4.1 The hypotheses

All laryngeal restrictions share the property that the contrast between one and two instances of a laryngeal feature is neutralised. This observation leads to Hypothesis 1, given in (11).

1. (11)

   

The most ambitious hypothesis about the role of perception in laryngeal co-occurrence restrictions is that not only the existence of restrictions, as predicted by Hypothesis 1, but also the variation between the three types of restrictions is based in perception. While all three types of restriction neutralise the 2 vs. 1 contrast, which is claimed to be the weakest, all other laryngeal contrasts are not necessarily equal. Differences between the various types of restrictions may reflect different thresholds for the distinctness of allowable contrasts. Specifically, assimilation may be due to an additional perceptual asymmetry, namely that the contrast between two and zero instances of a laryngeal feature is stronger than the contrast between one and zero instances of a laryngeal feature. Neutralisation of the 1 vs. 0 contrast, either in all pairs of roots or only in pairs of roots with homorganic stops, may reflect the relative weakness of these contrasts relative to the 2 vs. 0 contrast. The preference for forms with multiple ejectives over roots with a single ejective may be explained by the relative perceptual distinctness of contrasts. The two relevant hypotheses are stated in (12).

1. (12)

   

Hypothesis 2 suggests a general preference for assimilation, while Hypothesis 3 aims to explain why assimilation may target only homorganic pairs of stops, as is the case in mixed restriction languages. If Hypothesis 1 is correct, the 2 vs. 1 contrast should be more difficult than the 1 vs. 0 contrast, regardless of place of articulation. The predictions of these three hypotheses are summarised in the four-tiered perceptual hierarchy in (13), where ‘<’ means ‘less distinct than’.

1. (13)

   

The hypotheses centre on claims about the relative distinctness or strength of a contrast. For some contrast category A to be ‘weaker’ than some contrast category B, subjects should have more difficulty discriminating pairs of forms that fall into category A than B (i.e. incorrectly think that pairs of different roots are the same). The experiment tests these hypotheses by presenting subjects with pairs of CVCV nonce roots that differ only in whether the consonants are ejective or plain. Subjects are then asked to decide whether the roots they hear are the same or different from one another (an AX discrimination task). If Hypothesis 1 is correct, subjects will perform better on pairs like [k'ap'i] and [kapi] (2 vs. 0), [k'api] and [kapi] (1 vs. 0 heterorganic) and [k'aki] and [kaki] (1 vs. 0 homorganic) than on pairs like [k'api] and [k'ap'i] (2 vs. 1). If Hypothesis 2 is supported, subjects should be better at accurately distinguishing pairs like [k'ap'i] and [kapi] (2 vs. 0) than pairs like [k'api] and [kapi] (1 vs. 0). Finally, if Hypothesis 3 is true, subjects will perform better on pairs like [k'api] and [kapi] (1 vs. 0 heterorganic) than on pairs like [k'aki] and [kaki] (1 vs. 0 homorganic).

4.2 Stimuli

The stimuli for this experiment are pairs of CVCV disyllables manually spliced together from recordings of South Bolivian Quechua. The stops in pairs of stimuli differ only as to whether the consonants are plain or ejective.

4.2.2 Acoustic properties of the ejective–plain contrast in the stimuli

To see what acoustic cues potentially underlie the results of the study, an acoustic analysis was conducted on the edited stimuli. Acoustic analysis serves two purposes. First, it verifies that cues to the ejective–plain contrast are present in all of the experimental stimuli, eliminating the possibility that any perceptual asymmetries result from acoustic irregularities. Second, if subjects show difficulty in perceiving the ejective–plain contrast in certain categories, as is predicted, we will know what acoustic properties are not being accurately perceived.

For each initial and medial syllable used in stimulus creation VOT, burst amplitude and voice quality of the following vowel were measured. VOT was measured in milliseconds from the beginning of the burst to the onset of periodicity in the following vowel. Burst amplitude was measured in arbitrary units directly off of the waveform by calculating the difference between the highest and lowest points in the waveform. Voice quality is quantified as the difference between the amplitude of H1 and H2 in the first 30 ms of the vowel (Ladefoged Reference Ladefoged2003).

Each unique sequence was measured once. For example, while [k'i] from [sak'i] was spliced into multiple stimuli (e.g. [pak'i], [tak'i] and [kak'i]), this sequence was only measured once. For all initial sequences, as well as medial sequences with an ejective, there were six unique tokens to measure. For medial sequences with a plain stop, there were twelve tokens. This asymmetry arises because no initial and medial sequence from the same natural utterance were spliced together. All medial syllables with ejectives were taken from an utterance with initial [s]. Medial syllables with plain stops were cross-spliced, so that [ki] in the stimulus [taki] was spliced from [t'aki] and [ki] in the stimulus [t'aki] was spliced from [taki].

The measurements for each place of articulation were subjected to a two-way ANOVA, with position and laryngeal type as factors. For all three places of articulation, there was a significant effect of VOT and burst amplitude, along with a significant interaction between laryngeal category and position. For the alveolars, there was also a small main effect of vowel voice quality (p<0·04), but no interaction with position. The statistical results are reported in Table I.

Table I

Statistical analysis of acoustic corelates of contrasts in experimental stimuli.

Mean values for amplitude and VOT for each consonant in initial and medial positions are given in Table II.

Table II

Average values for VOT and burst amplitude in experimental stimuli.

Acoustic analysis shows, as summarised in Tables I and II, that ejective and plain stops are distinguished by burst amplitude and VOT. These cues are reliably preserved in the spliced stimuli, since CV transitions are kept intact.Footnote ² Any difficulty in accurately discriminating between ejectives and plain stops, then, cannot be attributed to conflicting cues in the signal. Sample spectrograms are given in Fig. 1 for the syllables [k'a] and [ka]. Differences in the acoustics of contrasts depending on place of articulation and position are taken into account in the analysis of the perceptual results presented in §4.4 below.

Figure 1

Comparison of a voiceless unaspirated and ejective velar stop in Quechua.

Finally, the stimuli were tested for any long-distance interaction between ejective and plain stops. The main hypothesis is that the ejective–plain contrast is perceptually weaker in roots with another ejective stop. One possible explanation for such an effect could be acoustic – an ejective may alter the production of a stop elsewhere in the root, making ejectives and plain stops acoustically more similar in roots with another ejective. In theory, the acoustic effect could be that ejectives following other ejectives have weaker bursts and shorter VOTs than other ejectives (e.g. [p'] in [k'ap'i] vs. [kap'i]), or that plain stops following ejectives have louder bursts and longer VOTs than other plain stops (e.g. [p] in [k'api] vs. [kapi]). Given the phonotactics of Quechua, only the latter possibility can be tested; forms with two ejectives are absent in the Quechua lexicon, as are forms with medial ejectives following plain stops. The VOT and burst amplitude of plain stops in medial position were compared, depending on whether the initial stop was ejective or plain. Neither measure was significantly different in the two contexts, according to a two-tailed paired t-test (p>0·05 for both measures). The differences in VOT between the two contexts showed a slight trend (p<0·09) for longer VOTs following ejectives than plain stops. The average values between the two contexts, however, differed by only 4 ms (28 ms following an ejective stop vs. 24 ms following a plain stop). It seems unlikely that such a small difference would have any perceptual effect, especially considering that the contrasting ejectives in medial position have an average VOT of 108 ms.

4.3 Procedure

Subjects were presented auditorily with pairs of nonce roots and asked to decide whether the roots they heard were the same or different from one another. Subjects indicated their response by pressing one of two clearly marked keys on a standard USB keyboard. Subjects listened to the stimuli through a pair of high-quality headphones, while looking at a computer screen. The stimuli were presented while the computer screen was black, with a 300 ms interstimulus interval. After the second stimulus, a green line appeared on the screen. Subjects were told to indicate their response as quickly as possible once the green line appeared, but not before. All subjects thus listened to both stimuli in their entirety before indicating their choice of ‘same’ or ‘different’. Response time was limited to 1500 ms. If subjects did not respond within this time, they automatically went on to the next trial.

The subjects were 19 speakers of American English with no exposure to a language with ejectives. English speakers were chosen as subjects instead of native Quechua speakers to avoid a phonological bias in perception. Quechua speakers or speakers of any language with restrictions on ejectives, are cognizant of restrictions or trends in the distribution of these sounds, and this knowledge may affect their perception of nonce roots. Thus, while Quechua speakers may perform in a way consistent with the hypotheses, it would not be clear if this performance was due to the Quechua speakers' grammar, or to some basic property of the perceptual properties of ejectives in different environments. The English subjects were told that the sounds they were hearing were not from English, and would sound foreign. All subjects performed at above chance on the experiment as a whole, and made very few mistakes on the ‘same’ pairs. Each subject listened to five trial pairs to accustom themselves to the stimuli and task. The experiment took about 20 minutes, including a break halfway through, and subjects were paid for their participation.

5.1 Architecture of the grammar

The proposal in this paper is that laryngeal co-occurrence restrictions are driven by systemic markedness constraints on laryngeal contrasts. The particular set of contrasting root types in a given language is determined by the ranking of contrast markedness constraints that favour more perceptible contrasts over less perceptible contrasts, following the Dispersion Theory of Contrast (DT) framework of Flemming (Reference Flemming1995, Reference Flemming, Hayes, Kirchner and Steriade2004, Reference Flemming2006).

In DT, the grammar maps sets of input forms to sets of output forms, as opposed to the mapping of individual inputs to individual outputs in standard OT. The central hypothesis of the DT framework is that phonological patterns reflect grammatical pressures, formalised as systemic markedness constraints, to maximise the perceptual distinctness of contrasts. For the grammar to evaluate perceptual distinctness, sets of forms must be evaluated in parallel, and thus the grammaticality of a given form is dependent on what other forms it contrasts with.

The DT analysis of laryngeal co-occurrence restrictions presented in this section differs from a standard OT analysis in two ways: sets of inputs are mapped to sets of outputs, and the contrasts between output forms in a given set are evaluated by systemic markedness constraints. Systemic markedness constraints interact with standard input–output faithfulness constraints, which are evaluated as in standard OT. Systemic markedness constraints are universal, like standard OT markedness constraints, as they refer to language independent asymmetries in the perceptual distinctness of contrasts. The variation between languages is in the relative ranking of systemic markedness constraints with input–output faithfulness.

Some general motivation for systemic over non-systemic markedness constraints comes from segment inventories across languages, as discussed in Flemming (Reference Flemming, Hayes, Kirchner and Steriade2004, Reference Flemming2006). For example, the cross-linguistic preference for the peripheral vowels [i u] is formalised in standard OT as a fixed ranking of markedness constraints *[ɨ]≫*[i u]. This ranking, however, cannot account for the fact that the central vowel [ɨ] is preferred to the peripheral vowels [i u] in languages without a front–back contrast. Languages with only a height contrast in vowels never have [i] or [u] as the only high vowel; rather, the backness of a vowel varies depending on context, and tends to cluster around a central value. It is thus not the case that there is a difference in markedness of the three vowels [i ɨ u] in isolation, but rather that the contrast between a central and a peripheral vowel, [i – ɨ] or [ɨ – u], is more marked than the more distinct contrast between two peripheral vowels [i – u]. In the analysis of laryngeal restrictions, systemic markedness constraints take the place of standard OT markedness constraints on particular laryngeal configurations, like the Obligatory Contour Principle (OCP) or a non-local harmony constraint (see the specific proposals in Hansson Reference Hansson2001, Rose & Walker Reference Rose and Walker2004 and Mackenzie Reference Mackenzie2009). This analysis formalises the idea, presented earlier in §3, that certain combinations of laryngeal features are not marked in isolation; rather, certain contrasts in laryngeal configurations between roots are marked. The systemic analysis is explicitly compared with standard OT analyses of laryngeal restrictions in §6.

In both standard OT and DT, the role of the grammar is to map the rich base of all possible inputs to the set of attested forms in a given language. While in standard OT this mapping is modelled on individual forms, in DT the entire set of hypothetical forms is mapped onto the entire set of attested forms at once. The evaluation and interaction of constraints in DT tableaux are discussed further in the next subsection, once the systemic markedness constraints responsible for laryngeal co-occurrence restrictions have been introduced.

5.2 The constraints

Laryngeal co-occurrence restrictions reflect asymmetries in the perceptual strength of contrasts between roots with different laryngeal configurations. The experimental results reported in §4 reveal a three-tiered hierarchy in the perceptual distinctness of laryngeal contrasts in pairs of roots.

1. (15)

   

This hierarchy projects two systemic markedness constraints, defined in (16), corresponding to the two asymmetries in (15). I refer to the constraints that evaluate laryngeal contrasts between roots as laryngeal distance or LarDist constraints.Footnote ³

1. (16)

   

LarDist(2v1)-[F] penalises a minimal contrast in a given laryngeal feature between two roots each of which contains another segment specified for that feature. LarDist(1v0)-[F] is more general; it penalises a minimal contrast in a given laryngeal feature between two roots each of which contains another segment that can be specified for that feature. LarDist(2v1)-[F] penalises the weakest contrast on the scale in (15), while LarDist(1v0)-[F] penalises the next weakest contrast. There is no constraint against the 2 vs. 0 contrast, the strongest on the scale of laryngeal contrasts in (15) (cf. Gouskova Reference Gouskova2003).

LarDist(2v1) only penalises pairs of forms that contrast one and two instances of a laryngeal feature. LarDist(1v0), however, penalises pairs of forms with a 2 vs. 1 or 1 vs. 0 contrast in a laryngeal feature, or a contrast in the position of a laryngeal feature in the root, e.g. {k'api, kap'i}. Thus, pairs like {k'ap'i, k'api} (2 vs. 1), {k'api, kapi} (1 vs. 0) and {k'api, kap'i} (positional) all violate LarDist(1v0), while only a pair like {k'ap'i, k'api} (2 vs. 1) violates LarDist(2v1). The perception experiment in the preceding section found a hierarchy of perceptual strength between the 2 vs. 1, 1 vs. 0 and 2 vs. 0 contrasts, but did not explicitly test the relative strength of a positional contrast in laryngeal features. The hypothesis underlying the proposed constraints is that a positional contrast stands in the same place in the hierarchy as the 1 vs. 0 contrast; a pair like {k'api, kap'i} is presumed to be more easily distinguished than a pair like {k'ap'i, k'api}, but less easily distinguished than a pair like {k'ap'i, kapi}. Thus the grouping of a featural mismatch between segments, as in {k'api, kapi}, and a positional mismatch, as in {k'api, kap'i}, should form a natural class in perceptual terms. Future work must test this hypothesis.

(17) and (18) show how LarDist(2v1) and LarDist(1v0) evaluate pairs of roots that contrast in ejection. I characterise the collection of properties that distinguish ejectives from plain stops with the feature [ejective].

1. (17)

   

Only a pair like (17a), which contrasts one and two laryngeally marked stops, violates LarDist(2v1). For a pair of roots to violate LarDist(2v1)-[ejective], several conditions must be met. First, each root must have two segments that can contrast for [ejective]. In (17a), these segments are [k'] (x _R1) and [p'] (y _R1) in the first root and [k'] (x _R2) and [p] (y _R2) in the second. Second, the contrasting roots must be identical except that y _R1≠y _R2. This is true in (17a), as the only difference between the two roots is that y _R1 is an ejective [p'] and y _R2 is a plain stop [p]. Furthermore, x _R1 and x _R2 must both be [ejective], as is the case in (17a): x _R1 is [k'] and x _R2 is [k']. Finally, the roots must differ only in a single instance of [ejective]. In (17a), the only difference between the two roots is that y _R1 is ejective and y _R2 is not. This pair of roots thus violates LarDist(2v1)-[ejective].

1. (18)

   

LarDist(1v0)-[F] only penalises the 1 vs. 0 contrast in roots with another segment that contrasts for the given laryngeal feature. Consider the pair in (18b). This pair of roots violates LarDist(1v0)-[ejective], because each root contains two segments that contrast for [ejective] (x _R1 is [k'], y _R1 is [p], x _R2 is [k] and y _R2 is [p]), and the only difference between the two roots is that x _R1 is [ejective], [k'], and x _R2 is not, [k].

In (18), the contrast between an ejective and a plain stop is penalised in roots with another ejective or plain stop (the segments that contrast for [ejective]), as in the pairs {k'ap'i, k'api} and {k'api, kapi}, but not in roots with another nasal, as in the pair {k'ami, kami}. This distinction is necessary because languages with assimilatory restrictions neutralise the 1 vs. 0 contrast only in roots with two segments that contrast for the given laryngeal feature, as shown by the Chaha examples in (19). The examples here show that the contrast between [t] and [t'] is not possible in the context of another plain stop, [k] in (19a), but is possible in the context of a nasal, [m] in (19b).Footnote ⁴

1. (19)

   

It is thus necessary to distinguish between a 1 vs. 0 contrast in the context of another minimally contrastive consonant, e.g. {k'api, kapi}, and a 1 vs. 0 contrast in the context of a more dissimilar consonant, e.g. {k'ami, kami}.

While there is no experimental evidence supporting a perceptual distinction in contrast strength between the two environments in (19), the existence of a distinction (perceptual or otherwise) is an empirical fact. The requirement in languages with assimilatory restrictions is not that a laryngeal feature be realised more than once, but rather that a laryngeal feature be realised on all possible hosts in a given root. The hypothesised explanation for this fact is that perception of the 1 vs. 0 contrast is made more difficult by the presence of another potentially contrastive segment, i.e. the contrast in (19a) is weaker than that in (19b). Further experimental research is necessary to determine if this is in fact the case.

The constraints in (16) are in a stringency relation (Prince Reference Prince1997, de Lacy Reference de Lacy2002, Reference de Lacy2006). Any candidate that violates LarDist(2v1)-[F] also violates LarDist(1v0)-[F], but not vice versa. LarDist(2v1)-[F] penalises a contrast in [F] in the environment of another [F] segment anywhere in the root, e.g. {k'ap'i, k'api}. LarDist(1v0)-[F], however, penalises a contrast in [F] if there is another contrastive segment in the root regardless of whether that segment is marked for [F] or not. Contrasts like {k'ap'i, k'api} and {kap'i, kapi} both violate this constraint.

One additional systemic markedness constraint is necessary for the analysis, though it is not expressly supported by the experimental results. The general constraint LarDist(1v0)-[F] must have a more specific version, which penalises a 1 vs. 0 contrast between roots with specifically homorganic pairs of stops, e.g. {k'aki, kaki}. This constraint is necessary to account for mixed restriction languages where assimilation is required only between roots with homorganic pairs of stops. This constraint is defined in (20).

1. (20)

   

The hypothesis put forward earlier was that the distinction between homorganic and heterorganic pairs of consonants is perceptually based – the 1 vs. 0 contrast may be neutralised in only homorganic pairs of stops because this contrast is weaker in homorganic than in heterorganic pairs. While the experimental results reported above did not support the distinction between a 1 vs. 0 contrast in heterorganic and homorganic pairs of consonants in perceptual terms, future research into this area will hopefully reveal a perceptual or articulatory explanation for why assimilation may be restricted to only homorganic consonants, those pairs of segments that are made completely identical as a result of laryngeal assimilation. H-LarDist(1v0) and LarDist(1v0) stand in a stringency relation; the violations of H-LarDist(1v0) are a subset of those of LarDist(1v0), as can be seen in (21).

1. (21)

   

Systemic markedness constraints interact with input–output faithfulness constraints (McCarthy & Prince Reference McCarthy, Prince, Beckman, Dickey and Urbanczyk1995), which penalise contrast neutralisation. Consider (22).

1. (22)

   

Given a set of input forms showing minimal contrasts in [ejective], contrast neutralisation violates Ident[ejective], as shown in (23).

1. (23)

   

In (23a), the input form /k'ap'i/ is neutralised, violating Ident[ejective] once, due to the mapping of a form with two ejectives onto a form with only one ejective, either [k'api] or [kap'i]. In (23b), both forms with a single ejective are neutralised, incurring two violations of faithfulness because of the mapping of the plain stops in /k'api/ and /kap'i/ onto ejectives in [k'ap'i]. It could also be assumed that the forms /k'api/ and /kap'i/ are both mapped onto [kapi], still violating Ident[ejective] twice.

Given that the phenomenon of interest is the distribution of laryngeal features in roots, the analysis is presented with tableaux showing sets of roots that minimally contrast in laryngeal features. It should be kept in mind that the input in any tableau is the rich base, and that other constraints not shown in the tableaux are responsible for ruling out forms that are unattested for reasons orthogonal to laryngeal co-occurrence restrictions. The tableaux show only the smallest set of forms needed to represent the distribution of laryngeal features in the roots of a language.

5.3 The analysis of laryngeal co-occurrence restrictions

The variable ranking of Ident with LarDist(2v1) and LarDist(1v0) accounts for the difference between dissimilation and assimilation in laryngeal features. Three case studies illustrate the analysis.

5.4 Summary and discussion

This section has shown that all three types of laryngeal co-occurrence restrictions can be accounted for with the variable ranking of Ident and LarDist constraints. The variety of long-distance restrictions on laryngeal features is accounted for without constraints explicitly demanding non-local assimilation or dissimilation. Rather, both assimilation and dissimilation result from the interaction of constraints favouring perceptually distinct contrasts between roots. The ranking schema for dissimilatory, assimilatory and mixed restrictions are given in (40).

1. (40)

   

There are 24 possible rankings of the four constraints proposed for the analysis of co-occurrence restrictions, the predictions of which are now considered.

Dissimilatory languages result when LarDist(2v1) outranks Ident, and Ident outranks LarDist(1v0) and H-LarDist(1v0). The relative ranking of these latter two constraints is irrelevant, and thus two of the 24 possible rankings result in a language with a dissimilatory co-occurrence restriction.

Assimilatory languages emerge whenever LarDist(1v0) outranks Ident. The position of LarDist(2v1) and H-LarDist(1v0) in the hierarchy has no effect on the outcome, due to the stringency relation between LarDist(1v0) and the two other LarDist constraints. LarDist(1v0) outranks Ident, deriving assimilation, in twelve of the 24 possible rankings.

Mixed restrictions require that LarDist(2v1) and H-LarDist(1v0) outrank Ident, and that Ident outrank LarDist(1v0). The relative ranking of LarDist(2v1) and H-LarDist(1v0) is irrelevant, and thus two of the 24 possible rankings yield languages with mixed restrictions.

In six of the 24 permutations, Ident outranks all LarDist constraints. This results in a language without co-occurrence restrictions on laryngeal features. Such languages are attested; the data from Acoma in (9) above show that roots in this language may have two, one or zero ejectives.

In the remaining two ranking permutations, H-LarDist(1v0) outranks Ident, but the other two LarDist constraints rank below Ident. In this scenario, a language is predicted with place-dependent assimilation in laryngeal features but no restriction on heterorganic stops. This pattern is schematised in (41).

1. (41)

   

This type of pattern is attested in a gradient fashion for voicing in Muna (Coetzee & Pater Reference Coetzee and Pater2008). In Muna, heterorganic voiced and voiceless stops co-occur as often as expected by chance, but homorganic pairs of stops that disagree in voicing are underattested. Pairs of heterorganic and homorganic stops that agree for voicing are all well attested. The relevant O/E values are given in (42), taken from Coetzee & Pater's analysis of roots in van den Berg & Sido's (1996) dictionary, which they graciously shared with me.Footnote ¹⁰

1. (42)

   

Further research may uncover additional languages with this type of pattern. The factorial typology of the proposed constraint set predicts languages with the three well-attested types of laryngeal co-occurrence restrictions, as well as languages without co-occurrence restrictions, and languages with place-dependent laryngeal assimilation as in (40). §6 compares these predictions with the typology predicted by non-systemic accounts of laryngeal co-occurrence restrictions, and shows that non-systemic accounts both over- and undergenerate.

5.5 The potential role of misperception and coarticulation

An alternative to the systemic approach to laryngeal co-occurrence restrictions proposed here is the diachronic explanation of these phenomena presented in Ohala (Reference Ohala, Masek, Hendrick and Miller1981, Reference Ohala and Jones1993). Ohala outlines a perceptual account of long-distance laryngeal restrictions based on misperception. The basic idea is that laryngeal features like ejection often affect adjacent segments, e.g. glottalisation on the following vowel, thus causing an analytical problem for the listener. For example, a listener hearing a form pronounced as [k'ap'i] must decide whether glottalisation on [p'] is in fact due to this segment being specified as an ejective, i.e. related to the underlying representation /k'ap'i/, or is simply a result of coarticulation with preceding [k'], i.e. related to the underlying representation /k'api/. Co-occurrence restrictions are the phonologisation of misanalysis. Dissimilation in the lexicon results when forms with two ejectives are misanalysed as forms with a single ejective because the cues to the second ejective are attributed to coarticulation; assimilation results when coarticulation between an ejective and a non-ejective is analysed as two ejectives.

Ohala's explanation for long-distance interactions in laryngeal features relies on hypothesised coarticulation between the laryngeal feature of a given stop and surrounding segments. This hypothesis can be tested through acoustic analysis of the production of laryngeally marked segments. It must be shown first that ejectives, aspirates and implosives all have coarticulatory effects on surrounding segments, and second that the attested degree of coarticulation creates an ambiguous percept.

There are a few reasons to think that confusion about the number of laryngeal features in a root is not based in coarticulation. First, while the cues to ejectives and aspirates may indeed be realised partially on the following vowel, this coarticulation is often unidirectional. That is, in the majority of languages ejectives are postglottalised and aspirates are postaspirated (Ladefoged & Maddieson Reference Ladefoged and Maddieson1984). Thus it seems unlikely that breathy voice or glottalisation would be attributed to a following consonant, e.g. that something like [k'pi] would be interpreted as /k'ap'i/. It seems equally unlikely that breathy voice or glottalisation spreads all the way through to the right edge of the second consonant, e.g. that /k'api/ would be realised as [k'p], particularly given that in many languages a consonant may intervene between the two potential hosts for ejection or aspiration.

It is possible that coarticulation would lead to confusability in a language where ejectives and aspirates are preglottalised in coda position and postglottalised elsewhere, as is the case in many languages of the Pacific Northwest (see Steriade Reference Steriade1999, Howe & Pulleyblank Reference Howe and Pulleyblank2001). In a language of this type, a CVC sequence like [k'p] might be interpreted as /k'ap'/, as Ohala predicts. However, while it may be the case that coarticulation can lead to confusion about the number of laryngeal features in a root in some cases, this explanation certainly cannot cover all cases.

Ohala's account and the proposed systemic approach to laryngeal restrictions share the main idea that the ungrammaticality of forms with two laryngeally marked segments (*T'-K') or forms that disagree in laryngeal features (*T'-K) stems from the fact that these two types of forms are confusable with one another. The difference between the two approaches is that while Ohala attributes the perceptual confusability to the acoustic realisation of laryngeal features, the proposed analysis does not. The perceptual asymmetry documented in §4 arises in the absence of any coarticulation between stops; the cues to the ejective–plain contrast are identical across conditions. Further research is required to determine whether misperception could be due to coarticulation in some or all cases of long-distance laryngeal restrictions.

6.1 The analyses

Assimilatory restrictions on laryngeal features have been analysed in the broader context of long-distance consonant assimilation by Hansson (Reference Hansson2001) and Rose & Walker (Reference Rose and Walker2004). In their framework, similar non-adjacent consonants interact via a correspondence relation, and corresponding output consonants may be required to agree in certain features. Both output correspondence and identity between corresponding output consonants are governed by ranked and violable constraints. The necessary constraints to account for laryngeal assimilation are loosely defined in (43). For formal definitions and detailed discussion, see Hansson (Reference Hansson2001) and Rose & Walker (Reference Rose and Walker2004).

1. (43)

   

The constraints requiring correspondence between output segments refer to similarity, allowing for an account of general laryngeal assimilation, and assimilation between only homorganic pairs of stops, as in languages with mixed restrictions. In (42), two constraints of this type are shown. Corr-T↔K requires correspondence between all pairs of stops (e.g. [t-k]), but not pairs of non-stops (e.g. [m-l]) or pairs of one stop and one non-stop (e.g. [t-m]), and Corr-T↔T requires correspondence only between pairs of stops that agree in place (e.g. [t-t']).

Assimilatory restrictions, where all stops are required to agree in laryngeal features, result when Corr-T↔K and CC-Ident outrank input–output faithfulness, as shown in (44). Subscripts indicate correspondence relations between output segments.

1. (44)

   

Assimilation in languages with mixed restrictions applies only to pairs of homorganic stops. This pattern arises when the more specific Corr-T↔T and CC-Ident[ejective] outrank input–output faithfulness, which in turn outranks the general constraint Corr-T↔K, as shown in the three tableaux in (45).

1. (45)

   

The tableau in (45a) shows assimilation in laryngeal features between pairs of homorganic stops, while (45b) shows that under the same ranking, pairs of heterorganic stops are not required to assimilate. With this constraint set, roots with two heterorganic ejectives will surface faithfully, as shown in (45c). Additional constraints are necessary to enforce dissimilation between heterorganic stops in languages with mixed restrictions, as discussed below in §6.2.

In the correspondence-based framework, assimilation may be required of all pairs of stops or only pairs of homorganic stops, depending on the ranking of the constraints that require correspondence. In languages with assimilation, all stops stand in correspondence, while in languages with mixed restrictions, only homorganic stops stand in correspondence, as schematised in (46).

1. (46)

   

Additional constraints are necessary to account for dissimilation in laryngeal features, as found in languages with dissimilatory restrictions and languages with mixed restrictions.

MacEachern observes that dissimilation in laryngeal features is among the many phenomena that have been attributed to the Obligatory Contour Principle (OCP) (Leben Reference Leben1973, Goldsmith Reference Goldsmith1976, McCarthy Reference McCarthy1986), which states that adjacent, identical features are disallowed. Following the formulation given by Suzuki (Reference Suzuki1998), MacEachern adopts a set of OCP constraints that penalise multiple occurrences of the same laryngeal feature in a root. Dissimilation results when OCP outranks faithfulness to input laryngeal specifications, as shown in (47).

1. (47)

   

In (47), high-ranked OCP rules out candidates with multiple ejectives, the desired result for languages with a dissimilatory restriction.

6.2 Accounting for mixed restriction languages

A grammar with both long-distance assimilation through correspondence and the OCP can account for assimilatory and dissimilatory restrictions. The necessary rankings are shown in (48).

1. (48)

   

The ranking of OCP[laryngeal] below IO-Ident[laryngeal] in assimilation and of CC-correspondence constraints below IO-Ident[laryngeal] in dissimilation is necessary and is discussed in §6.3 below.

While the rankings in (48) can account for both dissimilation and assimilation, a problem arises when trying to account for languages with mixed restrictions with the same constraint set. Mixed restriction languages show dissimilation in pairs of heterorganic stops and assimilation in pairs of homorganic stops. These patterns require both the OCP and output correspondence constraints requiring place-dependent harmony to outrank IO faithfulness. Given such a ranking, however, the result is complete neutralisation of laryngeal features in roots with homorganic stops, as shown in the tableaux in (49).

1. (49)

   

Dissimilation in laryngeal features between pairs of heterorganic stops is correctly predicted in (49a). The OCP is high-ranked, forcing dissimilation. Corr-K↔P is ranked below faithfulness, and thus pairs of heterorganic stops are not required to stand in correspondence and high-ranked CC-Ident[ejective] is vacuously satisfied. The problem is encountered when pairs of homorganic stops are considered, as in (49b). Corr-K↔K is high-ranked, forcing the stops to correspond, and thus CC-Ident[ejective] requires laryngeal agreement, favouring the candidates [k'_x-k'_x] and [k_x-k_x]. Since the OCP also outranks input–output faithfulness, the winning candidate is that which satisfies all high-ranking constraints by deleting laryngeal features entirely, [k_x-k_x], as opposed to the intended winner [k'_x-k'_x].

An anonymous reviewer points out that the problem in (49) can be solved by proposing distinct OCP constraints penalising heterorganic and homorganic pairs of laryngeally marked stops. In languages with dissimilation, a general OCP[ejective] constraint against all pairs of ejectives outranks IO-Ident[ejective]. In languages with a mixed restriction, however, only the more specific OCP[ejective]-heterorganic outranks IO-Ident, and the general OCP[ejective] is lower-ranked. This analysis of mixed restrictions is shown in (50).

1. (50)

   

In (50a), OCP[ejective]-heterorganic forces dissimilation in roots with heterorganic stops. This constraint is not violated by pairs of homorganic ejectives, and thus the candidate with complete neutralisation in (50b), [k_x-k_x], loses to the desired winner with assimilation [k'_x-k'_x]. The place-based split in OCP constraints is formally similar to the place-based split proposed for LarDist(1v0).

The predictions of the various rankings of OCP, CC-correspondence and IO-Ident constraints are considered next.

6.3 Further predictions

A grammar with both the OCP and CC-correspondence constraints predicts two unattested patterns. First, if constraints enforcing dissimilation and assimilation in laryngeal features are both high-ranked, languages with complete neutralisation of laryngeal features in roots with two stops are predicted. Consider the pattern schematised in (51), which is derived from the ranking in the tableaux in (52).

1. (51)

   

2. (52)

   

If constraints demanding both assimilation (Corr-K↔P and CC-Ident[ejective]) and dissimilation (OCP[ejective]) outrank input–output faithfulness, as in (52), a laryngeal contrast will only surface in roots with one stop (52c). Inputs with two ejectives (52a) or one ejective and one plain stop (52b) map to outputs with no ejectives.

This is exactly the prediction that is not made by a systemic account of co-occurrence restrictions. The problem in (52) is that independent constraints penalise forms with two ejectives (the OCP) or one ejective (CC-correspondence), and both types of constraints can be fully satisfied by deleting all laryngeal features. With systemic constraints, only the contrast between forms with one and forms with two ejectives is marked, and thus there is no motivation for eliminating both types of forms.

A second unattested pattern is predicted if the OCP is divided into heterorganic and general versions, as is necessary to account for languages with mixed restrictions. If only OCP[ejective]-heterorganic outranks faithfulness, a language is predicted with dissimilation in pairs of heterorganic stops, and no restriction on pairs of homorganic stops. This pattern is schematised in (53) and derived by the ranking in the tableaux in (54).

1. (53)

   

2. (54)

   

In (54a), an input with two heterorganic ejectives is mapped to a output with only a single ejective, in order to satisfy high-ranked OCP[ejective]-heterorganic. Inputs with a heterorganic ejective–plain pair surface faithfully, as in (54b). The tableaux in (54c, d) show that both roots with homorganic ejectives and homorganic ejective–plain pairs surface faithfully. OCP[ejective]-heterorganic does not penalise pairs of homorganic ejectives. With CC-correspondence constraints ranked below IO-Ident, homorganic ejective–plain pairs are also grammatical. The tableaux in (54a, b, d) show two winners, which differ only in whether the output stops correspond or not. The choice between these two candidates depends on the relative ranking of low-ranked Corr-K↔P and CC-Ident[ejective].

To my knowledge, the language predicted in (54) is unattested. While there are languages where dissimilation only applies to heterorganic pairs of stops, these are all languages with assimilation in pairs of homorganic stops, i.e. languages with mixed restrictions. The prediction that forms with one or forms with two laryngeally marked stops are only marked when they contrast with one another, as predicted by the systemic analysis, is supported by the typology.