3.2.1. Lenition

The pattern of allophony targeting voiceless obstruents in word-initial position in Campidanese is typically described as weakening or lenition (Wagner Reference Wagner1950 1997; Virdis Reference Virdis1978; Contini Reference Contini and Andersen1986; Bolognesi Reference Bolognesi1998; Mensching & Remberger Reference Mensching, Remberger, Ledgeway and Maiden2016; Lai Reference Lai, Gabriel, Gess and Meisenburg2021b). Lenition is a descriptive term used to refer to the process which produces alternations of the kind in (1) – it does not as yet have any formal, theoretical status in this analysis. The primary ambition of this section is to lay out the facts concerning lenition in Campidanese, so that the disjunctions pointed out in §3.2 can be given a phonological explanation.

In (1), it was shown that lenition targets stem-initial voiceless stops in intervocalic contexts. The same position also triggers lenition of the fricatives /f s/, manifested as voicing, as in (6):

Lenition also operates on the first member of stop–sonorant clusters, as in (7):

In this same context, /l/ is also in an allophonic relationship with [ʁ] (or sometimes [ʕ]; see Molinu Reference Molinu2009), as in (8):

Since this alternation has the same structural description as those in (1) and (6), following arguments from Kisseberth (Reference Kisseberth1970) concerning the functional unity of phonological rules, I will consider all these alternations to be the result of a singular process of Lenition, which must be given formal status (see §5.3). The categorical nature of the alternation’s structural change, which affects both manner and place of articulation in a seemingly arbitrary way, requires an abstract phonological analysis that does not depend on phonetic facts (Chabot Reference Chabot2021; Scheer Reference Scheer, Honeybone and Salmons2015).

3.2.2. Lenition in non-sandhi positions

The word-medial position merits some discussion regarding the effect of Lenition. In this position, there are no alternations, but the distributional facts show a preponderance of spirants and voiced fricatives: [zriβɔ̃ĩ] ‘wild boar’, [diu] ‘finger’, [fo u] ‘fire’. Bolognesi (Reference Bolognesi1998) and Lai (Reference Lai2015b, Reference Lai, Gabriel, Gess and Meisenburg2021b) argue that these are the result of lexicalised sound changes, and not the result of synchronic lenition as in (1). Lai (Reference Lai and Russo2015a: 275) provides the most explicit argument to this effect, suggesting that since word-medial obstruents in items such as [proku] ‘pig’ are not realised as spirants, it shows that by the time the diachronic process of metathesis which changed Latin porcu $>$ ˈporku $>$ ˈproku was completed, any synchronic rule of lenition had already ceased to be productive.

Synchronically, this position introduces a number of difficulties. The first is that it establishes an active synchronic process which targets only word-initial onsets in intervocalic contexts, while word-medial onsets in the same context are spared. The second is that it introduces a number of phonemes, including //, which are distributionally limited to the intervocalic context, the very context which targets stops for spirantisation and fricatives for voicing. This includes /z/, which is not a phoneme in Campidanese (Lai Reference Lai, Gabriel, Gess and Meisenburg2021b: 606), since per Bolognesi (Reference Bolognesi1998: 28) it does not occur in absolute word-initial position. Words such as [kazu] ‘cheese’ suggest that there is an underlying /s/ which is being voiced, thus an active process of lenition targeting intervocalic obstruents. The same is true of words with suffixes such as 3sg in /pappa-t/ ‘to eat 3-sg’ or the plural /-s/ as in /faula-s/ ‘lies’, where the final morpheme of each form may surface with a following epenthetic copy vowel, such that they appear as [pap:aa] and [fauʁaza], respectively. Assuming that [z] and [] are not contrastive segments in Campidanese, their presence in these forms can be explained if they are the result of Lenition of /s/ and /t/, respectively.

A grammar that generalises over the facts in §3.2 while ignoring word-medial intervocalic voiceless obstruents is significantly more complex, with a rule of lenition that distinguishes between external sandhi (V#.CV) and word-internal intervocalic contexts (V.CV), along with an increase in the size of the phonemic inventory. If the rule that targets word-initial obstruents in external sandhi is also active in word-medial position, the rule itself is much simpler, and the distribution of spirants can be easily accounted for within the phonological grammar.Footnote ⁵ This is an application of the Free-Ride Principle discussed by Zwicky (Reference Zwicky1970), where non-alternating forms are assumed to be subject to an active phonological process in a grammar (see Krämer Reference Krämer2012: 41–42 for discussion). For this reason, this analysis considers Lenition to be active in word-medial positions. I will take up the case of [proku] and its representation in §3.2.4.

3.2.3. Non-targets of Lenition

Lenition does not target all obstruents in Campidanese. As discussed in §3.2, voiced stops may variably be reduced to zero in this context, but unlike /p t ʧ k f s l/, which are always targeted by Lenition when in the proper context, voiced stops show variation, with a number of lexical exceptions in which they never delete. Furthermore, the voiceless fricative /ʃ:/ and the voiceless affricate /ʦ:/ never undergo Lenition, even when in the proper context (Bolognesi Reference Bolognesi1998: 33), and the same is true of /dz/, /v/ and /ɖ:/ (Bolognesi Reference Bolognesi1998: 39). The nasals, /n m ɲ/, never lenite in external sandhi: [su niu] ‘the nest’. Word internally, /n/ does seem to lenite, but only when following the main stress-bearing vowel (Bolognesi Reference Bolognesi1998: 26). Given the essential role played by stress in the structural description of N-deletion and the fact that there are no alternations in intervocalic positions created by sandhi, this process is not the same as Lenition, and is not treated here.

3.2.4. Resistance to Lenition: Virtual Geminates

Recall that one of the arguments against synchronic word-medial Lenition is that words such as [proku] do not have word-internal spirants (Lai Reference Lai, Gabriel, Gess and Meisenburg2021b). I argued in §3.2.2, however, that the same process of Lenition that targets word-initial obstruents in external sandhi is in fact active in word-medial position. In this section, I argue that gemination is a manifestation of strength in that geminate stops never undergo Lenition.

In order to understand why Lenition does not target voiceless stops in words such as [proku], consider the data provided by Bolognesi (Reference Bolognesi1998: 149) in (9), in which voiceless obstruents are realised in surface forms:

Immediately, what stands out in (9) is that all of the lenition-resisting objects are realised as phonetically long. However, in Campidanese, the status of phonemic geminates is uneven: only for the sonorants /r n l/ does phonetic length always correspond to an underlying contrast between geminates and singletons (Virdis Reference Virdis1978; Bolognesi Reference Bolognesi1998). Phonemically long obstruents are variable and may be realised as short. For sonorants, Bolognesi (Reference Bolognesi1998: 161) provides some near-minimal pairs as in (10), though the deletion of /n/ and its effect on vowels makes the contrast between ‘hand’ and ‘big-msc.’ less obvious:

The contrasts in (10) suggest that geminate structure is active in the phonology of Campidanese. Lai (Reference Lai2015b, Reference Lai, Gabriel, Gess and Meisenburg2021b) notes, however, that for all other obstruents in Campidanese phonetic duration is not contrastive. Even in words such as those in (9), geminates may be realised as phonetically short, meaning that duration is not a reliable correlate for geminacy in Campidanese (Bolognesi Reference Bolognesi1998; De Iacovo & Romano Reference De Iacovo and Romano2015).

While it seems obvious that an increase in phonological timing should result in an increase in phonetic duration, phonological timing is above all a matter of phonological representations (see Davis Reference Davis, Oostendorp, Ewen, Hume and Rice2011 for discussion), and many factors related to performance can impact the manifestation of timing as duration (Clements Reference Clements, Wetzels and Sezer1986: 39). In Italian, for example, length is not always the primary phonetic correlate of gemination (Payne Reference Payne2005, Reference Payne2006). Geminates that are not expressed as phonetically long are what Ségéral & Scheer (Reference Ségéral, Scheer and Dziubalska-Kołacyk2001a: 311ff.) refer to as virtual geminates, objects that are associated with two positions on the skeleton, but whose surface realisation is identical to a corresponding singleton.Footnote ⁶

With no phonetic correlate on the consonant itself available for identifying phonological geminates, they can only be identified through phonological behaviour. The principal characteristic of geminates is that they never lenite (see Jones Reference Jones, Harris and Vincent1988: 321 and Bolognesi Reference Bolognesi1998: 33).Footnote ⁷ Geminate resistance to Lenition is a manifestation of Inalterability (Hayes Reference Hayes1986), a structural property inherent in geminates that protects them from Lenition. Since words such as [mak:u] ‘crazy’ never undergo Lenition, regardless of the phonetic length of the medial consonant, they must be geminate (see also Barillot & Ségégral Reference Barillot and Ségégral2005; Barillot et al. Reference Barillot, Bendjaballah and Lampitelli2018 for a comparable case in Somali). Whether a voiceless stop is realised as long or short does not affect its phonological status: speakers perceive them as the same object.

This suggests that any word such as [proku] which does not manifest surface length but which resists Lenition is a virtual geminate and that this fact must be discoverable by language learners. In the case of [proku], for example, the resistance to lenition of word-medial stops is enough for learners to recover geminate structure – the underlying form /prokku/.Footnote ⁸ The result is a phonological geminate, which is not realised with phonetic length, recoverable through its resistance to Lenition.

3.2.5. Fortition

As shown in (2), in the intervocalic configuration which triggers spirantisation of voiceless stops, voiced stops never spirantise. However, (4) exhibits a pattern of alternation in which voiced stops do in fact spirantise. This is what Bolognesi (Reference Bolognesi1998) calls ‘pseudo-lenition’, and what Katz (Reference Katz2021: 657f.) refers to simply as lenition.

I argue that spirantisation of voiced stops is not lenition in the phonological sense, but rather is the result of a process that targets voiceless stops as well, for which I will provisionally adopt the term Fortition. The effect of Fortition is most apparent where voiceless stops are concerned, since they are generally realised with phonetic length. Ultimately, I will argue that Fortition affects voiced stops, as well as voiceless stops, fricatives and sonorants, though only members of the latter must be realised with phonetic duration. Unexpectedly, voiced stops are realised as spirants when subject to Fortition. This will be shown through the examination of three related contexts in which Fortition is active in Campidanese. What unifies the three contexts is that in each, there is an empty timing position to the left of the targeted segment.

The first context to consider is word-initial following a final stop in a preceding word, such as the plural marker /-s/ or the 3sg verbal marker /-t/. In such cases, the final obstruent does not surface as a coda, and instead triggers either the insertion of a paragogic copy vowel if realised, or subsequent lengthening of the word-initial stop if elided (Contini Reference Contini and Andersen1986; Jones Reference Jones, Harris and Vincent1988; Molinu & Pisano Reference Molinu, Pisano, Rainer, Russo and Miret2016; Lai Reference Lai, Gabriel, Gess and Meisenburg2021b). In the description given by Jones (Reference Jones, Harris and Vincent1988: 322), in such circumstances, initial consonants are ‘reinforced’ or given a ‘geminate pronunciation’. Indeed, Bolognesi (Reference Bolognesi1998: 190) sees this as a fortition manifest as surface gemination, and provides the following examples:

The examples in (11) represent a subcase of Fortition, which I will refer to as compensatory lengthening. Lai (Reference Lai, Gabriel, Gess and Meisenburg2021b) calls this a synchronic process of fortition, by which an obstruent /p/ is realised with increased length [p:], as in (11a). The same is true of fricatives, as in (11b).

When a voiced stop is realised in parallel contexts, the result is a spirant, as shown in (11c). What (11) and (4) show is an interesting dual pattern: in compensatory lengthening contexts, voiceless stops geminate, while voiced stops spirantise.

The second context of Fortition in Campidanese is fed by a process of metathesis. Diachronically, metathesis characterises the evolution of Latin to Sardinian generally (Molinu Reference Molinu1999), but its effect was particularly salient in Campidanese (Virdis Reference Virdis1978). Lai (Reference Lai2013, Reference Lai2014, Reference Lai and Russo2015a) identifies three kinds of metathesis, each of which affects the rhotic phoneme /r/:

While (12a) and (12b) are diachronic processes, (12c) is active synchronically in words that begin with a vowel and are disyllabic: VrCV (Bolognesi Reference Bolognesi1998: 419). This process is key for understanding how Fortition works in Campidanese, and for what it tells us about prosodic structure and the synchronic lenition process (13):

As /r/ moves into the branching onsets shown in (13), it suppresses the realisation of the initial vowel in the article, and triggers lengthening of following voiceless stops (13a) and spirantisation of following voiced stops (13b), a dual patterning which parallels (11).

The third context of Fortition occurs after certain vowel-final prepositions and connectives which have lost an etymological final consonant in diachrony (Jones Reference Jones, Harris and Vincent1988): for example, /a/ ( $<$ ad or aut) ‘to/at’, /ɛ/ ( $<$ et) ‘and’, /nɛ/ ( $<$ nec) and a handful of others. Lengthening is thus triggered by unstressed monosyllables with etymological coda consonants (Bolognesi Reference Bolognesi1998; Lai Reference Lai, Gabriel, Gess and Meisenburg2021b). The lost etymological consonant is what Bolognesi (Reference Bolognesi1998) refers to as a ghost consonant, since it appears to mark the context for a certain subset of Fortition processes. As in some other languages of Italy, this appears to be a kind of Raddoppiamento fonosintattico (RF).Footnote ⁹ That is, following Fanciullo (Reference Fanciullo1986: 67), an initial consonant is realised as geminate if immediately preceded by an item specified in the lexicon to trigger RF, as in (14):

RF is common to all varieties of Sardinian and represents a kind of strengthening (Contini Reference Contini and Andersen1986). In RF, as in compensatory lengthening, there is a dual pattern: voiceless consonants are realised as geminate (14a), while voiced consonants are realised as spirants (14b).Footnote ¹⁰

3.2.6. Summary of the empirical situation in Campidanese Sardinian

To summarise, Campidanese is characterised by a process, Lenition, which spirantises voiceless stops but does not target voiced stops. There are three processes, however, which do produce spirantised voiced-stop realisations: compensatory lengthening triggered by the loss of a preceding word-final consonant (11c), compensatory lengthening induced by metathesis (13b) and RF (14b). In addition, these latter three processes all result in lengthening of voiceless stops, and so I refer to all three processes as Fortition. The pattern of lenition and gemination is schematised in Table 1.

Table 1 A summary of spirantisation and lengthening patterns in Campidanese Sardinian.

Considering the distribution of spirants and stops in Table 1, there is a direct link between the context of lengthening in voiceless stops and spirantisation in voiced stops. Fortition, then, which results in /p t ʧ k/ being realised as geminate, also involves phonological gemination of /b d ʤ g/, despite their phonetic realisation as spirants. To suppose otherwise is to interpret as an accident the fact that RF, compensatory lengthening and metathesis-induced compensatory lengthening all have complementary scope over voiceless and voiced stops. I argue they are the result of a singular process of strengthening which is reflected in prosodic structure (§4.2).

In a surface-based approach (cf. Katz Reference Katz2021), this conclusion is surprising, since spirantisation is a classic case of lenition. Kirchner (Reference Kirchner2000: 510) suggests that spirant realisations of geminates are suboptimal and violate constraints which select output candidates for articulatory ease as well as perceptual faithfulness, and therefore can never be selected by a grammar.Footnote ¹¹ This is no doubt what leads Bolognesi (Reference Bolognesi1998: 165) to argue that voiced stops spirantise precisely because ‘they cannot give rise to geminate structure’, as voiced postlexical geminate structures are ill-formed. In order to prevent the grammar from producing such structures, Molinu (Reference Molinu1999: 169) imposes a constraint on the grammar which blocks voiced geminates in postlexical phonology, arguing that the RF process which produces geminates in Lugodorese instead gives rise to ‘variantes non-géminées et spirantisées’ in Campidanese. That is, a spirant is explicitly not a geminate, since a constraint in the grammar interdicts the gemination of voiced obstruents. An equivalent constraint blocks gemination in the analysis of Katz (Reference Katz2021: 666).

I argue that this is a classic case of ‘substance abuse’ (Hale & Reiss Reference Hale and Reiss2000b, Reference Hale and Reiss2008) – a misuse of the phonetic facts in the building of the analysis. This has two unfortunate results in Campidanese. The first is the loss of generalisation entailed by analysing spirantisation and gemination in a disjunctive way as a function of the output. The second is the bloating of the synchronic grammar in order to prevent voiced stops from geminating. The solution to these problems, I argue, is to recognise that compensatory lengthening, metathesis-induced compensatory lengthening and RF produce geminate structures from all stop inputs, voiceless and voiced alike. The correct view is to analyse all three Fortition processes as a singular, unified process that results in a phonological geminate, resistant to Lenition as revealed by voiceless stops and surfacing as spirants in the case of voiced stops. This conclusion discards entirely the phonetic properties of the segments in question, and emerges only from consideration of phonological structure and behaviour.

Following a general principle established by Hyman (Reference Hyman1970), the advantage of positing such abstract structures is the explanatory value they provide:Footnote ¹² patterns of spirantisation and gemination are the result of two prosodic effects, one of Lenition and one of Fortition. Lenition is a melodic process that targets voiceless stops in a weak prosodic position, triggering the loss of melodic material. Fortition is a prosodic effect that spreads melodic material and results in phonological gemination in all cases. The result is a unified analysis of strength and weakness in Campidanese, summarised in Table 2. In this view, both weakening and strengthening are still metaphorical notions – their formal status is in their prosodic representations, and how each process falls out from prosodic structure (§4).

Table 2 A summary of positional effects in Campidanese Sardinian.

The facts in Campidanese suggest some notion of phonological weakness inherent in the context that conditions obstruent lenition, and strength inherent in the context that resists the lenition process. That is, weak contexts allow lenition, and strong contexts produce geminate structure. The observation that segments that resist Lenition are geminate does not, in and of itself, constitute an explanation for their exceptional status – it merely recapitulates the distribution of stops and spirant allophones – nor does it satisfy the requirements for a theory of lenition (see §2.2). To do these things, an adequate theory of phonological representations and computations is required.

4.2.1. The analytic tool: Strict CV phonology

In order to capture the three distinct prosodic positions, this analysis makes use of the basic machinery of Strict CV phonology (Lowenstamm Reference Lowenstamm, Durand and Laks1996; Scheer Reference Scheer2004b, Reference Scheer2012), a development of the Government Phonology program (Charette Reference Charette1990; Harris Reference Harris1990; Kaye et al. Reference Kaye, Lowenstamm and Vergnaud1990; Harris & Kaye Reference Harris and Kaye1990). In Strict CV phonology, prosodic structure is built not out of hierarchical arborescent structures, but out of lateral relationships between constituents on a CV tier. This means that the three different positions of strength and weakness are distinguished by the different configurations of lateral relations between members of the CV tier.Footnote ¹⁵

In Strict CV, the skeletal tier is built from an invariant alternation between C and V positions.Footnote ¹⁶ Relations between C and V positions are defined by two lateral forces, government and licensing, and the difference between C and V lies in their distinct licensing powers. Co-occurrence restrictions between adjacent segments are not due to hierarchical syllabic structure, but to lateral relations between the segments; ‘branching onsets’ or ‘onset and coda’, for example, are not in a relationship derived from an arboreal hierarchy, but one derived strictly in terms of the lateral relationships between them (Kaye et al. Reference Kaye, Lowenstamm and Vergnaud1990). Syllabic structure, then, is not a phonological primitive per se, but a derived property of adjacency relations.

Roughly speaking, government is a force which serves to weaken or inhibit melodic material, while licensing reinforces it. Both originate at the right edges of words, propagating back – V positions with melodic material may govern and license constituents to their left. There is a hierarchical relationship between governing and licensing: they cannot both exert influence on the same segment (Scheer Reference Scheer2012). If a segment is potentially subject to both lateral forces, it will be subject to government. In turn, licensing will influence the next available segment. When a CV unit is full of melodic material, the V position will contract a relationship of licensing from any following licenser, while the C position will contract one of government.

In Strict CV, all morphological boundary information – information that communicates with the interfaces and is translated from morphology – must be in the form of CV units, meaning that representations contain an empty initial CV at the left edge of words. The difference between a CV unit and a conventional # is that a CV unit is a true phonological object through which morphology is translated into phonology, while # is an arbitrary diacritic whose only function is to mark morpho-syntactic boundaries. This initial CV may enter into a lawful lateral relationship just like any other CV position (Lahrouchi Reference Lahrouchi2018; Lowenstamm Reference Lowenstamm, Rennison and Kühnhammer1999; Scheer Reference Scheer2009, Reference Scheer2012). As a phonological object, it exerts an influence on the structure of lateral relationships in phonology.

The nature of the strict ordering of CV elements in lateral phonology, along with CV-interpreted morphosyntactic information, means that there is potential in any representation for a number of empty C or V positions – positions with no associated melody. Empty V positions are known as empty nuclei. All languages impose restrictions on the number of empty nuclei in a given prosodic representation. In order to remain unexpressed, empty nuclei must contract a relationship of government, which they get from any filled V position that follows. Consider the representation in (16), which shows the interaction between the forces of government, licensing and the empty positions in the initial CV:Footnote ¹⁷

In (16), C $_2$ contains lexically specified melodic material, and contracts a relationship of government from the following vowel, V $_2$ , which in turn licenses V $_1$ , since the latter also contains lexically specified melodic material. V $_1$ , on the other hand, must govern the empty V position of the initial CV, and thus licenses C $_1$ . This means that C $_1$ is in the position of the coda mirror; being [+Lic], it is not subject to lenition. Thus, [+Lic] is a formal configuration of prosodic structure that reinforces and does not diminish melodic primes; in Campidanese, this lateral configuration is phonologically stable. The position of the intervocalic C $_2$ is an onset, like C $_1$ , but has a distinct status in this representation, since it is governed, but not licensed. This position, [+Gov], is where Lenition occurs – here /k/ is realised as [].

Since filled V positions license and govern, and codas are followed by empty V positions, they do not enter into any lateral relationship. The result is a [ $-$ Lic, $-$ Gov] position, with reducing power to license melodic primes. This weak position is subject to severe cooccurrence restrictions, with only /r/, homorganic nasals, or the first part of a geminate being allowed to surface as a coda in Campidanese. For example, a word such as /fatat/ ‘do-3sg’ has the following underlying representation:

The representation in (17) has an empty final V position. If an empty nucleus cannot contract a lateral relationship, it must be expressed.Footnote ¹⁸ In Campidanese, this restriction results either in the deletion of the coda or in the realisation of a paragogic copy vowel identical to the stem-final vowel in ungoverned empty V positions. The representation in (18) is the surface realisation of (17):

Lateral relations thus not only affect phonological computation, but also determine the well-formedness of a string (Scheer Reference Scheer2012: 145). In order to remain unexpressed, a V $_1$ must be governed by a following V $_2$ ; if V $_2$ is empty, it cannot govern V $_1$ . This predicts that there may not be a sequence of two consecutive empty nuclei: an empty V $_2$ cannot govern a preceding V $_1$ , which thus cannot remain empty and must be phonetically expressed. In classical terms, the result is epenthesis, as when the post-consonantal copy vowel surfaces in Campidenese. In this way, well-formedness does not come from outside of phonology in the form of arbitrary constraints – rather, it is the result of lateral relations (see Lai Reference Lai and Russo2015a,Reference Laib for a discussion of other similar effects elsewhere in Sardinian).

The second solution for the problem presented by /fatat/ is to delete the final consonant. This solution is available when there is a following C-initial word, since final and internal codas do not have the same status in Campidanese. The latter precede a governed nucleus, which thus cannot be expressed and so cannot contract any kind of lateral relationship with any final consonant. The association between this consonant and its position on the CV tier then moves to the following consonant, resulting in compensatory lengthening (represented by the dashed association line), as in /fatat luna/ ‘the moon is shining’, represented in (19):

In (19), government and licensing proceed as normal from the right edge. The /l/ in C $_4$ is normally subject to Lenition, but is protected here because it is licensed. The empty position at V $_3$ attracts government from V $_4$ , remaining empty, and consequently C $_3$ is a coda position, a position which imposes severe distributional restrictions on segmental material in Campidanese, as mentioned above. Since the stop in C $_3$ cannot be associated with the timing tier, the association it projects moves to C $_4$ , resulting in geminate structure. V $_2$ contracts lateral relationships as normal, and V $_1$ governs the V position in the initial CV instead of C $_1$ , resulting in another strong position.

What is the content of this restriction on coda licensing? Harris (Reference Harris1990) and Cyran (Reference Cyran, Carvalho, Scheer and Ségéral2008, Reference Cyran2010) argue that the amount of melodic material in a segment corresponds to its substantive complexity: the more melodic material in a segment, the more complex it is. Cyran (Reference Cyran2010) argues that substantive complexity has consequences for prosodic structure, since the more melodic material a segment has, the more licensing strength it requires. Such complexity scales mean that positions that are not licensed cannot host as much melodic material as those that are. In Campidanese, positions that contract no lateral relationships – codas – are weak, licensing only minimally complex segments. In particular, codas may license nasal segments homorganic for place and the first parts of geminates, structures that have in common the ‘sharing’ of melodic material with following segments. In these cases, the strength of the structures is reflected by the sharing of segmental material that would otherwise be prohibited in [ $-$ Lic, $-$ Gov] positions (Honeybone Reference Honeybone, Carr, Durand and Ewen2005b).

In sum, lateral relations define the three positions of prosodic strength, with each receiving a unique prosodic identity. Importantly, they define the coda mirror as [+Lic], making it formally distinct from the intervocalic position, which is [+Gov]. Finally, since codas do not contract any lateral relationships, being [ $-$ Lic, $-$ Gov], their power to license contrasts is reduced. The three positions are summarised in Table 4.

Table 4 The lateral relations of the three positions of strength in Campidanese.

Distinguishing among these three positions means that the context of Lenition in Campidanese can be given a unified context: [+Gov]. Only segments with this lateral configuration can be targeted by Lenition; any other configuration is spared from its effects.

4.2.2. Prosodic position and geminate inalterability

We are now in a position to see how prosodic position interacts with Lenition and Fortition and gives rise to geminate inalterability – and how the notions of strength and weakness may be understood in a substance-free approach.

First, let us consider the case of Lenition, which targets intervocalic obstruents, as in (20):Footnote ¹⁹

In (20), V $_3$ licenses V $_2$ , which in turn licenses V $_1$ , resulting in both C $_3$ and C $_2$ being [+Gov] – weak positions. Since those positions are filled by segments subject to Lenition, each is realised on the surface as its corresponding weak allophone.

One prediction made by this theory is that word-medial intervocalic positions, such as C $_3$ , are also [+Gov], and thus potential targets for Lenition. While I presented conceptual arguments for considering spirants in this position to be the result of lenition in §3.2.1, to those arguments can now be added a theoretical one: this position is weak by virtue of its prosodic structure, so any potential target in this position is expected to be subject to Lenition. This is exactly how a learner is able to recover singletons in this position, since any surface stop in a word-medial position is a geminate, any surface spirant is a singleton.Footnote ²⁰ Thus, the simplifications to the grammar that come from assuming Lenition is active in this position fall out from the lateral relations of intervocalic consonants.

In Strict CV, the relationship which characterises obstruent–sonorant (TR) clusters is known as infrasegmental government (see Scheer Reference Scheer2004b, Reference Scheer2012). Infrasegmental government (represented as T $<$ =R) is a specific kind of government which holds between constituents of branching onsets, and has the effect of suppressing exponence of the V position between the two members, which remains empty. In contrast to other empty V positions, the empty V positions in TR clusters are good lateral actors, and may contract government and licensing relationships with other CV positions.

In (21), we see that V $_2$ is circumscribed by infrasegmental government and thus is a good lateral actor, governing C $_2$ and triggering lenition at that position:

This brings us back to geminate inalterability. As argued in §3.2.4, in Campidanese, phonetic duration is not a consistent correlate of phonological geminates – rather, it is resistance to Lenition that is the defining characteristic of phonological geminates. Infrasegmental government means that the empty V position between two consonants in a tautosyllabic TR cluster does not act like an empty V position between heterosyllabic consonants. Strict CV predicts that any C position which is [+Gov] is weak, and since geminates are not a target of Lenition, they must have a different lateral configuration, corresponding to prosodic strength. In all geminates, there is an empty V position between two C positions. This empty V must be governed to remain empty, as we can see for lexical geminates, as in makku ‘crazy’ in (22):

In (22), V $_3$ governs the empty position V $_2$ and licenses C $_3$ , leaving C $_3$ in the strong [+Lic] position and protecting it from Lenition. C $_2$ , in turn, contracts no lateral relationship, leaving it in the coda [ $-$ Lic, $-$ Gov], which is also immune to lenition. Since in the case of (22) the coda is the first element of a geminate, the weak position C $_2$ is able to license the melodic material inherited from C $_3$ . This gives us a formal representation of strength inherent in lateral relations: their interaction with multiply associated segments determines where Lenition is active and where it is not.

5.2.1. Fortition and the interface

The realisation of geminate /b d ʤ g/ as spirants is a surprising conclusion in any theory where phonological structure is recapitulated in phonetic substance. In the substance-free view adopted here, though, this is not relevant to phonology qua computation; rather, it is an example of the realisation of phonological structure being shaped by third-factor phenomena (Chomsky Reference Chomsky2005). In particular, this is an effect of the physical exigencies of voicing, which requires air to flow through the vocal cords and into the oral cavity (Ohala Reference Ohala1997).

During stop production, all exit valves are closed, and as air flowing through the glottis builds up in the buccal cavity, oral pressure nears subglottal pressure (Ohala Reference Ohala and MacNeilage1983: 194ff.). When this happens, the air flow through the glottis is diminished and voicing is no longer possible. The longer the stop is held, the greater the likelihood that voicing will be extinguished. This results in a strong tendency for long voiced stops – geminates – to become voiceless.

However, Ohala (Reference Ohala and MacNeilage1983) notes that the tension between voicing and stops can be relieved in other ways, notably by ‘unstopping’ the stop – changing it to a voiced fricative or approximant. This third-factor effect would have been an actor on voiced geminates at a diachronic stage when voiced geminates were realised with increased duration, as they are synchronically in Logudorese (Ladd & Scobbie Reference Ladd, Scobbie, Local, Ogden and Temple2003). If spirantisation is an effect of increased buccal cavity pressure and release that has been phonologised (see Hyman Reference Hyman and Juilland1976, Reference Hyman and Yu2013), then it will produce a synchronic pattern such as the one in Campidanese.

Melodic content is interpreted in spell-out through a mapping between underlying representations and surface forms. Since spell-out is sensitive to prosodic structure, and since each mapping is a look-up function, it can interpret singletons (exemplified in (33a)) and geminates (in (33b)) in phonetically disparate ways.

Something the representations in (33) make clear is that Fortition produces a geminate from voiceless obstruents, without effecting any change to the melodic representation of the underlying input segment. This follows from geminate inalterability: geminates are immune to such feature-changing operations in Campidanese. Put another way, the phonetic expression of geminate voiced stops as spirants is not a fact about the phonological computation senso strictu; rather, it is an effect of the interface. It is irrelevant to the formal computational theory, since the only phonological process in voiced stop gemination is the process that constructs association lines between the melodic material of voiced stops and empty C positions; their realisation as spirants is a fact about interpretation at the interface, not about phonological computation changing any melodic representation. This is in contrast to the process of Lenition which targets voiceless stops. Since Lenition actually changes melodic material through phonological computation, it is part of the phonological grammar, not an interface effect.

5.3.1. Lenition of voiceless stops

While Fortition operates on the prosodic tier, Lenition operates over the melodic tier, targeting /p t ʧ k f s l/ when they are in a [+Gov] prosodic context. Though /p t ʧ k f s l/ share no obvious phonetic properties that would suggest they constitute a natural class, the fact that Lenition can be given a unitary structural description suggests there is a single rule underlying Lenition in Campidanese, effecting a single structural change. The first objective of the computational analysis is to define the structural change that results in /p t ʧ k f s l/ being realised as [].

As noted by Katz (Reference Katz2021: 652), the alternation between voiceless stops /p t ʧ k/ and voiced fricatives [] has posed serious obstacles for output-oriented phonology; such alternations are an example of a saltation, a structural change where a category B is ‘jumped over’ on the way from A $\rightarrow $ C (Lass Reference Lass1997; Minkova Reference Minkova and Marle1993). Hayes & White (Reference Hayes and White2015: 267) define saltation in featural terms: an alternation between segments A and C is a saltation if there is a distinct segment B which shares all the features A and C have in common, but which does not participate in the alternation.

I suggest that the attention placed on saltatory alternations is misguided, and that they do not pose any particular difficulties for synchronic phonological computation. In part, the attention paid to saltatory alternations is based on a conceptual argument, reflected in the intuitions of many phonologists, that a change A to C is somehow more extreme than a change A to B. Given an alternation /p/ $\rightarrow $ [β], the assumption that this is a saltation is made with no evidence that this particular structural change actually involves more than a single change of feature, such that /p/ $\rightarrow $ [b] involves a single change but /p/ $\rightarrow $ [β] involves more than one. The assumption seems to be based on the description of /p/ and /b/ in phonetic terms, but not in phonological ones. The definition of saltation used changes depending on the feature system used; in substance-free approaches which assume no single, universal feature system (Dresher Reference Dresher2014; Odden Reference Odden2022), there is no reason to assume that the structural change that maps /p/ $\rightarrow $ [β] is more than a single feature. The explanation for the relative rarity of saltatory alternations is found in diachrony: saltation is never the result of a single sound change (Minkova Reference Minkova and Marle1993; Lass Reference Lass1997), but rather the result of a cumulative rule telescoping (Wang Reference Wang1968; Hyman Reference Hyman1975; Kenstowicz & Kisseberth Reference Kenstowicz and Kisseberth1977). The saltation pattern of Campidanese has similarly been attributed to an effect of diachrony and contact, rather than a one-step sound change (Lai Reference Lai2020: 251).

5.3.2. The lateral

As discussed in §3.2.1, the alternation between the lateral and its uvular or pharyngeal allophone must also be viewed as part of the larger pattern of Lenition. It is of particular interest because of its phonetically arbitrary nature: nothing in the structural description or the phonetic identity of /l/ explains why it should be realised as a uvular or pharyngeal fricative (Chabot Reference Chabot2021; Scheer Reference Scheer, Honeybone and Salmons2015) .

There is some diatopic variation between the uvular and pharyngeal realisations of the output segment, though Contini (Reference Contini and Andersen1986: 521) describes the output as a distinct pharyngeal fricative. This alternation, illustrated in (34), is present in many, but not all, varieties of Campidanese (see Virdis Reference Virdis1978; Contini Reference Contini and Andersen1986; Molinu Reference Molinu2009). In Genoni, speakers are very aware of this alternation and their production of a uvular or pharyngeal phone, and consider it to be one of the distinctive markers of their dialect of Campidanese.

Molinu (Reference Molinu2009) also provides examples of this alternation in data from her fieldwork in Genoni, some of which are shown in (35):

Given that this alternation takes place in external sandhi at word boundaries, is categorical, and is not explainable through recourse to phonetic functionalism, any argument against it being a phonological effect seems doomed to fail. It must be a grammatical effect since it is arbitrary in phonetic terms, and it must be a phonological process since it occurs in external sandhi and has a well-defined phonological context: intervocalic positions subject to government, as illustrated in (36).

In short, the alternation in (34) and (35) is a product of phonological computation that takes place in the context of [+Gov] positions, a process formalised as (37):

The alternation between /l/ and [ʁ] or [ʕ] is phonetically unnatural, since the structural change seems to have no plausible explanation in the triggering context. This computational change is a demonstration of the power of the computational system, which operates in a way that is insensitive to phonetic constraints – it is substance-free.

5.3.3. The phonological rule of Lenition

The regularity of the pattern exhibited in §3.2.1, the fact that it is triggered in external sandhi, and its well-defined phonological context all militate for its status as a phonological rule; any theory of phonology must be able to treat it as such. What is required is an explicit statement of the general computational properties at work. I have already suggested that Lenition occurs in positionally weak contexts, resulting in the loss of melodic material from segments which undergo it. This accords well with the diachronic observation that lenition follows a clear and well-recognised trajectory towards the eventual disappearance of segments, as expressed in the famous personal communication from Theo Vennemann reported by Hyman (Reference Hyman1975: 165), which provides a teleological definition of lenition as a segment’s progression towards zero (see discussion in Honeybone Reference Honeybone, Carvalho, Scheer and Ségéral2008: 13–14). In the final step of this progression, when the segment is reduced to zero, there is a clear loss of melodic material. To account for this erosion of melodic material, Harris (Reference Harris1990, Reference Harris1994) and Harris & Lindsey (Reference Harris, Lindsey, Durand and Katamba1995) propose that lenition is always the loss of melodic material.

Thus, a segment that finds itself in a [+Gov] context is subject to licensing restrictions: melodic material is removed in this position. Lenition in Campidanese can now be given the formal definition in (38):

This rule says that the feature [fortis] is deleted when it is [+Gov]. In the Campidanese system, then, the loss of a single feature causes three distinct structural changes: /p t ʧ k/ are voiced and spirantised; /f s/ are voiced; and /l/ is realised as a uvular or pharyngeal fricative.