2.1 The typology of consonant harmony as agreement

Consonant harmony can be broadly defined as the assimilatory interaction between two consonants which are separated by at least a vowel, where intervening segmental material (vowels and any non-participating consonants) appears to be phonologically unaffected by the assimilating property. Though such phenomena are not very common, they are not as rare as often assumed. An important finding of recent surveys (Hansson Reference Hansson2001b, Rose & Walker Reference Rose and Walker2004) is that, contrary to earlier claims (Gafos Reference Gafos1999; cf. Shaw Reference Shaw, Paradis and Prunet1991), consonant harmony is by no means limited to features subordinate to the [coronal] node (e.g. [±anterior], [±distributed]), or to articulatory gestures involving the tongue tip/blade.Footnote ¹ Although coronal harmony is certainly the predominant type, long-distance assimilations are attested for almost any conceivable featural dimension (major place of articulation being a notable exception). These include laryngeal features, nasality, the velar–uvular contrast in dorsals, laterality and/or rhoticity, secondary-articulation features and even continuancy. The examples in (1) illustrate this remarkable diversity.

1. (1) Examples of consonant harmony (from Hansson Reference Hansson2001b)

   1. a. stricture agreement

      

   2. b. laryngeal agreement

      

   3. c. dorsal agreement

      

   4. d. secondary articulation agreement

      

   5. e. liquid agreement

      

   6. f. nasal agreement

      

   7. g. coronal (non-sibilant) agreement

      

   8. h. coronal (sibilant) agreement

      

The most striking aspect of consonant harmony is the seemingly non-local nature of the interactions. For this reason, well-known cases of consonant harmony figured prominently in works where autosegmental tier-based notions of locality were being developed, often relying on underspecification and specific feature geometries (e.g. Poser Reference Poser, van der Hulst and Smith1982, Steriade Reference Steriade1987, Ao Reference Ao1991, Shaw Reference Shaw, Paradis and Prunet1991, Odden Reference Odden1994; see also Vaux Reference Vaux1999). Some recent studies argue instead that the non-locality is a mere illusion, and that the segments intervening between trigger and target consonants are in fact permeated by the articulatory gesture associated with the spreading feature, such that all phonological spreading is strictly local (Ní Chiosáin & Padgett Reference Ní Chiosáin and Padgett1997, Gafos Reference Gafos1999). Similar claims have been made, and supported with acoustic and/or articulatory phonetic measurements, regarding transparent vowels in many vowel-harmony systems: /i/ in Khalkha Mongolian tongue-root harmony (Rialland & Djamouri Reference Rialland and Djamouri1984), /a/ in Kinande tongue-root harmony (Gick et al. Reference Gick, Pulleyblank, Campbell and Mutaka2006), /i(ː) e(ː)/ in the palatal harmonies of Hungarian (Benus et al. Reference Benus, Gafos and Goldstein2004) and Finnish (Gordon Reference Gordon1999; but see Kim Reference Kim2005). Though such studies serve as a reminder that labels like ‘transparent’ or ‘neutral’ must be employed with great caution, cases of vowel transparency remain which cannot plausibly be explained in this manner. One is the transparency of /i ɪ/ (vs. opacity of /u ʊ/) to rounding harmony among non-high vowels in many Mongolian languages (Kaun Reference Kaun1995, Vaux Reference Vaux1999, Hansson Reference Hansson2006).

An analysis in terms of strictly local feature or gesture spreading is similarly untenable for many types of consonant harmony (e.g. stricture or nasal agreement; see (1a, f)). Even for cases where (perceptually covert) spreading through intervening vowels and consonants seems intuitively plausible, available evidence often speaks against it. In Tlachichilco Tepehua (Watters Reference Watters1988), shown in (1c), uvulars trigger dorsal agreement in velars (/k … q/→[q … q]), but also allophonic lowering of adjacent high vowels (/i u/→[e o]); both might conceivably be viewed as manifestations of a single spreading process. However, when a high vowel intervenes between the trigger and target consonants, but is adjacent to neither, it conspicuously fails to undergo any such lowering (/lak-puːtiq'i-ni-j/→[laqpuːteenij] ‘recounted it to them’, /ak-pitiq'i-j/→[aqpiteej] ‘folds it over’).Footnote ²

On the basis of the non-locality of consonant-harmony interactions, a radically different approach is taken in a series of recent works: the assimilation is interpreted as featural agreement at a distance, rather than feature spreading or gestural extension. By a proposal originally developed by Walker (Reference Walker2000a, Reference Walkerb, Reference Walker, Kirchner, Pater and Wikeley2001), agreement is taken to be mediated by a formal correspondence relation linking certain segments within the output string (Hansson Reference Hansson2001b, Reference Hansson2007, Rose & Walker Reference Rose and Walker2004, McCarthy Reference McCarthy, Bateman, Reilly, O'Keefe and Werle2006).Footnote ³ In the analysis of consonant harmony as correspondence-based agreement, the main factor which helps establish a correspondence relation between a pair of output segments is their relative similarity. The more similar two consonants are, the more likely they are to become (potentially, via correspondence) subject to agreement in some feature [F], and interaction between a pair of less similar consonants is restricted in ways that do not apply to more similar pairs. Consider, for example, the root-internal laryngeal agreement observed in Zulu and Ndebele, shown in (1b) above (Khumalo Reference Khumalo and Steven1987, Hansson Reference Hansson2001b, Reference Hansson2004a). This agreement is restricted to stops and, more interestingly, interacts with a general ban against laryngeally specified velar stops [kʰ g] outside of root-initial position. When the co-occurring stops are heterorganic, the ban overrides agreement; cf. Ndebele /-pʰek-a/ ‘cook, brew’, /-dak-w-a/ ‘be drunk’, and the Zulu loanwords /iː-pʰáketʰe/ ‘packet’, /iː-bakêde/ ‘bucket’ (note the transparency of [k] in the latter two examples). When the stops are homorganic, and thus more similar, agreement in turn overrides the preference for medial [k]; cf. Ndebele /-kʰokʰ-a/ ‘pull, draw out’, /-gug-a/ ‘wear out’, and Zulu loanwords like /-kʰekʰe/ ‘cake’. For a formal analysis of this similarity effect in correspondence-based terms, see Hansson (Reference Hansson2001b); for analogous cases, see also Walker (Reference Walker2000a, Reference Walkerb, Reference Walker, Kirchner, Pater and Wikeley2001), Rose & Walker (Reference Rose and Walker2004).

2.2 Functional motivations for consonant agreement

Dependence on trigger–target similarity is one of several salient characteristics which consonant agreement shares with phonological speech errors (see Hansson Reference Hansson2001a, Reference Hanssonb, Rose & Walker Reference Rose and Walker2004). In the correspondence-based analysis, these parallels are claimed to reflect the functional grounding of such agreement in the cognitive exigencies of speech planning and speech processing, mainly at the level of phonological (and phonetic) encoding for speech production. Dissimilatory co-occurrence restrictions manifested as Obligatory Contour Principle (OCP) effects, which often have non-local manifestations and likewise display significant similarity effects, have also been claimed to be rooted in processing factors (see Frisch et al. Reference Frisch, Pierrehumbert and Broe2004). With respect to the role of similarity in the planning domain, psycholinguistic speech-error studies have shown that segments are far more likely to interact in slips of the tongue if they share a large number of properties (Fromkin Reference Fromkin1971, Frisch Reference Frisch1996). In connectionist models (Stemberger Reference Stemberger and Ellis1985, Dell Reference Dell1986), this follows from spreading activation. When two co-occurring consonants C₁ … C₂ share a large number of features, there is extensive overlap in the units activated for C₁ and C₂. The greater the overlap, the greater the potential for unwanted interference, such that properties of one segment may intrude on the other.

Directionality patterns provide another parallel. In speech errors, anticipatory interactions tend to outweigh perseveratory ones by a considerable margin, other things being equal (Dell et al. Reference Dell, Burger and Svec1997). This is not surprising, given that speech planning and production unfolds in time: a segment which is being planned (and hence activated) may interfere with a highly similar segment which is currently being executed, whereas there is less reason for an already-executed segment to influence an upcoming one. To quote Dell et al. (Reference Dell, Burger and Svec1997: 123), ‘when the language-production system is working well, it looks to the future and does not dwell on the past’. The same asymmetry is evident in the typology of consonant agreement. Among the approximately 110 cases surveyed in Hansson (Reference Hansson2001b), the directionality – where one could be determined at all – was found either to reflect constituent structure (inside-out or ‘cyclic’ directionality, understood as stem control; Baković Reference Baković2000) or else to be consistently regressive/anticipatory.

A third and perhaps more striking parallel involves the ‘Palatal Bias effect’, whereby in slips of the tongue postalveolars such as /ʃ ʧ/ intrude on alveolars like /s t/ far more often than the reverse (Shattuck-Hufnagel & Klatt Reference Shattuck-Hufnagel and Klatt1979, Stemberger Reference Stemberger1991). This asymmetry, a preference for /s/→[ʃ] over /ʃ/→[s], is mirrored in the cross-linguistic typology of sibilant harmony systems (the most frequent subtype of consonant agreement by far). Of the 17 sibilant harmony systems which Hansson (Reference Hansson2001b) identifies as being in some way asymmetric along the [s]/[ʃ] axis, all but one show an asymmetry in the direction consistent with Palatal Bias. Moreover, the same is true of the extremely rare subtype of coronal harmony which affects stops rather than sibilants. Here, too, we find /t/→[ʧ] rather than vice versa, e.g. in Kera [t-eːŋa] ‘dry (fem)’ vs. [ʧ-əːʧ] ‘small (fem)’ with /t … ʧ/→[ʧ … ʧ]. Although the underlying causes of Palatal Bias are not well understood, the strong tendency towards the very same kind of asymmetry among coronal harmony systems is yet another argument in favour of the view that phonological consonant-agreement patterns have their roots in the speech-planning domain.

2.3 Diachronic aspects of consonant agreement

Even if we accept the hypothesis that consonant agreement derives from cognitive processing factors, it remains unclear what ‘derives from’ means in this context. The issue could be approached in at least two ways, not necessarily mutually exclusive. One is to assume a more or less direct and synchronic connection, whereby the very existence of constraints generating intersegmental correspondence (and agreement) is dependent on their being ‘functionally grounded’ in such processing factors (cf. Archangeli & Pulleyblank Reference Archangeli and Pulleyblank1994, Hayes Reference Hayes, Darnell, Moravcsik, Newmeyer, Noonan and Wheatley1999). An alternative approach is to view the connection as an indirect one, residing in the diachronic domain: processing factors exert their influence by facilitating, inhibiting or otherwise influencing diachronic changes which affect and shape the lexicon, and in this manner help create (and maintain) the co-occurrence patterns which constitute agreement. There are indeed cases where non-local assimilations of precisely the relevant kind have occurred as sporadic historical changes. For example, in what is the earliest attested occurrence of the term ‘consonant harmony’, Jespersen (Reference Jespersen1904) mentions diachronic correspondences such as French chercher<cercher (cf. English search) and the ‘vulgar’ Danish and German pronunciations [ʃerˈʃant]∼[ʃerˈʒant] for [serˈʃant]∼[serˈʒant] ‘sergeant’.

Diachronic explanations of this sort are feasible in particular for co-occurrence patterns which constitute static morpheme-structure constraints (MSCs). Frisch et al. (Reference Frisch, Pierrehumbert and Broe2004) propose an account along these lines for dissimilatory OCP[Place] restrictions, which are argued to be an effect of similarity avoidance. They suggest that over time, processing constraints incrementally (and, one presumes, sporadically) shape the lexicon as it is being acquired and augmented by successive generations of speakers: lexical items which are less problematic in processing terms will tend to be favoured in acquisition, in borrowing, in the coining of novel forms and in active usage. Categorical or near-categorical patterns of segmental co-occurrence across the lexicon can eventually arise as the cumulative result of this incremental process over long periods.

Rose & Walker (Reference Rose and Walker2004) speculate that a similar diachronic mechanism may be responsible for the emergence of consonant-agreement patterns as well. This view presupposes that all consonant agreement originates in static root-internal co-occurrence restrictions, and that where it reaches beyond the confines of root morphemes, this is due to some kind of analogical generalisation: ‘in circumstances where affixation forms words containing consonant combinations excluded within morphemes, the condition could be extended by analogy within a language to operate over the entire word’ (Rose & Walker Reference Rose and Walker2004: 489; emphasis added). While this is not an implausible hypothesis, it is not sufficient as a general purpose account of the diachronic origins and evolution of consonant agreement in the world's languages. It is true, as noted in §2.1, that consonant agreement is often manifested as morpheme-internal co-occurrence restrictions. However, there are many individual cases which appear to have evolutionary histories wherein the agreement pattern does not originate as an MSC. If we are to gain a proper understanding of consonant agreement as a general phenomenon, and of the extent to which cognitive processing and speech-planning factors are implicated in its emergence, we must direct our attention to the diachronic aspects of individual consonant-agreement systems (to the extent that these are known or can be reconstructed). The immediate goal of such an enterprise, to which this paper is a contribution, is to chart the diachronic trajectories which may lead to the emergence of synchronic consonant-agreement patterns, and which may play a role in shaping the properties of such sound patterns as aspects of synchronic grammars.

Another reason for investigating the diachrony of consonant agreement is to shed light on certain typological inconsistencies and asymmetries. While the typology of consonant agreement is remarkably consistent, especially considering the wide variety of features on which agreement may be based (see Hansson Reference Hansson2001b, Rose & Walker Reference Rose and Walker2004), many empirical generalisations are contradicted by one or two recalcitrant cases. For example, Bantu nasal agreement is the sole exception to the generalisation that when the directionality of consonant agreement is not derivable from constituent structure, it is always regressive, never progressive (Hansson Reference Hansson2001a, Reference Hanssonb). Likewise, a few systems are problematic in that the class of participating segments is difficult to characterise in terms of relative similarity (Hansson Reference Hansson2004a; see §4 below). Finally, it remains to be explained why certain types of agreement are exceedingly rare while others are much better attested. Earlier claims that the only attested type is agreement in [coronal]-dependent features (Gafos Reference Gafos1999), and perhaps laryngeal agreement as well (Shaw Reference Shaw, Paradis and Prunet1991), are certainly an overstatement, since consonant agreement is also attested for features like [±nasal], [±retracted tongue root] and even [±continuant]. Nevertheless, coronal agreement is vastly more common than any other type, and even within this category there are as yet unexplained asymmetries, such that coronal agreement involving stops and nasals is extremely rare compared to that involving fricatives and affricates. For many consonant-agreement types, only a small handful of cases are attested, and for some only a single unambiguous case has been documented (e.g. [±continuant] agreement in Yabem; see Hansson Reference Hansson2001b: 137–139).

In a study of the diachronic aspects of one such rare type, obstruent-voicing agreement, Hansson (Reference Hansson2004a) finds that the majority of attested cases can be accounted for by two diachronic scenarios, which, though conjectural, are supported by circumstantial evidence. The first scenario involves the interaction and interdependence of voicing and tone (see Hansson Reference Hansson2004b): obstruent voicing gives rise to low tone on an adjacent vowel, which may in turn trigger a percept of voicing on an obstruent on the other side of that vowel, and the resulting match in voicing values is reanalysed as a phonological agreement effect in its own right. The second scenario applies to systems in which voicing agreement coexists with agreement in another laryngeal feature ([spread glottis] or [constricted glottis]), where participation of [voice] in the agreement system appears to be of secondary origin, resulting from the analogical generalisation of agreement in one specific laryngeal feature to the set of all such features. The few remaining cases that are not amenable to either account involve MSCs which show acute sensitivity to similarity, and where agreement results in total rather than partial segmental identity. These are therefore highly compatible with the kind of diachronic explanation outlined earlier, whereby consonant-agreement patterns may arise as functional pressures shape the lexicon incrementally over many successive generations of speakers/learners.

In the remainder of this paper, I examine in detail a another cross-linguistically rare type of consonant harmony, secondary articulation agreement, with a view to exploring the role of diachrony in explaining the observed synchronic patterns. As I hope to demonstrate, an adequate understanding of the systems in question is only possible once their individual histories are taken into account.

4.1 The manifestation of agreement in Northwest Karaim

In Northwest Karaim, what corresponds to the palatal vowel harmony found in most other Turkic languages is instead a consonant harmony involving agreement in palatalisation. Like the Baltic and Slavic languages by which it has been surrounded for centuries (Lithuanian, Polish, Ukrainian, Russian, Belorussian), Northwest Karaim displays a consonant inventory wherein practically all consonants can occur either palatalised (/dʲ vʲ mʲ lʲ/) or non-palatalised (/d v m l˭/). The Northwest Karaim vowel inventory, shown in (2), also diverges from the typical Turkic one (for example, front rounded vowels have at best an extremely limited distribution), and shows affinities with the vowel systems of neighbouring languages as well.

1. (2) Vowel inventory of Northwest Karaim

   

Most importantly, vocalic [±back] contrasts are neutralised in most environments for most vowels, as signalled by the parentheses in (2). First of all, the [e]:[ɑ] contrast exists only in word-initial syllables, e.g. in /kʰʲelʲ-/ ‘to come’ vs. /kʰɑl˭-/ ‘to remain’ (cf. Turkish /gel-/, /kal-/), but is neutralised in favour of back [ɑ] in all other positions, as in the ablative suffix /-DɑN/→[-dʲɑnʲ]∼[-dɑn] (cf. Turkish /-dEn/→[-den]∼[-dɑn]; segments not contrastively specified for palatality are represented with capital letters). Secondly, the contrasts [y]:[u] and [ø]:[o] are even more restricted, confined to absolute word-initial position (i.e. in vowel-initial words). Thus we find /øzʲ/ ‘self’ vs. /on/ ‘ten’ (cf. Turkish /øz-/, /on-/), as well as /ysʲtʲ/ ‘top, upper’ vs. /us/ ‘reason, intellect’ (cf. Turkish /yst/, /us/), whereas neutralisation to back [u o] is observed in all other environments, e.g. /tʰʲuzʲ/ ‘smooth’ and /tʰuz/ ‘salt’ (cf. Turkish /dyz/, /tuz/).Footnote ⁹

Finally, though both members of the pair [i]:[ɯ] do occur in most positions, word-initially we find neutralisation to [i], regardless of whether the word/root in question behaves as front-harmonic or back-harmonic for the purpose of palatalisation agreement. For example, the 1st singular possessive suffix shows both vowel qualities (alternating predictably), [-imʲ]∼[-ɯm], whereas word-initially we see consistent [i] in pairs like /inʲɑ/ ‘needle’ and /irl˭ɑ-/ ‘to sing’ (Turkish /iːne/, /ɯrlɑ-/). In this respect, the relationship between Northwest Karaim [i] and [ɯ] is remarkably similar to that of [i] vs. [ɨ]/[ɯ] in neighbouring languages (Lithuanian, Polish, Russian). For example, most analyses of Russian treat the two qualities as allophones of a single /i/ phoneme, with [i] occurring word-initially, after vowels and after palatalised consonants, and [ɨ] after non-palatalised (phonetically often velarised) consonants (see e.g. Padgett Reference Padgett2003).

The palatalised and non-palatalised consonants of Northwest Karaim are subject to agreement, such that the consonants of a given word are either all palatalised or none is. Suffix consonants agree in palatalisation with those of the preceding stem, much like suffix vowels in most Turkic languages harmonise in [±back] with a preceding stem vowel. Examples are shown in (3), with cognate Turkish forms given for comparison. (Note that some alternations show the additional effects of rounding harmony.)

1. (3)

   

The palatal glide /j/ does not group with the palatalised consonants, and can occur in palatalised and non-palatalised contexts alike, as the examples in (4) illustrate.

1. (4)

   

Though descriptive sources all agree in stating that front-harmonic words contain only palatalised consonants, and back-harmonic words only non-palatalised ones, the interpretation of Northwest Karaim as a consonant-based harmony system – stated in very explicit terms by Kowalski (Reference Kowalski1929)  – has not gone unchallenged. Some authors advocate a ‘syllabic harmony’ analysis, where the [±back] property characterises the syllable as a whole and is expressed by vowels and consonants alike (see Denwood Reference Denwood2005 for a similar view). For example, Musaev (Reference Musaev1964: 52) rejects a consonant-harmony interpretation outright, in part on the basis of the residual existence of [±back] vowel contrasts [e]:[ɑ] (in initial syllables) and [y]:[u], [ø]:[o] (in absolute word-initial position), and in part by claiming that the alleged back vowels [ɑ u o] in back-harmonic words are not phonetically back at all, but rather front vowels (which he represents as <ə ü ö>). Csató (Reference Csató and Brendemoen1999) advocates a similar position (see also Csató & Johanson Reference Csató and Lars1996), transcribing <ȧ ȯ > in front-harmonic words as opposed to <a o u>, and describing the former as front vowels ‘pronounced in a somewhat retracted way’, presumably referring to central(ised) qualities like [a ɵ ʉ] vs. genuinely back [ɑ o u].

The existence of some indeterminacy along these lines is not surprising. The considerable overlap of vowel frontness and consonant palatalisation in the acoustic signal, and the consequent perceptual interdependence of the two, inevitably gives rise to ambiguities of localisation to a certain degree. In various Turkic languages with palatal vowel harmony, such as Turkish, consonants are known to exhibit palatalisation in front-harmonic words (Waterson Reference Waterson1956; cf. Ní Chiosáin & Padgett Reference Ní Chiosáin and Padgett1997, Reference Ní Chiosáin, Padgett and Lombardi2001). The converse is true in the Slavic and Baltic languages spoken in the vicinity of Northwest Karaim, where the [±back] contrast is generally attributed to consonants. Coarticulatory fronting in adjacent vowels is thus one of the strongest cues to consonant palatalisation in Russian (Derkach Reference Derkach, Fant and Tatham1975, Evans-Romaine Reference Evans-Romaine1998). As in the case of Karaim, the ambiguity has prompted some to question the attribution of [±back] contrasts to consonants in these languages (e.g. Bratkowsky Reference Bratkowsky1980).

There are two reasons to reject the claim that phonological backness harmony in Northwest Karaim harmony targets vowels and consonants equally, and that it is therefore not a case of consonant agreement. Firstly, even Musaev (Reference Musaev1964) acknowledges that in the speech of the younger generation, rounded vowels in palatalised contexts (Csató's <ȯ >) are fully back (e.g. [kʰʲozʲ] ‘eye’, a pronunciation which he represents as <kjoz> rather than <kjöz> or <köz>), even in absolute initial position (e.g. [juv(ʲ)] instead of [yvʲ] ‘house’). Musaev attributes this to influence from nearby Slavic (and, one presumes, Baltic) languages. Secondly, in examining the acoustics of Karaim vowels in recordings from Csató & Nathan (Reference Csató and Nathan2002), Nevins & Vaux (Reference Nevins, Vaux, Csirmaz, Lee and Walter2003) find many instances where fully back tokens occur between palatalised segments. This is especially true in initial and final syllables, which are also characterised by greater vowel duration than medial ones, e.g. in [kʰʲotʲ-ʉɾ-ʉlʲ-gʲunʲ] ‘lift yourself up’. They argue that it is precisely the short duration of word-medial vowels (leading to undershoot and increased coarticulatory overlap) which is responsible for the observed centralisation, not that the vowels are being actively targeted by a spreading [−back] feature (or corresponding articulatory gesture). In any case, the fully [+back] quality of the vowels of the peripheral syllables in this form, as well as in many other instances, clearly indicates that the harmony is genuinely a non-local palatalisation agreement among consonants.

4.2 Diachronic aspects of Karaim palatalisation agreement

The diachronic trajectory by which Karaim consonant-palatalisation agreement arose is largely uncontroversial. As a harmony system, it is a descendant of the [±back] vowel harmony found in most Turkic languages. The evolution of a consonant-palatalisation harmony from an inherited palatal vowel harmony was no doubt brought on in large part by the sociolinguistic situation of intensive and prolonged contact with languages in which the locus of phonological [±back] contrasts was consonants rather than vowels. In their reconstruction of the history of Northwest Karaim palatalisation agreement, Nevins & Vaux (Reference Nevins, Vaux, Csirmaz, Lee and Walter2003) posit the four-stage scenario summarised in Table I; my formulation differs from theirs only slightly. In the rightmost column, ‘y’ and ‘u’ are generic labels for front and back vowels respectively, whereas ‘U’ stands for a vowel which either lacks a [±back] specification or whose surface [±back] value is otherwise predictable from context.

Table I

Evolution of Karaim palatalisation agreement (after Nevins & Vaux Reference Nevins, Vaux, Csirmaz, Lee and Walter2003)

Two steps are central in this reconstructed history. The first of these (the I>II transition) is the appearance of consonant palatalisation in the context of [−back] vowels. This would have originated as a coarticulatory effect, but was possibly entrenched as a categorical phonologised interaction at some point. The second step (the II>III transition) is a covert reanalysis whereby [−back] specifications were attributed to consonants rather than to vowels, made possible by the inherent ambiguity of output configurations like [CʲyCʲyCʲ] with respect to the phonological source of the [−back] feature. A significant contributing factor here would have been the prevalence of phonemic C:Cʲ contrasts in the sound systems of the surrounding Baltic and Slavic languages, and the dependence in these languages of (phonetic) vowel fronting on (phonological) consonant palatalisation rather than vice versa. In effect, the inherited palatal harmony was ‘transphonologised’ from the vocalic to the consonantal dimension. Note that although the II>III transition is effectively the point at which consonant-palatalisation agreement arose as such, the true nature of the system is still largely covert at stage III, due to the vowel-fronting effects attributable to palatalised consonants. It is only when these (now redundant) fronting effects start disappearing at stage IV that the non-local and consonant-based character of the harmony system emerges overtly.

The synchronic properties of Karaim consonant agreement are all attributable to its origins in the metamorphosis of a pre-existing vowel harmony system. These include directionality (progressive, stem-to-suffix), but also some features which are highly unusual for consonant-agreement systems (see §2.1). For example, the agreement shows little or no sensitivity to relative trigger–target similarity, and is not limited to any specific subset of the consonant inventory, defined by such criteria as same place, manner and/or laryngeal features; instead, every consonant in the language falls within its purview. This is entirely as predicted by the above diachronic analysis, given that the vowel harmony ancestral to the agreement was oblivious to the featural composition of neighbouring consonants (and certainly insensitive to whether the consonants flanking a given vowel did or did not happen to have certain properties in common). Another striking feature of Karaim palatalisation agreement which follows from the diachronic analysis is the exemption of the palatal glide [j] from the agreement system (see (4) above). This seems surprising, as [j] embodies the very element which defines the secondary articulation on consonants like [mʲ] or [dʲ]. While one might imagine construing this as a ‘similarity effect’, restricting agreement to [+consonantal] segments, this interpretation is hardly insightful. The diachronic explanation is simpler and far more revealing: in Turkic languages in general, [j] is neutral and transparent with respect to [±back] vowel harmony. Since [j] occurred freely in front-harmonic and back-harmonic contexts all along (that is, in the predecessors of forms with and without consonant palatalisation respectively), no aspect of the distribution or behaviour of [j] has changed whatsoever.

5.1 The phonology of Tsilhqot'in pharyngealisation

Tsilhqot'in is unique among the Athabaskan languages, in that a pharyngealised/non-pharyngealised (or retracted/non-retracted, or ‘flat’/‘sharp’) contrast occupies a central place in its phonological system (Krauss Reference Krauss1975, Latimer Reference Latimer1978, Cook Reference Cook1983, Reference Cook1993). Tsilhqot'in is something of a frontier language, located on the southwestern edge of the Northern Athabaskan subgroup, and the prominent role played by pharyngealisation in its sound system has close parallels in the phonology of gutturals and vowel retraction in many neighbouring languages of the Salish, and to some extent Wakashan, families. There is thus clearly an areal component to the phonology of pharyngealisation in Tsilhqot'in (Cook Reference Cook1978, Reference Cook1993). The similarity ends there, however, as the specific details of the sound patterns have no direct counterpart in any of the neighbouring languages.

5.2 Pharyngealisation agreement in a diachronic perspective

There is little doubt that Tsilhqot'in sibilant agreement is an inherited sound pattern rather than an innovation. Coronal agreement with many of the same properties (regressive directionality, unboundedness) is attested in many other Athabaskan languages, not only in the Northern branch (e.g. Tahltan; Hardwick Reference Hardwick1984, Shaw Reference Shaw, Paradis and Prunet1991), but also in the Southern or Apachean branch (e.g. Navajo; Sapir & Hoijer Reference Sapir and Hoijer1967, McDonough Reference McDonough1990, Reference McDonough1991), and to a limited extent in the Pacific Coast branch (e.g. Tututni; Golla Reference Golla1976). As an MSC, sibilant agreement is reconstructed as far back as Proto-Athabaskan-Eyak (Krauss Reference Krauss1964): sibilants of the reconstructed *ts, *tš and *tšʷ series did not co-occur in roots except with members of the same series.Footnote ¹⁶ We have no evidence that sibilant agreement operated in heteromorphemic contexts at that stage, but this may be partly due to limitations of the comparative and internal reconstruction methodology. In any case, deep correspondences between the consonant-agreement patterns attested in different subgroups suggest that agreement alternations in prefix coronals are at least an early innovation.

Table II

Cognate sets (Krauss Reference Krauss, Hargus and Rice2005). Tahltan forms after Hardwick (Reference Hardwick1984).

Far more important than geographic/genetic distribution and superficial similarity is the fact that the particular coronals which interact in Tsilhqot'in pharyngealisation agreement correspond directly to those involved in most other Athabaskan coronal agreement systems. In spite of their striking dissimilarity from a synchronic phonetic-phonological point of view, the pharyngealised vs. non-pharyngealised alveolar contrasts in Tsilhqot'in, the dental vs. alveolar contrasts in Tahltan and the alveolar vs. postalveolar ones in Navajo and Apache are all historically cognate with one another, reflecting the Proto-Athabaskan *ts and *tš series respectively. This is shown schematically in (16) for a few representative languages, and illustrated by cognate sets from the same languages in Table II.Footnote ¹⁷ Since the Proto-Athabaskan *tš:*tšʷ contrast was merged in all relevant languages, I have conflated the two under *tš(ʷ) for simplicity. The reconstructed front velar * series is included, since these consonants shifted to sibilants in many daughter languages (see below) and thus had the potential of entering into the coronal agreement system. For each case in (16), the participating segments are boxed; dashed lines enclose those coronals which show only partial or secondary involvement.Footnote ¹⁸ In all cases, the two series which are centrally involved in agreement in (some of) the daughter languages are precisely those which diachronically reflect Proto-Athabaskan *ts and *tš(ʷ).

1. (16) Diachronic correspondence classes of agreeing sibilants

   

In Navajo and other members of the Apachean subgroup, the * series merged with the *ts series, and is indistinguishable from the latter with respect to agreement; cf. PA *ːtš'∼*ːš ‘knot’>Navajo /šaːž/ (instead of otherwise expected */saːž/). In many languages of the Northern subgroup, the fronting of the * series instead yielded [−anterior] sibilants (/tš/, etc.). This forms part of a general and pervasive sound shift which affected the majority of languages in the Northern Athabaskan subgroup, what Leer (Reference Leer, Jelinek, Midgette, Rice and Saxon1996) calls the Great Northern Series Shift (see also Howe & Fulop Reference Howe and Fulop2005):

1. (17) Great Northern Series Shift (Leer Reference Leer, Jelinek, Midgette, Rice and Saxon1996: 197)

   Sibilants become thibilants (interdentals) and shibilants become sibilants. … Palatal onsets become shibilants and uvular onsets become velars.

Many of the Northern Athabaskan languages spoken in the vicinity of Tsilhqot'in have merged, or are in the process of merging, the reflexes of the *ts and *tš series into a single alveolar /ts/ series. Consequently, these languages show no evidence of coronal agreement. These include Sekani (Hargus Reference Hargus1988), Babine-Witsuwit'en (Story Reference Story1984), Dakelh (Carrier; Morice Reference Morice1932, Story Reference Story1984), the Halfway River dialect of Beaver (Randoja Reference Randoja1989) and Fort Liard Slavey (Rice 1989). The fact that the merger is ongoing in some of these languages, and that where it is incomplete the contrast is a dental vs. alveolar one ([tθ]:[ts] or [t]:[ts]), strongly suggests that the Tsilhqot'in opposition of [+RTR] vs. [−RTR] alveolars goes back to an earlier stage where some kind of dental/alveolar contrast characterised the coronal series in question.

Tsilhqot'in sibilant agreement, as a sound pattern, is thus clearly inherited from Proto-Northern-Athabaskan (or perhaps a later, intermediate diachronic stage). What makes the Tsilhqot'in instantiation of this inherited consonant agreement unique – its being based on [RTR] rather than some coronal-dependent feature like [distributed] or [anterior] – is the result of a general shift in the coronal inventory which did not occur in any of the related languages. By this later change, dental obstruents ([tθ] or [t], etc.) became pharyngealised ones ([tsˁ], etc.). Howe & Fulop (Reference Howe and Fulop2005) discuss this change in the larger context of acoustically motivated sound shifts in Athabaskan languages, and posit an intermediate pre-Tsilhqot'in stage where dentals had acquired an enhancing feature of labialisation ([tθʷ], etc.). Their main motivation is to achieve the introduction of the Jakobsonian feature [flat], whose phonetic manifestation then shifted from labialisation to pharyngealisation. This may not be necessary, however, as connections between dentality and pharyngealisation (uvularisation, velarisation) are attested elsewhere; see the earlier discussion of St'át'imcets /z / in §5.101. Whatever the details of the [tθ tθ' dð θ ð]>[tsˁ tsˁ' dzˁ sˁ zˁ] change, areal convergence with neighbouring Interior Salish languages (St'át'imcets/Lillooet, Secwepemc/Shuswap, Nlaka'pamux/Thompson) likely played a contributing role. These languages all lack dental vs. alveolar contrasts, whereas they do have a pharyngealisation contrast in coronal sibilants (van Eijk Reference Eijk1997).

To summarise, then, we have seen that the emergence of secondary articulation agreement in Tsilhqot'in represents the transmutation of a pre-existing harmony system, whereby sibilant agreement in terms of a coronal-dependent feature like [±distributed] instead came to be based on the tongue-root feature [±RTR]. This was a consequence of the (areally motivated) shift of dentals to pharyngealised alveolars which took place in pre-Tsilhqot'in, and which has set Tsilhqot'in apart from all its Athabaskan relatives. The appearance of consonant pharyngealisation agreement as such is due to the remarkable coincidence of such a sound shift occurring in a language which already had a productive and pervasive coronal agreement between dental ‘thibilants’ and alveolar sibilants. In a sense, the Tsilhqot'in consonant-agreement system could be said to represent the fortuitous and unlikely collision of a typically Athabaskan coronal agreement system with a (very loosely speaking) typically Interior Salish coronal inventory.

5.3 The relationship between agreement and ‘flattening’

As noted in §5.1.2, many characteristics are shared by regressive Ŝ-flattening and sibilant agreement, including those which set regressive Ŝ-flattening apart from the other vowel-retraction processes. The parallels are summarised in (18); see Cook (Reference Cook1993: 166).

1. (18) Comparison of sibilant pharyngealisation agreement with regressive Ŝ-flattening

   1. a. Both are triggered by sibilants specified for [±RTR].

   2. b. Both apply in a regressive/anticipatory fashion.

   3. c. Both are unbounded, spanning the entire word.

   4. d. Both are unaffected by intervening segments with a conflicting value for [±RTR], which are thus transparent (sharp dorsals in the case of Ŝ-flattening, flat or sharp dorsals in the case of sibilant agreement).

   5. e. Sibilant agreement targets consonants, while Ŝ-flattening targets vowels, but the feature involved in both processes ([±RTR]) is perceived on vowels.

In addition to these striking parallels, we saw in §5.1.3 that there is extensive overlap between the manifestations of the two sound patterns. The effects of [+RTR] sibilant agreement are masked completely by regressive Ŝ-flattening. If a sibilant acquires an output [+RTR] specification by agreement with a [+RTR] sibilant later in the word (S…Ŝ→Ŝ…Ŝ), this is manifested phonetically through retraction of any vowels near the target sibilant. But since regressive Ŝ-flattening is unbounded and never blocked, the same vowels would be retracted in any case, due to direct spreading from the agreement-triggering [+RTR] sibilant. Due to this overlap, there can never be any surface evidence for the pharyngealising version of agreement. The only recoverable evidence for the existence of sibilant agreement comes from its depharyngealising [−RTR] version (Ŝ…S→S…S), in that this process bleeds Ŝ-flattening. The limited evidence base for sibilant agreement as such may be leading to its disappearance. Andrews (Reference Andrews1988) reports many forms where agreement with a later [−RTR] sibilant fails to bleed flattening (though some examples may be due to the ambiguity of [ə] as an allophone of /ε/; see §5.1.2).

I propose that neither the list of parallels in (18) nor the overlap and masking just described are coincidental. Instead, they are indicative of a causal link in the diachronic dimension between these two sound patterns: regressive Ŝ-flattening in its current form represents the analogical reanalysis, or ‘telescoping’, of the combined effects of non-local sibilant agreement and local vowel retraction. The proposal can be summarised as follows. At an earlier stage, all vowel flattening was local, affecting only the vowels of adjacent syllables, or perhaps reaching past a lax vowel into a non-adjacent syllable (as in progressive Ŝ-flattening today). In forms with more than one sibilant, unbounded regressive sibilant agreement and local vowel flattening (emanating bidirectionally from said sibilants) typically had the combined effect that any and all vowels preceding an agreement-triggering (i.e. ‘rightmost’) [+RTR] sibilant ended up surfacing as retracted. For example, if we were to hypothesise that regressive Ŝ-flattening operated as a mirror image of what we see in progressive Ŝ-flattening, all four retracted vowels in (14c) /næ#tε-ŝε-s-d-duẑ/→[natʌzˁʌsˁdozˁ] would be accounted for by the combination of sibilant agreement and such short-range Ŝ-induced flattening. The source of retraction on those vowels is thus ambiguous between a local [+RTR] sibilant (the agreement target) and a distal one (the agreement trigger, which furthermore was the obvious source of the lexical [+RTR] specification). I suggest that such ambiguous contexts were reanalysed as involving unbounded vowel flattening triggered directly by the last (and underlyingly [+RTR]) sibilant, and that this reanalysis is what has lent regressive Ŝ-flattening its current properties and special status. An overt consequence of the reanalysis is that regressive Ŝ-flattening is now applied in an unbounded fashion in all contexts, even in forms which contain no additional sibilants.

Although this analysis is somewhat conjectural, there do exist concrete examples which show reanalysis of sibilant Ŝ:S distinctions due to ambiguities in localising the phonological source of [+RTR] effects on vowels. Krauss (Reference Krauss1975) notes that all stems which were historically ŜVQ have been reanalysed as SVQ in Tsilhqot'in. When the source of stem-vowel retraction was ambiguous, the [+RTR] specification was thus consistently attributed to the following (Q-series) consonant alone. As Krauss (Reference Krauss1975: 37–38) puts it, ‘preceding a flattened vowel, flat sibilants are in no way themselves phonetically distinguishable from sharp. Modern Chilcotin, instead of remembering the history, now treats all such initials as sharp rather than flat’. As in the reanalysis scenario described above, this restructuring would also have led to altered pronunciations of certain forms, as in the examples in (19), from Krauss (Reference Krauss1975: 38). Having been recast as a S-series sibilant, the historically Ŝ-series sibilant of a ŜVQ root now fails to trigger regressive vowel flattening, and triggers [−RTR] agreement on any preceding sibilants.

1. (19)

   

The development of unbounded regressive Ŝ-flattening out of a combination of unbounded sibilant agreement and local vowel retraction is shown in (20) and (21). Multiple-sibilant forms like (20) provided the pivot for analogical reanalysis. Propagation of [+RTR] by sibilant agreement is shown at the bottom, with local [+RTR] spreading at the top. I assume the original version of regressive Ŝ-flattening to have been like its progressive counterpart, spanning two syllables if the closer vowel was lax (see §5.1.2).

1. (20)

   

In multiple-sibilant forms like those in (20) (as well as in single-sibilant forms that were relatively short, or ones where the sibilant was centrally located within the word), the reanalysis is entirely covert: the surface realisation is the same before and after the change. For forms containing only one ([+RTR]) sibilant, and in which that sibilant was located relatively far from one or more of the vowels present in the form, the reanalysis has some overt consequences, in that certain vowels are affected by retraction which otherwise would not have been. One such example is shown in (21); note that for simplicity the regular contraction of /tε-ʁε-/ to /tæ/ (realised as [tæ]∼[ta], depending on the [±RTR] context) has been built into the ‘underlying’ representation in the diagrams.

1. (21)

   

The proposed diachronic analysis of Tsilhqot'in sibilant agreement, and its interplay with (regressive) Ŝ-flattening, has several explanatory advantages. First of all, the account explains the asymmetric (and synchronically problematic) profile of the regressive Ŝ-flattening process relative to the other vowel-retraction processes, especially as regards its unboundedness and the transparency of intervening [−RTR] consonants. Secondly, the diachronic analysis explains the parallels between sibilant agreement and regressive Ŝ-flattening (with respect to directionality and unboundedness), rather than viewing these as purely coincidental. Thirdly, it explains how sibilant agreement and regressive Ŝ-flattening have come to coexist in Tsilhqot'in despite the considerable redundancy arising from their overlapping effects (which may in turn be leading to the disappearance of sibilant agreement as such).

A fourth advantage of the diachronic analysis, pertaining to the interpretation of Tsilhqot'in sibilant pharyngealisation agreement as reflecting an earlier agreement based on a dental vs. alveolar contrast, is the light it sheds on how pharyngealisation agreement in Tsilhqot'in fits into the greater picture of coronal agreement in Athabaskan (and cross-linguistically), despite its synchronic uniqueness. A fifth and final advantage is that the proposed account permits us to formulate in more concrete and precise terms the notion that pharyngealisation-related aspects of Tsilhqot'in phonology may be partly due to areal contact with neighbouring Salish languages (Cook Reference Cook1993: 150). More specifically, the change of dental ‘thibilants’ into pharyngealised sibilants, as well as the local flattening (retraction) effects on nearby vowels by Q- and Ŝ-series consonants, can likely be blamed on language contact or areal convergence. By contrast, the existence of a consonant-agreement system involving coronals is an inherited feature, while its instantiation as pharyngealisation agreement is a mere by-product of the change just described. The complexities of the vowel-flattening processes as a whole, and the special status of regressive Ŝ-flattening within that set, are an entirely Tsilhqot'in-internal phenomenon, arising from the interaction and overlap between [RTR]-related vowel-retraction effects and the inherited consonant agreement which has come to be based on [±RTR] as well.

This study of Tsilhqot'in sibilant pharyngealisation agreement reveals the benefits of diachronic-evolutionary explanations for synchronic sound patterns. A better understanding of the consonant-agreement system is gained: why it involves the particular sibilants it does, why the operative feature is [±RTR], why it shows the directionality effects that it does, and so forth. We have also deepened our understanding of the complex evolutionary trajectories that can produce consonant-agreement patterns which are otherwise rare or unattested. Finally, the diachronic analysis yields explanations for other elements of the synchronic phonology of Tsilhqot'in. In short, diachrony provides us with important clues to two synchronic puzzles: vowel flattening (its internal complexity and interaction with consonant agreement), and the very existence of such a typologically rare phenomenon as long-distance consonant-pharyngealisation agreement.