3.1 Linear order: Lexical, phonological and/or phonetic representation

Sequences of linearly ordered acoustic events which, however, appear to instantiate only one single phonological unit have left a noticeable footprint in the literature. Early discussion includes Martinet (Reference Martinet1939) and Hoard (Reference Hoard1967: 34ff.). The latter offers a solution that complies with the contemporary view on melodic organisation: like all other segments, complex segments are a feature matrix, which, however, is multi-columned. In the case of an affricate, the features [−cont] and [+cont] appear in different columns within the matrix (see also Hoard Reference Hoard, Bell and Hooper1978). The next step were autosegmental representations, which were developed since the mid-1970s. They allowed for a straightforward formulation of this insight: Clements & Keyser (Reference Clements and Keyser1983: 34f., 85ff.) set the standard autosegmental representation of affricates as distinct stop and fricative branches that are dominated by a single skeletal position.

A different take first appears in Fujimura & Lovins (Reference Fujimura, Lovins, Hooper and Bell1978), who advance the idea that acoustic events which occur in sequence may be unordered in the phonology and linearised only upon phonetic interpretation (in their view, by phonetic realisation rules). While this concerns what they call the syllable core (a vowel plus a consonantal satellite on each side), the idea is adapted to (light) diphthongs by Kaye (Reference Kaye and Goyvaerts1985: 289 – ‘[b]y universal convention the less sonorous of the two elements associated to the same point is produced first in the speech chain’), and to consonantal complex segments by Sagey (Reference Sagey1986: 52ff.). The latter author distinguishes between two categories: contour segments (like affricates and prenasalised stops) whose linear order is specified in the phonology, and complex segments (like labio-velars: kp) which are unorderd in the phonology and receive their linear structure upon phonetic interpretation. Sagey's diagnostic for distinguishing the two types of monosegmental items are what she calls edge effects, i.e. the behaviour of the linearly distinct components in regard of phonological processes: affricates behave as stops in rules sensitive to their left edge, but as fricatives in rules sensitive to their right edge. Hence the order of the two branches must be available in the phonology. By contrast, labio-velars, for example, show no sensitivity to either edge: they always behave like both labials and velars. Therefore Sagey concludes that their linear order is absent from the phonology. This echoes a point made by Anderson (Reference Anderson1976: 343): ‘the segmental idealization is no longer quite so pervasive, since segments have internal structure which may be manipulated by rules’. In Anderson's data, phonological processes could concern either the nasal or the non-nasal part of a prenasalised consonant.

Pace Sagey, Lombardi (Reference Lombardi1990) argues that there are no edge effects: Sagey's data are either reanalysed or interpreted as the result of phonetic, rather than phonological, activity. Lombardi proposes that affricates (and only affricates: she is not concerned with other types of segments) are monosegmental items which bear a stopness and a continuancy feature that are unordered in the phonology. Linear ordering occurs upon phonetic interpretation for universal reasons (there are no mirror-image affricates like st, ʃt, etc., and there is no contrast between affricates and their hypothetical mirror images), and edge sensitivity is thus predicted to occur in phonetic processes. In addition, Lombardi levels anti-edge effects against phonological ordering, for example, cases where a rule applies after stops and affricates. This shows that the phonology has access to both the continuant and the non-continuant component of affricates.

Also engaged in the debate on whether or not melodic primes are ordered in complex monosegments is Rennison (Reference Rennison and Cyran1998), who views the component that is realised last in two-membered complex items as floating, or ‘lazy’. That is, the component that is first pronounced is the ‘regular’ consonant, which is then modified by a new melodic prime, the ‘lazy’ one, that may either be lexically present or acquired from context (e.g. in assibilation). In Rennison's perspective, the linear order is thus programmed in the phonology: the component that is realised first is specified as such in the lexicon, and may be supplemented in further pronunciation either by other melodic primes that are lexically present (floating), or that are provided during phonological computation.

The study of complex segments in general, and the view that features are unordered in particular, has also consequences for the kind of melodic primes that are used: monovalent, rather than binary features are necessary in order to run unordered complex segments. An early contribution in this direction is based on the study of prenasalised consonants: Anderson (Reference Anderson1976: 332) argues ‘against any solution to the problem of prenasality in which a single feature has the entire segment as its domain’ because the difference between N, ⁿC and Cⁿ is one of timing, i.e. the moment at which the velum is lowered. ‘For this reason any feature which purports to characterise the entire segment uniformly, like [continuant] or [sonorant], is apparently destined to fail’. Lombardi (Reference Lombardi1990: 377f.) discusses the same issue from a different perspective: how could [−cont] and [+cont] occur simultaneously in the same segment, i.e. an affricate in her case? On the basis of other equipollent features such as [±anterior], Lombardi concludes that opposite values of the same feature cannot co-occur within a single segment. She therefore introduces two distinct features that do the same labour as [−cont] and [+cont]: [stop] and [cont]. An affricate bears both of them (in unordered fashion), and as a consequence both must be monovalent, since [stop] is the equivalent of [−cont].

Steriade (Reference Steriade, Huffman and Krakow1993: 461; 1994: 210ff.) arrives at the same conclusion for different reasons, namely the fact that the only reliably attested segmental contours – for nasality and continuancy – occur with released plosives, and also that plain consonants, i.e. unaspirated and unglottalised, are always good landing sites for aspiration and glottalisation. If they bore [−spread] and [−constricted] specifications, they would be expected, at least in some cases, to block the spreading of the positive value of these features (since positive and negative values of the same feature cannot co-occur). Against this background, Steriade (Reference Steriade, Huffman and Krakow1993: 459f.; Reference Steriade, Cole and Kisseberth1994: 216f.) introduces both universal and parametric conditions for the co-occurrence of features within a single segment. Universal in her view are physiological impossibilities such as the simultaneous realisation of an aspirated and a glottalised release: [+spread] and [+constricted] cannot be simultaneously pronounced. On the other hand, languages choose parametrically which features are allowed to hook onto a single segment. Steriade (Reference Steriade, Cole and Kisseberth1994: 217) therefore argues that ‘even features that are not simultaneous on the surface must pass a test of mutual compatibility in order for the structure encompassing them to count as monosegmental’.

Underlying this perspective is the idea that some feature combinations within a single segment are uninterpretable, either universally or on a language-specific basis. There are two ways to implement this idea, though. In the one we have seen, uninterpretable feature combinations are illegal and cannot exist in the system. An alternative view is that the uninterpretability of a given feature set causes a linear differentiation upon phonetic interpretation: the features are uninterpretable simultaneously, but can be pronounced in a sequence. Based on Jakobson, Fant & Halle (Reference Jakobson, Fant and Halle1952), this line of thought is followed, for affricates, by Hirst (Reference Hirst, Guéron, Obenauer and Pollock1985: 93), Hualde (Reference Hualde1988), La Charité (Reference La Charité1993), Rubach (Reference Rubach1994), Kim (Reference Kim1997), Clements (Reference Clements, Fujimura, Joseph and Palek1999) and Kehrein (Reference Kehrein2002: 7ff.). For these authors, there is no [+cont] component in affricates at any derivational stage, or in the lexicon; affricates are regular stops whose only peculiarity is to possess a [+strident] feature. Since [−cont, +strident] cannot be executed simultaneously, the segment is spelled out as a sequence of acoustic events at the phonetic level. With reference to feature co-occurrence in (English) onset clusters, Hirst (Reference Hirst, Guéron, Obenauer and Pollock1985: 92) extends this perspective to (English) muta cum liquida (pl, tr, gl, etc.), which according to him are a single segment that is specified for [−cont, +son, −nas]. Since no segment can be an oral sonorant stop, the features at hand cannot surface unless their components are separated in a linear order.Footnote ²

There are thus two distinct views on the question why phonologically unordered features appear as a sequence on the surface. On one view, the reasons are purely phonetic (and universal) and have nothing to do with phonological specification; the stop component is pronounced before the fricative component (e.g. Lombardi Reference Lombardi1990). On the other view, the reasons are also phonetic but anchored in the phonology – uninterpretability of feature sets stems from the fact that affricates are strident stops.

Let us now turn to the question of how complex segments arise (whatever their internal structure). Are they present in the lexicon, or are they distinct segments underlyingly which merge in the course of the derivation? That lexically distinct items merge into complex segments at least in some languages and at least for some types of complex segments is shown by the trivial fact that in many Bantu languages the nasal and the non-nasal component of prenasalised stops belong to distinct morphemes (e.g. Herbert Reference Herbert1977). The issue is extensively discussed by Steriade (Reference Steriade, Cole and Kisseberth1994), who concludes that some mergers are obligatory and universal, while others are governed by parametric variation. Aspirated stops C^h instantiate the former case: they may occur as such in the lexicon, but if a C and an h happen to form a tautosyllabic cluster, they always merge. This is, Steriade (Reference Steriade, Cole and Kisseberth1994: 215) argues, the reason why no language opposes C^h and C–h. On the other hand, parametric variation is responsible for the fact that n+t^h merge into a prenasalised stop ⁿt^h in Tlacoyalco Popoloca (Oto-Manguean), but not in the closely related Western Popoloca (Steriade Reference Steriade, Cole and Kisseberth1994: 272). Van de Weijer (Reference van de Weijer1996: 158) argues along the same lines regarding prenasalised stops. Like Steriade, Hirst (Reference Hirst, Guéron, Obenauer and Pollock1985: 92) holds that muta cum liquida are single segments in some languages (English in his case), and like her, he admits that the same sequences may be bisegmental in other languages (e.g. when they are broken up by vowels, as in Moroccan Arabic). At variance with Steriade, however, Hirst does not provide for a mechanism that merges lexically distinct segments into a single segment during phonological computation: his monosegmental muta cum liquida are already monosegmental at the underlying level.

3.2 Diagnostics for identifying monosegmental items

In this section we take a closer look at the kind of arguments advanced to distinguish monosegmental from bisegmental items. A very common argument concerns the distribution of complex segments, which is one of singletons, rather than of clusters: for example, complex segments may occur in positions where clusters are impossible (Anderson Reference Anderson1976: 331; Steriade Reference Steriade, Huffman and Krakow1993: 458; Clements Reference Clements, Fujimura, Joseph and Palek1999: 271; Hualde Reference Hualde1991).

Another common argument looks at inventories: for example, in a language where voiced prenasalised stops are the only voiced stops, their interpretation as a cluster would leave the language without voiced stops (that are independent of a preceding nasal) (Anderson Reference Anderson1976: 331). Along the same inventory-based line is what Clements (Reference Clements, Fujimura, Joseph and Palek1999: 271) calls ‘noncompositionality: a sequence of sounds, say a stop+fricative string, cannot be plausibly decomposed into a sequence of two independently-occurring phonemes’.

Contrast is also often quoted. For example, when prenasalised stops contrast with singletons (stops and nasals for instance), but not with clusters, one may conclude that they are monosegmental as well (Anderson Reference Anderson1976: 311). Clements (Reference Clements, Fujimura, Joseph and Palek1999: 272) calls his argument based on contrast ‘noncommutivity: the stop and/or fricative component [of an affricate] does not commute with another phoneme’. Still based on contrast, but this time with clusters, is what Clements (Reference Clements, Fujimura, Joseph and Palek1999: 272) calls ‘contrastivity: some languages (e.g. Czech, Polish) have lexical contrasts between affricates and the corresponding stop+fricative sequences’. The contrast found in Polish between czy [] ‘whether’ and trzy [tʃɨ] ‘three’ (and a number of other minimal pairs of the kind) is often quoted when this point is made (e.g. Hoard Reference Hoard1967: 36f.; Anderson Reference Anderson1976: 311; Herbert Reference Herbert1977: 440; Ewen Reference Ewen, van der Hulst and Smith1982: 52; Clements & Keyser Reference Clements and Keyser1983: 34f.).

Another argument concerns syllabification and sonority sequencing. Word-initial prenasalised stops, for example, would violate the latter, were they separate segments (Ewen Reference Ewen, van der Hulst and Smith1982: 42, 54; Hualde Reference Hualde1991; van de Weijer Reference van de Weijer1992: 133; Reference van de Weijer1996: 54). The corresponding argument that Clements (Reference Clements, Fujimura, Joseph and Palek1999: 271) calls tautosyllabicity is about syllabification: ‘affricates syllabify as syllable onsets in contexts where bisegmental stop+fricative sequences usually syllabify as coda+onset sequences’.

Another factor for deciding whether a sound sequence is mono- or bisegmental revolves around the aforementioned feature co-occurrence. Steriade (Reference Steriade, Cole and Kisseberth1994: 217) provides a definition of what a single segment is which relies partly on theory-internal considerations, but involves one theory-neutral factor: a sound sequence that has more than one place of articulation cannot be a complex segment. Hence muta cum liquida pr, kl, tr, etc. do not qualify. Note, however, that Steriade's notion of single segmenthood is more complex than that (and may appear inconclusive): she operates with two distinct notions, ‘featurally monosegmental’ and ‘structurally monosegmental’. An item may be featurally non-monosegmental, but qualify for structural monosegmenthood: this is the case of muta cum liquida. According to Steriade (Reference Steriade, Cole and Kisseberth1994: 213, 221), these have indeed the same structure as simplex (released) stops (i.e. an A₀ followed by an A_max position in her system, where A stands for aperture position): they are structurally monosegmental. At the same time, though, muta cum liquida are bisegmental regarding their featural makeup, which contains two distinct sets of place features.

Among all arguments for monosegmenthood, though, the one drawing on phonological processes is certainly most pervasive. Sagey (Reference Sagey1986: 69ff.) offers a number of relevant examples. One (73ff.) is built on Kinyarwanda (Bantu), where compensatory lengthening of the vowel preceding prenasalised (/ku-ngana/→kuu^ŋgana ‘to be equal’), but following labio-velar consonants (/ku-gu-ir-a/→kug^wiira ‘to fall on’) is observed. This is explained by the fact that two lexically independent items have merged into a complex segment, and the x-slot that was vacated by this operation is colonised by the adjacent vowel. Clements (Reference Clements, Fujimura, Joseph and Palek1999: 272) contributes two patterns to this argument: ‘inalterability: affricates fail to undergo rules of cluster reduction that apply to bisegmental consonant sequences; [and] inseparability: the stop+fricative sequence cannot be broken up by epenthesis, reduplication, speech errors, and so forth’. Other cases are quoted by Hualde (Reference Hualde1991: 127) and van de Weijer (Reference van de Weijer1996: 134ff.). The latter author mentions the Sahidien variety of Coptic, where the definite article appears as p- (masc.), and t- (fem.) when the noun begins with a single consonant (p-to ‘the earth’, t-soone ‘the sister’), but is realised as pe-, te- when followed by a cluster (pe-thab ‘the leaven’, te-hbuur ‘the left’). Like single consonants, the affricate does not provoke the vocalisation of the article (p-čo ‘the armpit’, t-čamee ‘the calm’), which shows that it counts as a single segment. Another source of evidence is reduplication. Van de Weijer (Reference van de Weijer1996: 135) reports a case from Ewe (Niger-Congo) where only the first consonant of a word-initial cluster is reduplicated (ɲra – ɲa-ɲra-la ‘to rave, raver’, fle – fe-fle ‘to buy, bought’). When the initial cluster begins with an affricate, the entire affricate is copied (á – á-á ‘sell, selling’), and this is also the case when the affricate is alone ( – - ‘to grow, grown’). Affricates are thus treated as single segments.

Steriade (Reference Steriade, Cole and Kisseberth1994: 204f.) argues against the bisegmental status of what is traditionally interpreted as a branching onset: ‘one expects all prosodic constituents … to be the domain of phonological rule application. But the subconstituents of the onset … never play this role of prosodic domain’. McCarthy & Prince (Reference McCarthy and Prince1996: 70) make an argument from templatic morphology: ‘[i]n Modern Hebrew and in Arabic, the respective affricates c and j spread intact to multiple positions in the skeleton: kicec “he cut” … , hajjaj “he made the pilgrimage to Mecca”. To derive these results it is crucial that affricates constitute a single melodic element that spreads’ (Lombardi Reference Lombardi1990: 409 and van de Weijer Reference van de Weijer1992: 133 also make this point).

Lowenstamm (Reference Lowenstamm and Ploch2003: 340ff.) presents an argument along the same lines, which, however, offers a diagnostic for the mono- or bisegmenthood of the same phonetic object. Quadriliteral forms of Chaha (Ethio-Semitic) identify, among other things, by bearing ɪ in V₁. Hence the regular form mɪsäkär-ä ‘to testify, 3sg masc.’ and fɪc′änäq-ä ‘to squash, 3sg masc.’ (ä denotes a low schwa). In q ^wɪräc′ä ‘to take a handful and close the hand, 3sg masc.’, however, one consonant seems to be missing. Lowenstamm argues that the palatal ejective stop c′ may represent two distinct lexical items (and two distinct items in the consonantal inventory of the language): in fɪc′änäq-ä, it is a monosegmental unit (/t^y′/), while it identifies as a bisegmental item (/t′y/) in q ^wɪräc′ä. Steriade (Reference Steriade and Lacy2007: 140) reports a number of parallel cases where a language contrasts a surface sequence XY according to whether it is mono- or bisegmental: kw vs. k^w in reconstructed Indo-European, tʔ vs. the ejective t^ʔ in Yokuts (native American, California).

Lowenstamm (Reference Lowenstamm and Ploch2003: 343ff.) further argues that muta cum liquida are never branching onsets in any language: branching onsets do not exist, and muta cum liquida are either monosegmental or heterosyllabic (like in typical Semitic languages). An argument he makes against branching onsets as such is their inertness with respect to stress: stress assigning algorithms may or may not be sensitive to branching rhymes, but never make a difference between simplex and what is supposed to be branching onsets.

On the phonetic side, Byrd (Reference Byrd1996a, Reference Byrdb) has found that the segmental unit displays relative stability of timing and coordination: there is less gestural variability and overlap (in timing) within segments than among sequences of segments. Also, Byrd (Reference Byrd1996a: 161) has observed that speech rate has less impact on segments than on sequences of segments. Related to this is the obvious point that monosegmental items should be shorter than bisegmental clusters (especially according to autosegmental standards where the former occupy one x-slot, while the latter extend to two timing units). The duration of familiar complex segments such as affricates, prenasalised stops, etc., however, is not significantly longer than the one of simplex consonants (e.g. Ewen Reference Ewen, van der Hulst and Smith1982: 42, 60; Sagey Reference Sagey1986: 78; Maddieson Reference Maddieson1989; van de Weijer Reference van de Weijer1996: 158f.).

Finally, Fromkin (Reference Fromkin1971: 33) makes a point regarding affricates on the basis of a corpus of English speech errors: ‘[i]t is interesting to note that while [str], [pl], [kr], [bl], [fr], etc., as well as final clusters, reveal the splitting of clusters into segments, not a single example in my own data, or the English examples cited by others, shows a splitting of [tš] or [dž] into sequences of stop plus fricative’.

3.3 Diagnostics for identifying the internal structure of monosegmental items

A related issue concerns the internal structure of items that are agreed to be monosegmental. It was reported in Section 3.1 above that there are two basic options: linear structure is either a phonological or a merely phonetic property. In the latter case, either purely phonetic reasons cause linearisation (e.g. aerodynamics), or linear realisation is a consequence of the attempt to execute features that cannot be simultaneously pronounced (for either universal or language-specific reasons).

Again, the evidence that is most pervasively used in order to distinguish between these options comes from phonological processes. Anderson (Reference Anderson1976: 337) writes that ‘[i]f segments with such internal structure can exist, however, we would also expect phonological rules to be able to create them, and in general we would expect phonological rules to be able to manipulate the size of the domain of a feature specification wherever that is a variable’. He then examines a number of relevant processes, and concludes, on page 343, that ‘segments have internal structure which may be manipulated by rules’. Recall from Section 3.1 that so-called edge effects, i.e. processes which involve only the left or right element of complex items, are the basis for Sagey's (Reference Sagey1986) distinction between contour and complex segments. This distinction was then counteracted by Lombardi's (Reference Lombardi1990) anti-edge effects. Other cases where phonological processes are used as a diagnostic for the internal structure of monosegmental items include Rosenthall (Reference Rosenthall, Dressler, Luschützky, Pfeiffer and Rennison1992: 252ff.), La Charité (Reference La Charité1993), van de Weijer (Reference van de Weijer1993: 89, 99), Rubach (Reference Rubach1994) and Kim (Reference Kim2001).

Another recurrent argument is overgeneration. In order to avoid the possible existence of objects (and hence contrasts) that are unattested, Sagey (Reference Sagey1986: 50) restricts branching within the feature geometrical tree to terminal nodes (which represents an evolution with respect to Clements' (Reference Clements1985) original feature geometric model). Clements (Reference Clements, Fujimura, Joseph and Palek1999: 272) writes that ‘[a]ny further representational distinction, such as one involving a feature [+continuant] (whether contoured or simultaneous), would make incorrect predictions concerning affricate behaviour, and massively overgenerate the number of potential contrasts predicted by the representational system’, and concludes, on page 277, that ‘special analyses of the sort represented in (1) [i.e. involving the feature [+cont]] are not just superfluous, but predict the existence of many more speech sounds than are actually attested’. Van de Weijer (Reference van de Weijer1992: 133) and Rosenthall (Reference Rosenthall, Dressler, Luschützky, Pfeiffer and Rennison1992: 250f.) are also concerned with overgeneration.

Another line of attack concerns the (universal or language-specific) interpretability of feature combinations that was already discussed in Section 3.1. On the one hand, the internal structure of monosegmental items is strongly marshalled by the universal impossibility to accommodate two conflicting values of the same feature in a single melodic makeup. Hence Lombardi (Reference Lombardi1990: 377f.), who argues that the correct representation of affricates involves the unordered cohabitation of [+cont] and [−cont], must express this distinction by two different features, [stop] and [continuant]. These are then necessarily monovalent. Steriade (Reference Steriade, Huffman and Krakow1993, Reference Steriade, Cole and Kisseberth1994) is also strongly committed to monovalent features, for the same reason. On the other hand, those who argue that linearisation into distinct acoustic events is a consequence of the obligation to pronounce simultaneously uninterpretable feature combinations in turn need to propose a specific featural content for monosegments that creates such a conflict. Hence the presence of [+strident] in affricates in Clements' (Reference Clements, Fujimura, Joseph and Palek1999) view, which causes linearisation since it cannot be pronounced together with [−cont]. Finally, considerations regarding the general feature-geometric setup may also determine the internal structure of monosegments. For example, Rosenthall (Reference Rosenthall, Dressler, Luschützky, Pfeiffer and Rennison1992: 250) rejects Sagey's above-mentioned ban on branching class nodes because prenasalised consonants need both a [+nas] and a [−nas] specification (he does not mention the monovalent way out that is commonly chosen, though).

Steriade (Reference Steriade, Cole and Kisseberth1994) makes a point based on the clustering possibilities of stops vs. non-stops. She first asks ‘why is it that the only reliably attested segmental contours – for nasality and continuancy – are found among released plosives?’ (page 210) and then argues that this is because released plosives have internal structure, i.e. a closure and a release position. Against this background, she explains that the curtailed clustering possibilities of non-stops is due to their lack of the release position: they are mono-positional and therefore cannot accommodate extra structure related to nasality or continuancy: ‘plosives have more clustering possibilities than continuants, because plosives are bipositional’ (Steriade Reference Steriade, Cole and Kisseberth1994: 203). The increased clustering possibilities of released stops thus motivate a bipositional internal structure of complex monosegmental items (see also Steriade Reference Steriade, Huffman and Krakow1993: 406).

Finally, van de Weijer (Reference van de Weijer1993: 94; Reference van de Weijer1996: 138f.) argues that in affricates, the continuant part has a tighter relationship with the place of articulation of the whole than the stop component, and that this should be represented in the feature geometric tree. His proposal is based on the ‘observation that affricates are typically found at place of articulation where fricatives are also found’ (van de Weijer Reference van de Weijer1993: 94).

4.1 Complex sounds: Genesis and general properties

In order to understand what a complex sound is, we must first understand what a ‘sound’ is in Duanmu's system: the author only talks about sounds, not about segments. On pages 14–17, two criteria are established that an item needs to meet if it wants to qualify as a sound: it must be independently pronounceable (hence a feature alone cannot be a sound), and it must be sequenced in time: simultaneously pronounced items cannot be part of different sounds.

Sounds are thus the minimal units that have linear status. On page 25, the author introduces complex sounds in the following manner: ‘[w]e can think of a complex sound as the merger of two (or more) sounds’. This indicates that complex sounds are absent from lexical recordings in Duamnu's system: the lexicon is a linearly ordered sequence of sounds. Unfortunately, though, and unlike other authors (e.g. Hirst Reference Hirst, Guéron, Obenauer and Pollock1985, Steriade Reference Steriade, Cole and Kisseberth1994; see Section 3 above), Duanmu is not explicit on the issue, and the reader cannot be sure that this conclusion is correct. All sources that I have come across agree that sometimes complex segments are complex already in the lexicon, and there is also evidence for contrasting bisegmental CX vs. monosegmental C^X in the same language (e.g. Lowenstamm Reference Lowenstamm and Ploch2003; see Section 3.2 above). To the extent that Duanmu really implies that all complex segments are non-complex in the lexicon, we can note that he stands alone with this position.

In Duanmu's system, lexically adjacent sounds may thus combine into complex sounds (that take up the place of one x-slot in (1) above) by virtue of some computational action (which is not further specified) that occurs prior to the construction of syllabic constituents. This is parallel to Steriade's (Reference Steriade, Cole and Kisseberth1994) merger operation (but no reference to this article is made).

Like a number of other authors, Duanmu argues that muta cum liquida and other types of what is traditionally held to be branching onsets may be monosegmental.Footnote ³ As far as I can see, though, nobody has ever proposed that sequences which do not qualify for tautosyllabicity (according to regular sonority sequencing), i.e. strings of falling or equal sonority, may also be monosegmental (e.g. Steriade Reference Steriade, Cole and Kisseberth1994: 215 strictly upholds the tautosyllabic condition for monosegmenthood). This is because whenever there is a diagnostic for closed syllables in a language, the strings at hand are heterosyllabic. Pace the literature and the evidence, though, Duanmu is also prepared to count sequences like lp as monosegmental. Issues raised by granting this option are discussed in the following section.

In sum, Duanmu's fusion operation is entirely unrestricted: it can in principle combine any number and any kind of adjacent sounds into a complex sound (25–31) as long as Duanmu's (Reference Duanmu1994) No Contour Principle is satisfied (which is discussed in Section 4.3 below).

4.2 Linearity is purely phonetic, hence lexical linearity is neutralised (/lp/ and /pl/=[pl])

We have seen in Section 3.1 that Sagey (Reference Sagey1986) distinguishes contour and complex segments according to linearisation: while the former have phonological ordering, the latter are only linearised during phonetic interpretation. Recall that according to Sagey (Reference Sagey1986: 51ff.), contour segments include affricates, prenasalised consonants and contour tones, while sequences of stops with distinct place of articulation like kp, pt or tk^w instantiate complex segments (Sagey Reference Sagey1986: 57f.). Hualde (Reference Hualde1988), Lombardi (Reference Lombardi1990) and Clements (Reference Clements, Fujimura, Joseph and Palek1999) are among those authors who deny the existence of any phonological ordering of the components of affricates, hence making this type of item also a complex segment (in Sagey's classification).

As far as I can see, though, there is no attempt to interpret the other prominent contour segment, prenasalised consonants, as a phonologically unordered structure. There is good reason for the absence of unordered interpretations. Lombardi (Reference Lombardi1990: 377) explains that linearisation of affricates occurs post-phonologically for universal phonetic reasons: there are no mirror-affricates where the first component articulated is a fricative, followed by a stop. The same cannot be said for prenasalised consonants: their mirror image, postnasalised consonants, exist as well. Hence linearisation cannot be decided by some universal phonetic mechanism; rather, it needs to be specified in the phonological representation. Steriade (Reference Steriade, Cole and Kisseberth1994: 206) (who upholds phonological ordering also for affricates) is explicit on this. In her system where released stops are made of two internal positions, closure (C) and Release (R) (in this linear order), [nas] linked to both C and R produces a regular nasal (m, n), [nas] associated only to C results in a prenasalised stop (^mb), while [nas] linked only to R is realised as a postnasalised stop (b^m).

Without discussing or mentioning these issues and the literature at hand, Duanmu takes on the position that all complex sounds are unordered in the phonology.Footnote ⁴ Following Articulatory Phonology (Browman & Goldstein Reference Browman and Goldstein1989), he thinks of complex sounds in terms of gestural overlap: ‘all features in a complex sound are simultaneous, in the sense that no sequential timing difference can be used distinctively’ (26). He also makes clear that linearisation is due to purely phonetic and universal properties: ‘the articulatory gestures for an affricate are made simultaneously, but the phonetic results are sequential, owing to the nature of articulatory aerodynamics’ (22). Hence Duanmu is not among those who hold that linearisation is done in the phonetics, but provoked by the cohabitation of a set of features that are uninterpretable simultaneously (e.g. Hirst Reference Hirst, Guéron, Obenauer and Pollock1985; Clements Reference Clements, Fujimura, Joseph and Palek1999, see Section 3.1).

Given this context, here is the list of sequences that are analysed as complex sounds in the book: labialised velars k^w (25), labio-velars (27), a vowel and a glottal stop V^ʔ (31), obstruent–sonorant clusters (branching onsets in traditional analysis) p^l, t^r, etc. (26, 50), but also the reverse items ^lp, ^rt (50, which Duanmu says are rare, on page 26). Beyond things that appear as a linear sequence of two items on the surface, complex sounds may also involve three surface items, or only one. On Duanmu's analysis, English r is in fact r^w in syllable-initial position, and when involved in a branching onset a triple complex segment is produced as in try: t^rwy (29). On the other hand, the simplex surface ʃ that may appear at the word boundary of this year is also a complex sound, made of s+yod (26), and nasal vowels may be an underlying VN sequence (30–31).

A consequence of Duanmu's claim that the components of complex monosegments are never ordered in the phonology is that lexical linearity is neutralised. Hence Duanmu makes the following point (about English): ‘why is there no lpum? The word ought to fit into CVC, if [l] and [p] can form a complex sound. The answer is that all articulatory features in a complex sound are simultaneous, so that [p]+[l] and [l]+[p] are identical. Therefore, lpum would be the same as plum’ (50, emphasis in original).

This reasoning does not go through, though. In order to see that, let us first consider the situation in languages where word-initial #RT clusters do occur: many Slavic languages and occidental varieties of Modern Arabic are cases in point (Scheer Reference Scheer, Compton, Goledzinowska and Savchenko2007).Footnote ⁵ In Czech, for example, #kl- and #lk- not only both occur word-initially, but also form ‘minimal pairs’: compare klát ‘to cut wood’ with lkát ‘to whine, to moan’. Following Duanmu's standards, one concludes that in such a language the R of word-initial #RT clusters cannot belong to the same syllable as the first vowel: it cannot form a complex sound with the obstruent since linearity in complex sounds is neutralised (/kl/ and /lk/ would both come out as [kl]). The only option for #RT clusters in Czech, then, is to be bisegmental: the #R remains unsyllabified. Due to anti-allomorphy (see Section 5.1), stray (or in traditional terms: extrasyllabic) consonants are legal as long as there are V-final prefixes (for extra #C) or V-initial suffixes (for extra C#). Unsurprisingly, Czech does have V-final prefixes, hence lkát can happily (or rather: must) identify as l.kát where the lateral remains unsyllabified. This is confirmed by Duanmu's analysis of word-initial clusters in Jiarong (Tibeto-Burman, Chapter 11), where large word-initial clusters occur that include #RT. Items such as #rpV, for example, are analysed as r.pV on the grounds of the aforementioned potential prefix-final vowel (226, 230).

The question, then, is why the initial #RT cluster of the putative English word /lpum/ could not be bisegmental along the same lines: English has V-final prefixes, hence an analysis in terms of an unsyllabified word-initial liquid, l.pum, is an option. Behind this issue is the question what it is that makes some languages tolerate #RT clusters, while others refuse to accommodate them. Recall that Duanmu's explicit answer for the absence of English #lp is that the only option would be a complex sound, which, however, would be pronounced #[pl]. This means that for some reason, which Duanmu does not discuss, the bisegmental (and heterosyllabic) option for #lp is barred in English. One guesses, then, that it is this prohibition for word-initial #RT clusters to acquire bisegmental status that drives the parametric variation at hand: #TR-only languages bar bisegmental status for #RTs, while languages where #RTs occur on the surface allow for this syllabification.

But in any case, the reason that Duanmu gives (50) for the absence of lpum in English cannot be the real reason, according to his own standards.

4.3 The No Contour Principle and the exclusion of prenasalised consonants from the class of complex sounds

It was mentioned in Section 3.3 that the study of complex segments has yielded consequences for their featural makeup, and for feature theory in general. Namely the impossibility for a segment without internal linear structure to accommodate the plus and the minus value of the same feature was the motor for introducing privative features: instead of being [+cont] and [−cont], an affricate in Lombardi's (Reference Lombardi1990: 378) analysis bears the monovalent features [stop] and [continuant]. The same insight was already made by Sagey (1986: 65): ‘the absence of an articulation is never specified’ (Sagey's solution were monovalent class (place) nodes, though). Steriade (Reference Steriade, Huffman and Krakow1993: 406f.; Reference Steriade, Cole and Kisseberth1994: 210ff.) is also along these lines.

In a sense, Duanmu's (Reference Duanmu1994) No Contour Principle (NCP), introduced on pages 25–31, is an implementation of this insight – but again the book offers no discussion of the relevant literature, or of eventual differences with previous formalisations.

Instead of prohibiting the presence of the same feature (with opposite values) within a segment (that lacks internal linear structure), Duanmu argues for a weaker version where several instances of the same feature are allowed within a (complex) sound – but cannot co-occur under the same articulator: ‘an articulator cannot take the same feature (F) twice within one sound’ (26). Articulators in Duanmu's system are similar to non-terminal nodes in a regular feature geometric tree: the six articulators assumed are Vocal-cords, Soft-palate, Tongue root, Dorsal, Coronal and Labial.Footnote ⁶

Duanmu's example to show what the NCP rules out is the merger of k and m. ‘Since [k] is [−nasal] and [m] is [+nasal], there is a conflict’ (27): the complex sound must be either entirely nasal, or entirely non-nasal. Duanmu argues that k^m (k before m), ^mk (m before k) and km (k and m simultaneous) are ruled out by the NCP since they involve simultaneous [+nasal] and [−nasal] action of the articulator Soft-palate. Were [+nasal] and [−nasal] realised by two distinct articulators, though, the result would be a viable complex sound. For example, if [+nasal] depended on the Labial articulator, and [−nasal] were dominated by Soft-palate, k and m could engage into a legal complex sound. The fact that km is declared illegal by Duanmu thus allows the reader to conclude that the feature [nasal] cannot depend on any other articulator than Soft-palate. This is never discussed or made explicit in the book, but it makes physiological sense since the velum is located in the area of the soft palate.

(Complex) sounds that bear two instances of the same feature are thus legal when the instances at hand either occur under two distinct articulators, or have the same value. Examples are k^w ([+stop] is realised by Dorsal and [−stop] by Labial (25)), kp ([+stop] is realised by Dorsal and Labial (25)) and ʃ ([+fricative] is made by both Coronal and Dorsal (25)).

Thus, given the NCP, affricates could not be complex sounds when standard features are assumed, since they have only one place articulator: [+cont] and [−cont] would be realised by the same articulator, Coronal, for example, in the case of . It was mentioned at the outset of this section that this is exactly the issue that Sagey (Reference Sagey1986), Lombardi (Reference Lombardi1990) and other authors were struggling with: their solution was to introduce two distinct features for the description of the stop and fricative components of affricates, [stop] and [continuant], and to make them privative. One senses that this is also what makes Duanmu adopt non-standard features: , for example, is made of the Coronal articulator, which dominates [+stop] and [+fricative]. But again there is no discussion in the book where these choices would be explained, and the reader is not referred to any literature either, despite the obvious precedent.

Finally, a direct consequence of Duanmu's feature-geometric setup, and also of the fact that linearity is banned from complex sounds (see Section 4.2 above), is that prenasalised consonants, an archetypical complex segment in the literature, are excluded from the set of possible complex sounds: they would involve [+nasal] and [−nasal] under Soft-palate. This is quite dramatic a departure from several decades of study of prenasalised consonants and the massive evidence that speaks in favour of their status as complex segments (see Section 3.2 above). Duanmu, however, takes this step in one sentence on page 26 (‘there is no compelling evidence that pre-nasalized stops … exist as single sounds, as I have argued in Duanmu (Reference Duanmu1994)’), without discussing any piece of evidence, without providing any argument (beyond the fact that his own theory excludes prenasalised consonants), and without informing the reader that there is a plethoric literature, and massive evidence, arguing in favour of a complex segment status for prenasalised consonants. The only literature item quoted is his own 1994 article.

4.4 When exactly do neighbouring sounds combine into a complex sound?

The previous sections have shown what a possible complex sound is, and how it comes into being. Let us now consider the question when exactly the potential of a well-formed complex sound is actuated, i.e. when exactly mergeable items merge. Recall that in Duanmu's system the lexicon is made of a linear string of simplex sounds, and that complex sounds are the result of a merger operation. This places an important burden on computation that is usually supported by static inventories. That is, the fact that a language does not have a given segment, say, English the affricate , becomes an issue of undergeneration in Duanmu's system: if is a possible complex sound, and if t and s (or s and t since linear order does not matter) may be adjacent in the lexicon, why are they never merged? The author addresses this issue on the occasion of his discussion of tl, dl (30), whose absence in word-initial position in English is as suspect as the absence of . Duanmu's answer is that languages do not have to use all opportunities to make complex sounds. In other words, the absence of in English is an accidental gap, as much as the absence of word-initial tl, dl (‘English does not need to use all possible complex sounds’ (30)).

While in the case of the statement is unsurprising (in fact the equivalent of the regular inventory-based view), it is incorrect for tl, dl. A trivial and well-known fact about word-initial tl, dl is that they are pervasively absent in many languages, and that equivalent random gaps of, say, #pl, #bl or #kr, #gr, etc. are not on record. Even if nobody has a good explanation why #tl, #dl are missing, it is not the case that their absence is accidental. This is confirmed by the observation that those languages where #tl, #dl do occur also accommodate word-initial clusters that violate sonority sequencing (i.e. #RT, #RR and #TT clusters). For example, this is the case in many Slavic languages (Russian, Czech, Polish, etc., see e.g. Scheer Reference Scheer, Compton, Goledzinowska and Savchenko2007), and in occidental varieties of Modern Arabic (see e.g. Kaye Reference Kaye1990; Scheer Reference Scheer2004: 462–465 on Moroccan Arabic). A reasonable conclusion is that tl, dl are unable to form branching onsets: when they occur, they represent coda–onset sequences, and this is the reason why they cannot exist in word-initial position in languages where initial clusters are restricted to tautosyllabic #TR. This is confirmed by their word-internal behaviour in English: in relevant American varieties, the t of atlas, Atlantic, etc. is glottalised rather than aspirated (e.g. Kenstowicz Reference Kenstowicz1994: 251). However, glottalisation occurs only in codas.

In his discussion of the absence of in English (30), Duanmu exposes his general attitude towards overgeneration: there is nothing wrong with it. Complex segments that are well-formed and could occur but do not represent accidental gaps, and there is nothing more to say (this position is also defended by Hale & Reiss Reference Hale and Reiss2008). It should be noted, however, that this not only applies to overgeneration within a given system, but also to ‘universal’ overgeneration: Duanmu's system may well derive well-formed complex sounds that are absent from the record of known natural languages. On page 50, he concludes that ‘[i]n other words, any stop can combine with any fricative to make an affricate’. Hence affricates as outlandish as ⌢, , , and the like are held to be just as regular, grammatically speaking, as and .

In any event and whatever the analysis of #tl, #dl, Duanmu needs to establish a language-specific list of those mergers producing complex sounds that do, and those that do not, occur. English, for example, must contain a statement that disallows t and s (or s and t) to merge into . There must also be some equivalent statement in the traditional perspective where the absence of segments is not due to computation, but to lexical inventories. So far so good, but further questions arise. In those cases where the merger of a given pair of sounds into a complex sound is allowed in a particular language, does this merger always occur when it can? Or, in other words, is any relevant input affected by the computation that creates complex sounds, or is the application of this computation selective? And if so, according to which criteria does it selectively apply?

These are questions that the static lexical alternative does not have to answer. Recall the oft-quoted Polish example from Section 3.2 above that illustrates situations where the same sequence of surface items contrasts according to whether they are mono- or bisegmental: czy [ɨ] ‘whether’ vs. trzy [tʃɨ] ‘three’. On the traditional view, infants acquire the sound inventory of their language during first language acquisition. The inventory is then fixed, stored and disconnected from actual phonological computation that relates input to output forms. In such a system where (in the Polish case) , t and ʃ are three distinct items of the inventory, the fact that a surface sequence XY is monosegmental in some words, but bisegmental in others, is a simple case of lexical accident, as is the occurrence of any other item of the inventory.

As far as I can see, this is not any different in more recent approaches that attempt to unify the definition of (static) inventories and (dynamic) input–output computation. For example, Steriade (Reference Steriade and Lacy2007: 145f.) argues that ‘it is … inaccurate to say that constraints on alphabets … only apply to define the underlying inventory: some of these constraints are persistent (Myers Reference Myers1991) and prohibit the same feature combinations throughout much or all of the derivation’. The unification of contrast and computation also lies at the heart of Clements (Reference Clements2003, Reference Clements, Raimy and Cairns2009) and Dresher (Reference Dresher2009), and the idea is developed in Government Phonology under the heading of Licensing Constraints (Charette & Göksel Reference Charette and Göksel1994; Kaye Reference Kaye and Dziubalska-Kołaczyk2001; Scheer Reference Scheer2010b: 2530ff.). In such an environment, the unique set of instructions (constraints) defines what a possible computational event is, and what a possible item of the static inventory is. But as in the traditional approach, the distribution of items of the inventory over lexical recordings is accidental: in Polish, , t and ʃ will still be three distinct lexical items whose distribution is unpredictable.

Hence the modern take on inventories, like its predecessor, does not need to make any specific statement regarding the question of what makes XY and X^Y distinct: they are distinct lexically. In Duanmu's system, where complex sounds are banned from the lexicon, however, some additional statement needs to be made in the lexicon in order to differentiate what will appear on the surface as a mono- vs. a bisegmental item. Again, the author does not address this critical issue directly: Duanmu does not mention the literature which provides for a merger mechanism that creates complex segments, but also allows for complex segments to occur in underlying forms. And there is no discussion of well-known cases of the Polish kind: the reader is left guessing what type of lexical diacritic Duanmu could use in order to make surface XY distinct from surface X^Y.

The book only provides an indirect indication: Duanmu is explicit on the fact that the merger computation may be selective. One thing that it depends on is speech rate: ‘VNC can be pronounced as three sounds, if one speaks carefully, or it can also be pronounced as two sounds, where VN merges into one sound’ (30–31); ‘[Vʔ]→[V^ʔ] is found in Shanghai [Chinese], where [Vʔ] can be pronounced as one sound [V^ʔ] if one speaks fast, or as two sounds if one speaks carefully’ (31). This speech-rate–based merger, however, is not applicable to other cases of complex sounds: the monosegmental status of affricates such as in German is not negotiable. Here the author will have to say that the merger always applies, no matter what the speech rate. We are thus left not only with language-specific merger operations, but also with a distinction between those operations that are speech-rate–dependent, and others that are absolute. But even within the latter, this cannot be the full story since, recall, linearity does not matter: an underlying sequence /ts/ should be merged into a complex sound and appear as [] as much as its mirror image /st/. In the same way, in a language where the transformation of /VN/ into the complex sound occurs depending on speech rate, the same surface result should be achieved with an /NV/ input. This is obviously not the case: German has plenty of st clusters, which are pronounced [st], rather than [], and no French NV sequence ever produces a nasal vowel on the surface.

But one senses from the discussion on page 50 that there is more to the question under which conditions exactly two mergeable items merge. Duanmu writes that ‘[o]ne might also ask why help is not realised as [hεp^l], where [l] and [p] form a complex sound [p^l] in the coda. The answer is that [p] is already supported by “anti-allomorphy” … , and so there is no need to incorporate it into a syllable’ (50). Hence merger turns out to be a last resort strategy: it is only when there is no other means to accommodate a sound that it tries to hook onto its neighbour. On the same page, the merging operation also loses against another competitor, sonorant syllabicity: ‘one might ask why there is no medial coda [p^l] or [k^l], which could result from a hypothetical word neplsa or tiklny. A possible answer is that, because [l] can be syllabic, neplsa or tiklny can be trisyllabic’ (50). Duanmu does not formalise this competition in any way, though.

We may thus note that at least some merger operations are in competition with other options, and that the driving force behind this competition is the accommodation (and ultimately pronunciation) of consonants. All this is just conjecture since the author is not explicit on the existence of competition, or on the way this competition is managed (constraint interaction, etc.; the absence of computational aspects of phonology in the book is discussed at greater length in Section 6 below).

Finally, it appears that merger is a last resort operation only in some cases. Compare the lp in help with affricates. The latter always merge no matter what: were their merger also last resort only, a word such as butcher would come out with independent t and ʃ where t closes the root syllable: but.ʃer. This is because the t can happily be accommodated in the preceding syllable, so there is no ‘need’ to engage into a merger operation. This is not the case, though: is an affricate also in butcher.

The overall picture thus requires the following variation for the merger operation: (i) statements regarding which items are allowed to merge in a given language; (ii) statements regarding which merger operations are subject to speech rate variation; (iii) statements regarding which merger operations are mandatory (e.g. affricates); and (iv) statements regarding which merger operations are in competition with other syllabic solutions, and how this competition is decided. This is probably all technically workable, but one wonders what the import is of creating this kind of machinery, which is only a direct consequence of the idea that complex sounds are absent from the lexicon. It was mentioned earlier that this ban is entirely unmotivated in the book, that the author stands alone with it, and that arguments in favour of the existence of complex segments in the lexicon that are made in the literature are not discussed. As it stands, Duanmu's system thus adds to the computational burden of the grammar, and must introduce some lexical diacritics in order to describe contrasts of the Polish vs. tʃ kind. It is hard to believe that this is the right way to go in absence of supporting arguments, and in presence of an alternative that does not need any of the extra machinery.

4.5 Stress-based syllabification?

Duanmu proposes that syllabification is stress-based. In Section 4.5 of the book, he argues against the standard Maximal Onset Principle (MOP), ‘according to which consonants between vowels should be syllabified as the onset of the following vowel as far as possible’ (56). The reader does not learn, however, how it is decided what is and what is not possible in the standard approach; and – in a book called Syllable Structure – sonority sequencing is not mentioned at all.

Duanmu argues that there is no real evidence for the MOP, since English aspiration and flapping can be accounted for by different means (a rare occasion where evidence from phonological processes is considered). The alternative builds on what the author calls the Weight-Stress Principle (WSP): stressed rhymes should be heavy, while unstressed rhymes should be light (59). Syllabification, then, is a function of stress: át.om vs. a.tóm.ic. On these grounds, p, t, k are aspirated word- and foot-initially (a.t ^hómic), and t, d are flapped when they occur elsewhere (á[].om). This mechanism appears to be circular: one wonders how stress is distributed, since the stress-assigning algorithm depends on syllable structure, which, however, does not exist by the time stress is assigned, precisely because it is built on stress.

But beyond this issue, Duanmu seems to really propose stress-based syllabification not as a mere alternative analysis for English, but as a general principle that drives syllabification in all languages, given that this proposal is made in the general part of the book where CVX theory is introduced as such, not in one of the language-specific chapters. In other words, the author intends to refute the MOP by showing that there is no evidence for it in English. He thus leaves unmentioned all those languages where stress does not play any role in segmental phenomena, and where syllabification is a strict function of the MOP. In (Southern varieties of) French for instance, mid vowels are tense if and only if they occur before TRV as in m[e]trique ‘metrical’, but lax before any other cluster, e.g. p[ε]rdu ‘lost’).Footnote ⁷ Therefore, if Duanmu's analysis of English is correct, the MOP is not operative in this language – but this does not tell us anything about other languages, let alone about the validity of the MOP in syllable theory.

That Duanmu really means stress-based syllabification to be universal is confirmed on page 236, where three core properties of CVX theory are listed: one of them says that ‘[s]yllabification is determined by the Weight-Stress Principle, according to which the rhyme of a stressed syllable is VX and that of an unstressed syllable is V’. Even if all languages were like English, one wonders what systems with fixed initial (like Czech) or final (like French) stress would look like: heavy syllables could only occur at the left or at the right edge, respectively – a description that does not match the structure of the languages at hand.

5.1 The potential V analysis

Let us now turn to the second mechanism used by Duanmu in order to account for consonants that appear not to fit the CVX template: anti-allomorphy. Anti-allomorphy identifies as the following instruction: ‘[k]eep a morpheme in the same shape regardless of the environment’ (47). Duanmu thus entertains a version of what, since the 19th century, has been known as analogy and since then variously, as quoted by Lass (Reference Lass1980: 71), as ‘the principle of one-meaning-one-form’ (Anttila Reference Anttila1972: 92ff.), ‘form–meaning biuniqueness’ (Ohlander Reference Ohlander1976) and allomorphy minimisation within a paradigm (Kiparsky Reference Kiparsky and Dingwall1971). More recently, the idea has been present under the headings paradigm uniformity, output–output correspondence (Kenstowicz Reference Kenstowicz, Durand and Laks1996; Burzio Reference Burzio, Durand and Laks1996; Benua Reference Benua, Beckman, Urbanczyk and Dickey1995, Reference Benua1997 and following) and Optimal Paradigms (McCarthy Reference McCarthy and Downing2005). As elsewhere, though, Duanmu does not provide discussion of the background: the only reference mentioned is Burzio (Reference Burzio, Durand and Laks1996) (46). While nobody doubts the reality of analogy, it is often hard to tell patterns that are due to analogical activity from those that are the result of regular (synchronic or diachronic) phonological processes (see e.g. the historical overview in Downing, Hall & Raffelsiefen Reference Downing, Alan Hall, Raffelsiefen and Downing2005a: 2ff.).

Since neogrammarian times, analogy and phonological processes are viewed as the two basic forces that shape sound patterns by counteracting each other: the former restores regularity by levelling out the irregularity introduced by the latter into paradigms of morphologically related words (e.g. Jespersen Reference Jespersen1976 [1922]: 92f., 277f.; Anttila Reference Anttila1972: 94ff.). Viewed from the other end, this means that exceptions to regular sound laws or regular phonological computation may be due to the normalizing action of paradigms, which compares surface forms of the different cells and tends towards invariable morpheme shape. This is what Duanmu calls anti-allomorphy (a term coined by Burzio Reference Burzio, Durand and Laks1996: 130ff.): analogical action prevents morphemes, in his case roots, as we will see, from appearing in different shapes on the surface.

For example, Duanmu argues that anti-allomorphy is the reason why the -p is pronounced in help although it is not integrated into any syllable (50, and this is meant literally: not at the underlying level, at no derivational stage and not on the surface either). Indeed, the -p in help cannot be syllabified in CVX since it is supernumerary with respect to the CVX template, and -lp cannot be a complex sound (it would be pronounced [pl] since linearity does not matter within complex sounds). This is what Duanmu means when he says on several occasions in the book (e.g. page 3) that supernumerary consonants at word edges ‘are explained by morphology’.

Since anti-allomorphy is an interaction of two morphologically related forms which produces the non-variation of a given morpheme, there must be some version of the morpheme help where the -p appears rightfully, i.e. as a duly syllabified item. This condition also precludes the existence of monsters with 24 unsyllabified consonants anywhere, including in the middle of morphemes. That such monsters do not exist, and that supernumerary consonants cannot occur morpheme-internally, is assured by what Duanmu calls the potential V analysis (46). That is, word-final extra consonants can only survive if there are V-initial suffixes that can ‘support’ them at least in some grammatical forms. Hence the -p of help can survive by way of anti-allomorphy only because there are V-initial suffixes such as in help-ing where the -p is properly syllabified in the onset of -ing. Were there no V-initial suffixes (or, one supposes, if none of them attached to the specific root help), the -p could not survive in help. The same applies to the left edge of the word: supernumerary initial consonants can only survive if the language has V-final prefixes.

5.2 Is the syllable inventory of a language predictable from morphology?

The potential V analysis thus states that supernumerary consonants at word edges are only legal if V-final prefixes and/or V-initial suffixes occur. In Duanmu's approach, though, the correlation also holds true in the other direction: the presence of a certain affix type predicts the occurrence of edge clusters. Hence, ‘unlike previous analysis, which attributes the variations [of edge clusters] to syllable parameters, the CVX theory derives the variations from morphology’ (51).

The author then provides a typology that compares his morphological with the regular phonological account of cross-linguistic variation of the size of monosyllables. For example, the regular double parameter setting ‘no branching codas and no branching onsets’ describes a language without extra consonants at either edge; with Duanmu's instruments such a language is characterised as ‘no V-final prefixes, no V-initial suffixes’. Another equivalence, describing a language that allows for supernumerary consonants at the left edge, is ‘having branching onsets’, which in Duanmu's system corresponds to ‘having V-final prefixes’ (51).

These equivalences do not stand up to fact, and they are also inconsistent with the workings of complex sounds. For example, the last equivalence mentioned above is wrong since languages that allow an extra left-edge consonant – i.e. one that is not syllabified – may well have no branching onsets at all: this is the case in Moroccan Arabic, for example, where both #TR and #RT (as well as #RR and #TT) occur (e.g. brɨd ‘to freeze’, rbɨT ‘to tie’) but TR clusters always behave as heterosyllabic sequences (the preceding vowel shows closed syllable effects, see Kaye Reference Kaye1990; Scheer Reference Scheer2004: 462–465). That is, word-initial #RTV clusters must be bisegmental and heterosyllabic (r.bɨT): this conclusion was already reached in Section 4.2 above for Czech and Jiarong. In sum, thus, Moroccan Arabic needs V-final prefixes in order to support supernumerary left-edge consonants, but has no branching onsets – while Duanmu holds that the presence of the former implies the latter.

The same equivalence is also inconsistent with what we know of complex sounds since it treats word-initial branching onsets as clusters whose first member is supernumerary and requires a V-final prefix in order to exist. However, earlier and later in the book TR clusters are said to be complex segments (e.g. in Chapter 8, where all English #TR clusters are analysed like this), which match one single x-slot and hence do not produce any supernumerary consonant. Therefore it is not clear why languages like English where TR clusters are complex segments should be in need of V-final prefixes.

The impact of the preceding needs to be measured in the light of the fact that the typology discussed turns out to be the heart of Duanmu's theory: it is designed to bring home the ambition stated in the subtitle of the book (‘the limits of variation’), and the author aims to show that it matches the empirical record better than regular parameters on syllabic constituency. The claim behind this is that phonologists have missed a cross-linguistic generalisation: ‘[t]he main claim of the CVX theory is that there is a correlation between the morphology of a language and the maximal size of monosyllabic words’ (52). Given the importance that the author grants to this claim, as well as its singularity in linguistic thinking, it may come as a surprise that Duanmu does not make any attempt to support his typological generalisations by cross-linguistic evidence, other empirical material or some case study.

In absence of supporting evidence and given the conflicting record, the reader will have to conclude that there is no such correlation between edge clusters and the occurrence of this or that affix type. That is, the presence of supernumerary consonants (to the extent that one agrees on exactly which consonants are concerned) at word edges is not predictable from the inventory of affixes.

This being said, Duanmu's claim is explicit and excludes certain configurations (still on page 51). Therefore its conditions of falsification can be easily stated. What should not exist if Duanmu is right are languages which have the same set of affixes, but different syllable structure. As far as I can see, though, there are plenty of cases of this kind on record: for example, languages that are closely related genetically and therefore possess the same affix inventory, but follow different syllabic patterns. Polish and Belarusian is such a pair: on the basis of the same set of affixes (in any case V-final prefixes and V-initial suffixes are shared by both languages), the former shows all possible sonority slopes in word-initial clusters (#TR, #RT, #RR, #TT; see Scheer Reference Scheer, Compton, Goledzinowska and Savchenko2007), while the latter is a #TR-only language like English or Spanish, where initial clusters strictly obey sonority sequencing. The presence of V-final suffixes in Belarusian, however, should ‘support’ stray word-initial consonants.

One may assume that Duanmu's answer is along his general line ‘there is nothing wrong with overgeneration’ (recall Section 4.4 above): Belarusian grammar does not object to word-initial RT clusters, but for some reason the language does not exploit this potential. For one thing, this argumentation would empty Duanmu's claim of any empirical content since there would be no way anymore to know whether a language should be counted as having branching onsets (or any other syllabic configuration for that matter) or not: a language that is strictly CVCV on the surface could be argued to actually have the potential of forming branching onsets, but for some unknown reason does not actuate this option. In fact all languages could be argued to have the highest syllabic complexity since one never knows whether the non-existing patterns are just accidental gaps.

But in the Belarusian case, even this would not work since there is positive evidence to the end that the language does not tolerate word-initial stray consonants: when a CVC root occurs in zero grade (because a V-initial suffix triggers a vowel-zero alternation on the root vowel, a typical pattern in Slavic), an i-prothesis appears. Compare, for example, lev ‘lion NOMsg’ with i-lv-a ‘lion GENsg’ (Scheer Reference Scheer, Dočekal and Ziková2009a). If the presence of V-final prefixes guaranteed the viability of the word-initial #lv cluster, the genitive singular could happily be pronounced lv-a (as is the case in Polish: lew – lw-a ‘id.’).

Another objection against the existence of a correlation between the phonological properties of affixes and syllable complexity is the simple question how many languages there are where V-final prefixes or V-initial suffixes are absent. This being the critical predictor of Duanmu's system, what kind of cross-linguistic variation do we expect if all languages are the same regarding the conditioning factor? Again, one would expect the different patterns predicted (seven are listed on page 51, but the list ends with ‘etc. ‘) to be illustrated at least by one language, which means that on the affix side Duanmu would have to come up with languages that have 1) V-final prefixes and V-initial suffixes, 2) V-final prefixes but no V-initial suffixes, 3) no V-final prefixes but V-initial suffixes, 4) neither V-final prefixes nor V-initial suffixes. The existence of patterns 2, 3 and 4 is hard to imagine among natural languages.

Finally, one wonders what the predictors ‘V-final prefixes’ and ‘V-initial suffixes’ become in languages where syllabification is done across word boundaries (so-called connected speech). Languages where this is notoriously the case include Sardinian (e.g. Bolognesi Reference Bolognesi1998), Middle Italian dialects (e.g. Marotta Reference Marotta, Brandão de Carvalho, Scheer and Ségéral2008, Scheer Reference Scheer2009b), the aforementioned Belarusian (Scheer Reference Scheer, Dočekal and Ziková2009a), and to a certain extent English. In relevant dialects of the latter, t-flapping also concerns word-final t's, which undergo the process iff the following word is V-initial. Compare, for example, hit[] Ann with hit*[] Mary (e.g. Nespor & Vogel Reference Nespor and Vogel1986: 46f., 224f.). We know independently that flapping occurs only in onsets (otherwise the t is glottalised, see Kenstowicz Reference Kenstowicz1994: 251), which means that the t in the external sandhi cases mentioned must sit in the onset of the following word. Now in such a language affixes should not play any role since supernumerary edge consonants can always be supported by a preceding V-final or a following V-initial word. The prediction, then, should be that there is no variation in syllable structure among these languages at all. This is not the case, though. External sandhi languages are not reported to have any particular syllabic typology: their choices regarding syllabic complexity go over the same range of possibilities as those of languages where the domain of syllabification is strictly bound by the word.

5.3 Law of Finals: If a language has internal codas, it must also have word-final consonants

The book works with another cross-linguistic generalisation that is controversial and faces a number of counter-examples, but again without inspecting well-known patterns from individual languages. Duanmu claims that there is a correlation between word-final and word-internal rhymes. He calls on Vennemann (Reference Vennemann1988) and Blevins (Reference Blevins2004) (and ultimately on Haugen 1956) in order to motivate the ‘Law of Finals (LOF): Word-final rhymes and word-medial rhymes should resemble each other’ (55) (there is also a Law of Initials LOI, which does the symmetric labour for the left edge/onsets).Footnote ⁸

The idea that edge clusters define possible (internal) onsets and possible (internal) codas is standard in the literature. But Duanmu transforms an implication (‘if at an edge X exists, then X is a permissible (internal) onset/coda’) into an equivalence whereby internal onsets/rhymes also define what kind of cluster occurs at edges. Hence if a language has word-internal codas, it should also have word-final codas (i.e. word-final consonants). By contrast, in case it does not possess word-internal codas, all words should be vowel-final.

A well-known counter-example is Italian, where all words are vowel-final, but word-internal codas regularly occur (altro ‘other’, parco ‘park’, etc.). It is true that word-final consonants occur in words that have recently joined the language (recent loans, acronyms, etc., e.g. internet ‘Internet’, see e.g. Passino Reference Passino2008, Krämer Reference Krämer2009). These may either be analysed as an instance of a specific grammar (loanword phonology, e.g. Itô & Mester Reference Itô, Mester and McCarthy2004). Alternatively, the pattern ‘internal codas but no word-final consonants’ is surface-true for earlier stages of Italian, when word-final consonants were actively eliminated, but internal codas survived: Vennemann (Reference Vennemann1988: 24) states that ‘Italian has eliminated (excepting some loan words and truncated forms) all word-final consonants, some by vocalisation as in (31) but most by loss, a few by adding a protective vowel’.

More generally, the correlation between internal codas and word-final consonants has been considered in the literature under the heading of the status of the latter (as a coda or a non-coda). Harris (Reference Harris1994: 162) writes that ‘[a]ccording to the final-coda view, the typology is exhaustively characterised in terms of a straightforward bifurcation into open-syllable “CV” systems versus closed-syllable “CVC” systems. … The facts, however, contradict this taxonomy. The internal-coda and final-consonant dimensions are actually independent of one another, as a result of which there are four rather than two major types of syllabic system’. Harris then quotes Italian and Telugu (Central Dravidian) as instances of systems with internal codas but no final consonants, and Luo (Western Nilotic, Kenya) as a representative of the reverse pattern (final consonants, but no internal codas).

5.4 Existing allomorphy: Constraint interaction?

Finally, we turn to the trivial issue, not addressed in the book, of the bare existence of countless examples of allomorphy. Were anti-allomorphy a rule without fail, morphemes should never appear in different shapes on the surface.Footnote ⁹ The reader senses that Duanmu does not mean to simply say that anti-allomorphy strikes when morphemes do not vary, and is violated when they do. This would void the mechanism of any empirical content: the observation would merely be restated.

Rather, anti-allomorphy looks like a constraint in the sense of OT, but Duanmu does not correlate it with other constraints (as the literature on paradigm uniformity would), or formalise his cross-linguistic typology in terms of constraint interaction.Footnote ¹⁰ The following section discusses the absence of computational aspects of phonology at greater length.

6.1 Evidence from phonological processes is by and large absent

A remarkable property of the book is the almost complete absence of phonological processes, something that may be called staticism.Footnote ¹¹ Duanmu's main concern is how to accommodate supernumerary consonants. Even though the behaviour of these consonants in phonological processes is a classical diagnostic for their syllabic status, the book does not take into consideration this kind of evidence. Even when the issue is pointed out by a reviewer of the book (the lateral of the word file is velarised, so it must be a coda, see the following subsection), he rest of the book does not engage with evidence from phonological processes beyond a reaction on the particular issue regarding file. Blevins (Reference Blevins2010: 290) also mentions this concern in passing: ‘[i]f Syllable structure is to be taken seriously, the same kind of phonological and experimental evidence [including thus phonological processes] needs to be presented for the CVX model’. The discussion below provides some illustration.

For example, staticism drives the section on syllable boundaries (in Chapter 3, 52–64). Duanmu considers how the beginning and the end of syllables can be determined – a classical question in syllable theory. He starts by arguing against speaker intuitions, and then writes ‘suppose there is a total lack of speaker intuition, should we conclude that there are no syllable boundaries? I suggest the answer is no’ (53). The reader may now expect that the alternative source of evidence that Duanmu proposes are phonological processes. This is not the case. Instead, the aforementioned Law of Initials (LOI) and Law of Finals (LOF) are developed. They are indeed a classical type of evidence as well (though only in their implicational guise, not in Duanmu's equivalence version, see Section 5.3). But significantly, the author does not explore phonological processes, which provide an obvious arbitral award.

Another illustration of staticism is the discussion of English word-internal monomorphemic sequences that look like VXC rhymes. On pages 44–45 and 150–154, relevant items are presented as a problem for CVX theory since they are too big to fit under VX: symptom (VNC) and council (VVN) illustrate the two patterns. The solution proposed is based on the observation that these items always involve a homorganic NC cluster. Since these clusters are produced by the same place articulator, they do not qualify for complex sound status (see Section 4.3 above). What can be a complex sound, though, is the VN sequence. Therefore Duanmu contends that ym in symptom and oun in council are complex sounds, that is nasal vowels. This indeed makes the sequences small enough to fit into CVX.

This analysis falls foul of lenition processes in many languages: whenever there is a homorganic NC cluster and a relevant diagnostic for consonantal strength, the C is strong, i.e. shielded against lenition, exactly like other consonants in post-coda position where the coda is not a nasal (e.g. Ségéral & Scheer Reference Scheer, Zybatow, Szucsich, Junghanns and Meyer2008). Duanmu is explicit on the fact that all VNC sequences where N and C are homorganic can be analysed as (‘another case of interest is VNC clusters, where the nasal N and the consonant C are homorganic (sharing the same place of articulation). In this case, VNC can be analyzed as C’, 30). The author thus predicts that in English and all other languages where VVN.CV sequences need to be shrunk because they are too big for CVX, the C of the NC cluster is an ordinary intervocalic onset: .CV. This is contrary to fact, though, whenever a lenition process provides a diagnostic for intervocalichood. In English, for example, t-lenition systems show various effects on etymological t's (making them unreleased, a flap, a glottal stop or an s, see Harris Reference Harris1994: 194ff.). In a number of varieties, t is completely preserved from any damage in post-consonantal position, no matter what the preceding consonant (chap.ter, win.ter, coun.ting, quar.ter, poul.try, eas.ter). This is unexpected for Duanmu, who predicts that t behaves like an intervocalic consonant in counting and the like. Now there are also systems on record where t is damaged in intervocalic position and after nasals (e.g. English spoken in New York: Harris & Kaye Reference Harris and Kaye1990). In these varieties lenition also occurs in words like quarter, but, contrary to Duanmu's prediction, it is not the nasal vs. non-nasal distinction of the preceding consonant that governs the lenition pattern, but the opposition sonorant vs. non-sonorant. That is, in New York t is flapped in winter and quarter, but not in after, custard, chapter and doctor. This distinction ‘post-coda weak after sonorants but strong after obstruents’ is recurrent cross-linguistically (Ségéral & Scheer Reference Scheer, Zybatow, Szucsich, Junghanns and Meyer2008).

6.2 Supernumerary consonants identified as codas by phonological processes

Another context in which the absence from the book of reference to phonological processes matters is the unsyllabified status of supernumerary edge consonants. It takes a reviewer of the book manuscript to introduce processing (on page 47 Duanmu reports that he was asked during the reviewing process whether the fact that the lateral in file is velarised does not constitute evidence in favour of its coda status). In Duanmu's system the lateral cannot be a coda because the X of CVX is already occupied by the second part of the diphthong. Therefore the lateral ought to be stray (i.e. supported by anti-allomorphy). The author accepts the evidence regarding the darkness of the lateral, but argues that the distribution of clear and dark l does not oppose onset and coda l, but rather laterals that are dominated by onsets vs. rhymes. Since in his system the lateral in file cannot be a coda, the author makes it a separate nucleus, which is dominated by a rhyme and therefore velarised. On this analysis, then, the word file has a syllabic l and is therefore bisyllabic: [f[aj]_N]_σ [[l]_N]_σ.

Beyond the particular case of English dark l (on which more below), the argument concerns a more general pattern: word-final consonants may be subject to coda effects, in which case Duanmu's stray analysis must be wrong. Indeed, processes such as l-darkening, l-vocalisation, r-vocalisation, deaspiration, final devoicing or others of the kind are not known for making any difference between those word-final consonants that, according to Duanmu, are true codas (i.e. which occur after short vowels and therefore instantiate the X of CVX), and those that remain unsyllabified and are supported by a potential V (i.e. which occur after a long vowel or a short vowel plus another consonant and therefore do not fit into CVX).

Let us examine two cases in greater detail, English l-darkening itself and German r-vocalisation. Lavoie & Cohn (Reference Lavoie and Cohn1999) study the peculiar properties of file-like words in English, i.e. items where a word-final liquid, r or l, is preceded by a long vowel or a diphthong. The intuition of the authors, which coincides with a number of pronouncing dictionaries, is that file-type words are longer than one syllable items such as beat or bead, but not quite as long as two syllable items such as bottom. They therefore call them sesquisyllabic words, i.e. words which count as one and a half syllables, and analyse them as threemoraic monosyllables (as opposed to bimoraic monosyllables for the beat/bead pattern, and a bisyllabic structure for the bottom-type items). Lavoie & Cohn support their intuition with evidence from (i) a forced-choice test with six American natives who had to decide how many syllables a written stimulus had, (ii) metrical structure (chant and verse), and (iii) the measurement of duration. The results of the latter show that the vowel of file-type words is significantly longer than the vowel of beat/bead-type words.

On the face of it, this looks like good news for Duanmu, since the extra length of file, both measured and perceived, is a fair match of his bisyllabic analysis. Unfortunately, though, this correlation breaks down when we examine exactly which words belong to the file-type in Lavoie & Cohn's data: sesquisyllabic words are those which have a word-final lateral preceded by a diphthong (aj, aw, oj like in pile, owl, foil) or a long monophthong, except ɔɔ and what the authors consider to be /o/. Hence peel, pale, pearl, pool, pile, owl and foil are extra-long monosyllables, while Paul [pɔɔl] and pole /pol/ [pəwl] are regular monosyllables. The situation with word-final r is yet different: only preceding diphthongs produce extra length (pyre, hour, foyer); all monophthongs remain regular monosyllables (peer, pear, par, poor, pour). Hence it is not the case that all l-final words preceded by a long vowel or a diphthong produce the extra length that suits Duanmu's bisyllabic analysis. This is, however, what Duanmu will have to claim since l-darkening, our diagnostic for the syllabic status of word-final l, does not care for the kind of vowel that precedes (long, short, back mid, etc.): all word-final laterals are dark (e.g. Spencer Reference Spencer1996: 214ff.). Therefore, the bisyllabic analysis that Duanmu is forced to recur to sits easily with some long-vowelled items such as file, but cannot make sense of others, such as Paul and pole.

A parallel process is r-vocalisation in German, where r is pronounced as some kind of low schwa [ɐ] in internal and final codas (e.g. Wiese Reference Wiese1996: 252ff.). The r of a stem like leer ‘empty’ surfaces in vocalised form in word-final position (leer [leeɐ] ‘empty’) and before consonants (leer-te [leeɐtə] ‘(s/he) emptied’), but appears as a uvular fricative in intervocalic position (leer-en [leeʁ-ən] ‘to empty, inf.’). This pattern is entirely insensitive to whether the stem has a long (as in leer) or a short vowel (as in verwirr-en ‘to confuse’). Here as well, Duanmu will have to resort to the rhyme-based analysis of file, which in the German case means that the [ɐ] of leer [leeɐ] sits in the nucleus of a second syllable, and hence that the item is bisyllabic. Since ɐ is a vowel, its analysis as a nucleus is certainly not outrageous, but one wonders what its status is word-internally, then, as in leer-te [leeɐtə]. And in German, the bisyllabic analysis clearly runs against the intuition of speakers, who count unanimously one single syllable for leer and the like.

6.3 Syllabic consonants

Let us now look at another critical ingredient of the bisyllabic analysis of file that Duanmu is forced into: the lateral must be syllabic, i.e. [f[aj]_N]_σ [[l]_N]_σ. The trouble is that the syllabicity of consonants – at least on regular assumptions – is connected to the way the items are pronounced: a consonant is not syllabic just because the analyst has decided that it should. It was mentioned above that speakers have intuitions about how many peaks there are in a word. Also, syllabic consonants leave traces in the phonology: they are stress-bearing units, which disqualifies, for example, the lateral in Czech lhát ‘to lie’ and Polish lgnąć ‘to cling to sth.’ from syllabic status. In Czech, stress is initial and would fall on the lateral (but does not: the á is stressed, while in blbec ‘idiot’, for example, stress falls on the lateral); in Polish, stress is penultimate and would also fall on the lateral in lgnąć, were it a possible stress-bearing unit (here as well it falls on the ą). Relevant diagnostics for the syllabicity of consonants are discussed in Scheer (Reference Scheer, Zybatow, Szucsich, Junghanns and Meyer2008). The laterals in the Czech and Polish words at hand are thus of a different kind: like their syllabic cousins they are cluster-building and ‘supernumerary’ from the point of view of ill-formed-looking clusters, but still not syllabic. In the Slavic literature they are called trapped.

Beyond the case of file, Duanmu also declares consonants syllabic on other occasions when this suits the analysis. The Tibeto-Burman language Jiarong, studied in Chapter 11, has regular #TR clusters, but also abundantly displays #RT, #RR and #TT sequences. In a first step, Duanmu's account is based on anti-allomorphy: C₁ of any #C₁C₂ is supported by prefix-final vowels and hence remains unsyllabified. Unlike in English, where all #TR clusters are interpreted as complex sounds (159-179), in this language even #TR clusters are analysed as bisegmental (although there is no apparent reason, and no discussion), and the T is thus supported by a potential V as well.

In a second step, clusters bigger than #CC that cannot be accounted for by anti-allomorphy are declared to have a syllabic consonant: this is the case for words and clusters such as zndai ‘wall’, #wrŋ-, #vrd-, etc., which Duanmu analyses as z dai, #wŋ-, #vd-. As it stands, this move certainly saves CVX from Jiarong monster clusters, but there is no evidence that the consonants which Duanmu needs to make syllabic really are. More general information about the occurrence and behaviour of syllabic consonants in Jiarong would be needed, and the application of regular diagnostics (stress, poetry, native speakers' intuitions about the number of syllables) would be required, also regarding the eventual difference between syllabic and trapped consonants.

6.4 Affix and ‘affix-like’ sounds are for free

On pages 49–50, Duanmu is forced into a further radicalisation of the allowance for stray consonants because of words like saved, risked, texts, sixths. These are even bigger than VXC where C is already supernumerary: the -d of sav-ed [sεjvd], for example, is on top of VXC – the diphthong takes up VX and the v is the extra C. While anti-allomorphy can account for the stem-final -v, it is without effect for the -d since no further V-initial suffix can ever be added. The critical observation is that all such monster clusters involve a morpheme boundary, and that thus morphology has a word to say. In Government Phonology, for example, the conclusion is that each morpheme involved sits in a cycle of its own (cycles are called domains in this theory), and that cycle-final empty nuclei may remain empty (Kaye Reference Kaye, Durand and Katamba1995). Hence sixths is made of three cycles and identifies as [[[sɪxø]θø]sø], where three domain-final empty nuclei occur in a row.

Duanmu's solution is also based on this morphology-driven insight, but simply places affix material outside of phonology altogether: the Affix Rule states that ‘[a]ffix or affix-like sounds can be pronounced, whether they can fit into a syllable or not’ (50). In other words, affixes are not concerned by phonology at all and therefore may be stray. As on other occasions, a functional explanation is advanced: if affix material were left unpronounced every time it cannot be syllabified, people would not be able to distinguish affixed from unaffixed forms, and the meaning carried by the affix would be lost (49).

Duanmu's Affix Rule (but not necessarily its functional motivation) expresses the idea known under the heading of the OT constraint Realize-Morpheme (or Realize-Affix), which, in Walker's (Reference Walker1998: 244) formulation, states that ‘[a] morpheme must have some phonological exponent in the output’. Relevant literature also includes Rose (Reference Rose1997) and Kurisu (Reference Kurisu2001) ; the book, however, does not mention the fact that the Affix Rule has a precedent.

But Duanmu goes still farther than that when he talks about perceived affixes and phonetic familiarity (49): since ‘German and English use coronal consonants as suffixes, all final coronals are perceived as suffixes’ (49, also 214).Footnote ¹² He therefore introduces a principle according to which ‘[s]ounds that resemble affixes can be treated as affixes’ (49). This is a version of what appears as aggressive suffixation in the literature (but this fact is not mentioned). Zuraw (Reference Zuraw2002: 396) reports that the term goes back to Hammond (Reference Hammond1999b), who ‘proposes that English adjectives whose final syllable resembles a suffix are stressed as though they actually were suffixed’.

Finally, according to Duanmu, phonetic familiarity can also explain the presence of extra word-final coronals in monomorphemic items: ‘because we are used to hearing coronal suffixes, other final coronals would not sound bad to the ear’ (49). The author thus contends that syllabification, at least in part, is a matter of sounding good or bad to the ear. This is probably not a trivial notion, and the reader wonders how it can be determined in practice whether a given sound, or a sequence thereof, is euphonic to the ear.