last syllable of the word. To explain these patterns goes beyond a purely
phonological theory. Optimality Theory with its assumption of constraint
interaction offers itself here since optimality-theoretic constraints refer to
all aspects of linguistic analysis rather than being restricted to a particular
component of grammar such as morphology or phonology within one
constraint hierarchy. The discussion in this book will center around the
topics listed in (38). The analyses provided can also be transferred into a
theory of non-equipolent features. However, the choice of feature theory

alone cannot answer the questions to be asked in this book and the related
discussion is therefore avoided where possible.
(38) The central patterns and issues in this book

– Morphological control and dominance
– Opaque vowels
– Transparent vowels
– Trojan vowels
– Phonological opacity
The next two chapters will set the theoretical basis for the case studies in
the second part of the book. I will first give a brief introduction to Optimal-
ity Theory and its particular subtheories of Correspondence, Alignment/An-
choring, and constraint coordination. After this I will give a schematic out-
line of my own proposal to deal with the phenomena under discussion and
the theoretic problems tied to them. In the second part of the book I will
illustrate how the theory works when individual languages are considered.

An Introduction To Vowel Harmony

Uploaded by

Copyright:

Available Formats

An Introduction To Vowel Harmony

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

An Introduction To Vowel Harmony

Uploaded by

Copyright:

Available Formats

Chapter 1

An introduction to vowel harmony

(1) Vowel harmony

[+F] [-F] [+F] [+F] [+F]

The generalisation that adjacent vowels agree with respect to a certain

1.1 Harmonic features

In this section, I will give a brief overview on which phonological features

(2) The IPA vowel chart

There are, however, other possibilities to express the differences and

1.1.1 Single feature harmonies

One type of harmony affects the dimension of backness or palatality. This

(3) Backness (or palatal) harmony in Finnish

b. Back vowel words

Roundness (or labial) harmony is attested in Southern-Payute, Khalkha-

(4) Roundness harmony in Turkish (combined with backness harmony)

Languages displaying solely labial harmony seem to be quite rare or non-

(5) Shona height harmony

(6) Yoruba ATR harmony among mid vowels

Geographically isolated occurrences of tongue root harmony – henceforth

1.1.2 Multiple feature harmonies

(7) Warlpiri harmony

However, if we consider (324c) as well it emerges that the harmonic feature

roundness. The conclusion is that roundness harmony is just a by-product

(8) Turkish backness and roundness harmony

contrast in the dimension of backness. So the active feature might also be

(9) Dagaare multiple feature vowel harmony

bo 'to look for' b8oro 'looking for'

b. mi 'to rain miire 'raining'

s,,r, 'to touch' s,,r( 'touching'

c. kpa 'to boil' kpaara 'boiling'

(10) Height and ATR harmony in Kimatuumbi

Hyman (2002), citing data from Paulian (1986a,b, 2001), describes a

(11) ATR, backness and roundness harmony in Kàlo¼1

b. /kù-fén-à/ → kù-fén-è 'dédaigner'

c. /kù-pós-à/ → kù-pós-ò 'aboyer'

d. /kù-yàn-à/ → kù-yàn-à 'to play'

(12) Yucatec Maya complete harmony and disharmony:

c. Transitive imperfective d. Perfective

Since rounding is redundant in the vowel systems of languages like Ainu

1.1.3 Restricted harmonies

(13) Yawelmani height uniform harmony

b. xat-hin 'eats, non-future'

c. xat-al 'might eat' d. xil-al 'might tangle'

Kaun (1994) gives a typology of triggering factors for rounding har-

1.1.4 Summary and discussion

(14) Types of vowel harmony

b. Feature combinations example language

The fact that roundness harmony predominantly co-occurs with backness

(15) A feature geometric organisation of vocalic place features

ATR/RTR high low back round

The idea behind such a representation is that according to association

(16) The Somali vowel system

(17) Somali vowel harmony

b. r((r-ka 'the village'

(18) Potential historical movement from ATR to backness

With the observation as background that being front is articulatorily pre-

(19) Natural feature combinations

To include the correlation of backness and roundness is not correct here if

c. s-tpúm-lxw 'hide with fur' u/o