The data used in this paper are drawn from the existing literature as well as my fieldwork conducted in June of 2023 in Biscay (Getxo, Lekeitio) and Navarre (Goizueta), Basque Country. Most of the fieldwork data exemplify processes described in the literature (Hualde 1988; 1991; Hualde & Bilbao 1992; Hualde et al. 1994; Côté 1999; 2000; Hualde et al. 2010; Hualde et al. 2019). Importantly, the presented data confirm some underreported cases regarding the behavior of strident clusters (see Section 3.2). Unless clearly indicated otherwise in a dataset, the data come from the fieldwork. Most of the presented IPA transcriptions have been confirmed by inspecting the spectrograms and waveforms. Example spectrograms generated in Praat (Boersma & Weenink 2024) are given in the Appendix.

Basque exhibits a considerable dialectal variation. However, most of the processes discussed in this article are common to the majority of varieties. Therefore, in the ensuing sections, unless indicated otherwise, the term “Basque” is used to denote the aforementioned dialects spoken in Biscay and Navarre. Since the paper does not aim to compare these varieties, individual examples are not associated with specific dialects.

In the ensuing analysis, apart from the general framework of Optimality Theory (OT; Prince & Smolensky 2004; McCarthy & Prince 1995), it is assumed that affricates are specified as [–continuant, +strident] (Jakobson et al. 1952). The discussion regarding the advantages and disadvantages of the alternative views, among which the most prominent is the non-linear complex segment approach (Lombardi 1990), where affricates are specified for both [+continuant] and [–continuant], is beyond the scope of this article (see LaCharité 1993; Rubach 1994; Clements 1999; Krajewska 2010). It is also assumed that phonological features are organized hierarchically on autosegmental tiers (Clements 1985; Sagey 1986). Finally, by the logic of autosegmental phonology, we are employing the Max/Dep-feature theory in the analysis (Lombardi 1995; 1998; LaMontagne & Rice 1995; Causely 1996; Walker 1997; Pater 1999).

Classic OT (Prince & Smolensky 2004; McCarthy & Prince 1995) has a well-known problem with expressing the directionality of processes. For instance, some languages require that VC₁C₂V clusters be simplified to VCV. The preserved member of the cluster, however, is almost unequivocally C₂ (Wilson 2001; though see Lamont 2015; 2016 for morphologically conditioned simplification). Classic OT is unable to generate this asymmetry since the constraints responsible for the mapping VCCV → VCV do not distinguish between C₁ and C₂. To illustrate this point, let us assume that simplification is driven by *CC (no clusters). The competing faithfulness constraint is MaxC (do not delete). *CC is satisfied by the deletion of either of the consonants. MaxC, on the other hand, equally penalizes the deletion of C₁ and C₂. As a result, under the ranking *CC >> MaxC, VC₁V and VC₂V tie.

In response to such issues, two basic theories are invoked: positional markedness (Itô 1988; Zoll 1997; Steriade 1997) and positional faithfulness (Beckman 1997; 1998; 1999; Casali 1997; Lombardi 1999). The debate whether one is superior to the other and/or whether both are necessary has not been entirely resolved (e.g., Lombardi 2001; Beckman 2004; Rubach 2008; Zhang 2020 for an overview). In the schematic example of VCCV simplification, VC₁C₂V → VC₂V may be a result of either a high-ranked positional faithfulness constraint protecting prevocalic segments (e.g., MaxC_PreV), or a positional markedness constraint militating against coda consonants (e.g., CodaCond). While the former analysis is readily carried out in a parallel framework, the latter necessitates a serial approach (McCarthy 2008).

Apart from the directionality of mappings, generic faithfulness and markedness constraints are unable to distinguish between stem-level, word-level and phrase-level processes (e.g., see Lexical Phonology; Kiparsky 1982; Rubach & Booij 1987) without abandoning a strictly parallel one-step OT model (cf. Derivational/Stratal OT; Rubach 1997; Kiparsky 1997). Many aspects of these domain-driven asymmetries can be captured by positional markedness and/or positional faithfulness relativized to different prosodic boundaries. The current paper shows two disadvantages of the positional markedness approach in the analysis of Basque cluster simplification: one theoretical and one empirical. Although the proposed constraint set involves anchoring constraints, which are not a classic example of positional faithfulness constraints, they still invoke the faithfulness rather than the markedness of constituents associated with specific structures. Therefore, in this sense, anchoring falls within the category of positional faithfulness.

Consider the data in (2) regarding stop deletion in Basque across morpheme and word boundaries.

The data in (2a) exemplify obligatory deletion of plosives before non-continuant segments across morpheme boundaries. In (2b) and (2c), word-final stops are optionally deleted before the following consonant-initial word (see also Hualde 1988; 1991; Hualde & Bilbao 1992).4

Past research treats the changes in (2a) and (2b) either as a dissimilation effect working against adjacent non-continuants (Hualde 1988; 1991) or as an effect of a Coda Condition against coda stops (Artiagoitia 1993). The former approach is insufficient since it does not account for the data in (2c), where stops are deleted before continuants. The latter, on the other hand, creates complications in accounting for word-final stops not followed by consonants, e.g., ba[t] ‘one’, which are not deleted (see Côté 2000 for a critique of the syllabic approach).

The mappings in (2) may be analyzed in OT as an interaction of a markedness constraint militating against non-prevocalic stops (Côté 2000) with MaxC and a positional faithfulness constraint, MaxC_PreV, listed in (3).

The markedness constraint T → V requires that non-continuants must be followed by vowels. The rationale for this constraint comes from the license-by-cue approach (Steriade 1999), where perceptual salience plays a pivotal role. Specifically, prevocalic stops have stronger phonetic cues than non-prevocalic stops (Jun 1995; 2004; Steriade 1999). As a result, the drive to avoid low salience may trigger various processes that conspire against relatively weak segments.

Consider the evaluations in (4) for /bat+naka/ → [banaka] ‘one by one’, /bost pert͡sona/ → [bospert͡sona] ‘five people’, /suret͡sat sormena/ → [suret͡sasormena] ‘creativity for you’ and /bat/ → [bat] ‘one’.

There are two problems with the ranking argument in (4). First, it fails to derive the attested output in (4iv). This is because the driver, T → V, makes no distinction between preconsonantal and phrase-final stops. Therefore, the ranking T → V >> MaxC predicts that preconsonantal stops and phrase-final stops are avoided to the same extent. This is an undesired result since plain stops at the end of utterances are not deleted: /bat/ → [bat], not *[ba]. Second, simplification across morpheme boundaries is obligatory, while simplification across word boundaries is optional. Therefore, in the latter case there is a possibility of reranking T → V below the faithfulness constraints. Such a demotion, however, proves problematic for the obligatory status of simplification across morpheme boundaries, e.g., /bat+naka/ → [banaka], not *[batnaka]. In other words, the model in (4) does not predict any discrepancies between stem-level and phrase-level processes and both must be either obligatory or optional.

In order to salvage the analysis, we may employ a set of prosodically conditioned constraints referencing different prosodic boundaries, as proposed in Côté (2000). Accordingly, T → V may be relativized word (W) or phrase (P) boundary, as summarized in (5).6

As argued by Côté (2000), simplification is stronger the lower the boundary. This relation is expressed in OT by a universal ranking T]_Ø → V >> T]_W → V >> T]_P → V. With the licensing constraints in the picture, phrase-final stops are no longer in danger of being eliminated on a par with morpheme and word-final stops, as shown in (6).

In (6i) and (6ii), simplification is observed since the licensing constraints requiring that stops at no/word boundaries must be followed by vowels outrank the faithfulness constraints. In contrast, no simplification is generated in (6iii), where MaxC is ranked above the driver, T]_P → V.

The split of the licensing constraints also resolves the problem with the optionality of simplification in different domains. While T]_Ø → V must be categorically ranked above MaxC, T]_W → V can be ranked either above or below MaxC, which yields optional simplification across word boundaries, as shown in (7). Since deletion is not observed at the end of Ps, MaxC must always dominate T]_P → V.

Let us now consider the data regarding the behavior of affricates in similar contexts, given in (8).

In (8a), affricates undergo obligatory spirantization before non-continuants across morpheme boundaries. Across word boundaries, spirantization is optional in the same phonological context, as shown in (8b). Finally, the data in (8c) document that the relevant context may also include continuant consonants or a pause (see also Hualde 1988; 1991; Hualde & Bilbao 1992; Hualde et al. 1994).

In order to account for spirantization, we need an additional constraint militating against the deletion of the feature [–continuant], Max[–con]. As shown in (9), the ranking argument developed so far with this additional constraint successfully generates the attested mappings.

With a low-ranked Max[–con], spirantization is observed whenever an affricate is not followed by a vowel. Since spirantization is optional at the end of Ws and Ps, (9i) and (9ii), Max[–con] may be promoted over T]_P → V or T]_W → V. This creates an effect similar to (7).

There are two problems with the analysis in (9). First, the model based on relativized markedness does not yield satisfactory results when tested using Gradual Learning Algorithm (GLA; Boersma 1997), as shown below. Second, the split of the driver C → V into (at least) three constraints effectively splits the process into three separate processes, thus obscuring the conspiracy against non-prevocalic [–continuant] segments.

Consider the results of a GLA model (run in OTSoft; Hayes et al. 2013) based on mock input frequencies (possible vs. impossible outputs) given in (10). Notice that such a model incorrectly predicts that phrase-final deletion of plosives applies on a par with phrase-final spirantization. Specifically, it does not distinguish between /arit͡s/ → [aris] ‘proper name’ and /bat/ → [bat] ‘one’ (not *[ba]).

The ranking values found by GLA based on the inputs given in (10) are listed in (11).

The model summarized in (10) and (11) incorrectly predicts stop deletion at the end of phrases. This is due to the fact that MaxC and Max[–con] are ranked similarly, too close to T]_P → V. As a result, T]_P → V may outrank both faithfulness constraints, generating the unattested /-t/ → *[-Ø].

A solution to both the theoretical and the empirical problem with the analysis outlined above entails shifting the weight of the asymmetry from positional markedness to positional faithfulness. In the ensuing analysis, we are going to employ a set of anchoring constraints that regulate the input-output correspondence between morphological and prosodic edges, as defined in (12).

The constraint in (12) is a template for a family of anchoring constraints. The relevant domains for the data in (2) include morphological roots and affixes under the umbrella term Morpheme in the position of S₁ as well as the prosodic domains of Word and Phrase in the position of S₂ (for a more fine-grained prosodic hierarchy, see Côté 2000). Since Morpheme will be common for all anchoring constraints used in the ensuing analysis, we are going to introduce the abbreviations in (13).

Consider the evaluation in (14), which includes the same input-output pairs as the evaluation in (4). To save space, only the relevant segments are listed.

The tableaux in (14) generate the attested outputs. In (14i), deletion is obligatory due to the ranking T → V >> MaxC. In (14ii) and (14iii), deletion is observed under the ranking T → V >> AnchR_W. However, this ranking may be optionally reversed, yielding no simplification phrase-internally. Finally, phrase-final deletion is banned by an undominated AnchR_P and candidate (14iv-a) is the only possible output. Deletion of prevocalic consonants is barred by an undominated Max_PreV (not included in the evaluation).

In order to account for spirantization in OT, the anchoring constraints in (13), which refer to segmental correspondence, are expanded to encompass individual autosegmental constituents. Consequently, the set in (15) emerges.

The constraints in (15) are a logical extension of the autosegmental correspondence on the one hand and the strength of boundary effects on the other. Consider the evaluations in (16) (relevant segments only). The constraints AnchR[–con]_W and AnchR[–con]_P are collapsed under a single entry AnchR[–con]_W/P.

In (16i), the spirantizing candidate, (16i-b), is preferred over the faithful candidate, (16i-a), due to the ranking T → V >> Max[–con]. Candidate (16i-c), which deletes the entire affricate, is rejected since it gratuitously violates MaxC. Spirantization across word boundaries in (16ii) is governed by the ranking of T → V vis-à-vis AnchR[–con]_W. If the licensing constraint outranks the anchoring constraint, spirantization is observed. Otherwise, the faithful candidate emerges victorious. Finally, phrase-finally in (16iii), spirantization applies under the ranking T → V >> AnchR[–con]_P. In all cases, deletion of prevocalic segments is prohibited by an undominated MaxC_PreV (not listed).

The ranking argument in (16) generates the attested outputs when it comes to plain stops, as shown in (17).

In (17i), deletion is obligatory. In (17ii), deletion depends on the ranking of the licensing constraint and AnchR[–con]_W and/or AnchR_W. Finally, phrase-final deletion is categorically banned by a high-ranked AnchR_P.

The upshot of the ranking argument presented in (17) is that it reflects a true cross-domain conspiracy driven by the requirement that non-continuants must be followed by a vowel. It also yields promising results when it comes to probabilistic constraint ranking, which is particularly insightful with regard to the obligatory vs. optional nature of processes in different domains. Consider the results of a GLA model based on mock probabilities (possible vs. impossible outputs), given in (18).

The ranking values found in this model are given in (19).

As visible in (18), the proposed constraint set is able to generate the input values without generating any unattested outputs in a probabilistic grammar. Comparing the model based on anchoring in (18) and the model based on conditional licensing in (10), it is evident that the former more accurately reflects the input values.8

Strident clusters constitute a special case of cluster simplification. Even though in Section 3.1 we have established that affricates are spirantized in the non-prevocalic position, a different result is observed when it comes to affricate-fricative and affricate-affricate clusters. Additionally, strident clusters straddling morpheme boundaries undergo dissimilation without deletion, which does not occur across word boundaries. Finally, fricative-fricative clusters undergo degemination. Consider the data in (20).9

The data in (20a) exemplify dissimilation of fricative-affricate clusters across morpheme boundaries, where the affricate loses the feature [+strident] in the output. As shown in (20b), affricate-fricative clusters across word boundaries are optionally simplified to a single affricate. Clusters of two affricates across word boundaries, (20c), are also optionally simplified to a single affricate, which preserves the place of articulation of the prevocalic segment. Finally, (20d) exhibits an optional simplification of fricative-fricative clusters (see also Hualde 1991: §5.3, §5.6; Hualde & Bilbao 1992: 20–21; Hualde et al. 1994: 33).

The data in (20) are problematic for the ranking argument developed in Section 3.1. First, recall that the final ranking generating the attested outputs of stop deletion and affricate spirantization involves a high-ranked MaxC_PreV. This, in turn, predicts that clusters in (20b) should simplify by deleting the non-prevocalic segment, /-t͡s s-/ → *[-s-]. Second, the markedness constraint employed in the analysis so far, T → V, is unable to drive the changes in (20a). Specifically, a requirement that stops (or even consonants in general) must be followed by vowels is not relevant in the mapping /-s+t͡s-/ → [-st-].

The patterns in (20), especially (20a), call for a driver militating against adjacent strident segments, OCP[+strid]. We must also consider additional output candidates that would not violate the high-ranked MaxC_PreV. Specifically, a mapping /t͡s s/ → [t͡s] in OT may be interpreted as segmental deletion, /t͡s₁ s₂/ → [t͡s₁], or segmental merger, /t͡s₁ s₂/ → [t͡s_1,2]. In the latter case, the two input segments correspond to a single output segment. A constraint militating against coalescence is Uniformity (Unif; McCarthy & Prince 1995).12 The asymmetry between word-internal and phrase-internal simplification is accounted for by the anchoring constraints targeting the feature [+strident], Anch[strid], which are analogous to the anchoring constraints targeting the feature [–continuant] in (15) of Section 3.1. The tableaux in (21) do not include candidates deleting prevocalic segments; I also assume a high-ranked Max(Place)_PreV, which assures that prevocalic place features are preserved in the output, thus eliminating mappings such as /-t͡s₁ t͡ʃ₂-/ → [-t͡s_1,2-]. The constraints AnchL[strid]_W and AnchR[strid]_W are collapsed under a single entry.

In (21i), the only possible winner is candidate (21i-b), which violates a low-ranked Max[–con] (not listed). In (21ii), (21iii) and (21iv), optional simplification is carried out by the ranking OCP[+strid] >> Unif. The reverse ranking yields no simplification, which is also attested. A high-ranked anchoring constraint militating against the deletion of the feature [+strident] at word boundaries generates the asymmetry between word-internal and phrase-internal simplification. Accordingly, an affricate following a strident segment becomes a plosive only at the stem level, (21i), but not across word boundaries, (21iii).

To conclude this section, let us examine the predictions of GLA for the dataset discussed so far using mock frequencies. Since we are considering coalescing candidates, such candidates are also included for the phenomena analyzed in Section 3.1. In (22), I omit outputs that delete prevocalic segments/Place features (they have been included in the actual model as they affect the final ranking).

The ranking found in the model is given in (23).

The ranking values in (23) do not reflect the hierarchy proposed by Côté (2000), where phrase-internal processes are more likely than phrase-final processes. Specifically, some of the anchoring constraints targeting phrase-final constituents are ranked below the anchoring constraints targeting the same constituents phrase-internally, e.g., AnchR[strid]_W significantly outranks AnchR[strid]_P. This might be due to the mock nature of the model. The goal of the presented GLA models is to test whether the proposed constraint set can theoretically generate all and only attested outputs in a single probabilistic evaluation. Therefore, the quantity of mappings as well as almost categorical frequencies do not reflect the actual grammar. Importantly, the same GLA model with established a priori rankings (Anch_P >> Anch_W) also yields satisfactory results, reflecting the input frequencies in the generated frequencies. This means that the proposed constraint set has a potential of accounting for the real corpus data. Quantitative evaluation of the model, however, falls beyond the scope of this paper.

The GLA model in (23) also attests to the existence of two conspiracies governing cluster simplification in Basque. When considered within a single grammar, the simplification of clusters involving two strident segments is triggered by OCP[+strid], while the remainder of cluster simplification is driven by T → V.

The competing positional markedness model proves inferior also when strident clusters are included, as shown in (24).

The ranking found for this model is given in (25).

It is clear that the results in (24) are not satisfactory. While phrase-final deletion is no longer a problem, the constraint set in (25) generates some unattested mappings when it comes to strident clusters. Specifically, the asymmetry between word-internal and phrase-internal simplification is not captured. The model incorrectly predicts optional simplification across morpheme boundaries, with /-s+t͡s-/ yielding three possible outputs: *[-s₁t͡s₂-], [-s₁t₂-] and *[-t͡s_1,2-]. Additionally, it incorrectly predicts word-initial deaffrication, /-t͡s₁ t͡ʃ₂-/ ~ *[-s₁ t₂-] as well as word-final affricate deletion, /-t͡s₁ s₂-/ → *[-s₂-]. The principal issue with the constraints employed in (24) is that they are unable to distinguish between edge and non-edge stridents. The analysis may be salvaged by employing the relevant anchoring constraints targeting strident segments (Anchor[strid]). However, such an operation seems to create a redundancy from the perspective of the model based on anchoring alone, which yields satisfactory results by using anchoring without the additional set of conditional licensing.

The following subsection verifies the predictions of the positional faithfulness model with regards to vowel epenthesis, another strategy of avoiding non-prevocalic stops in some dialects of Basque.

As discussed in Côté (1999; 2000), Ondarroa Basque exhibits a-insertion instead of consonant deletion in clusters, which constitutes a different strategy of avoiding non-prevocalic stops. Consider the data in (26) adduced in Côté (2000).

Similarly as in the case of cluster simplification discussed in the preceding sections, vowel insertion across morpheme boundaries, (26a), is obligatory. Phrase-internally, (26b), epenthesis is optional and fully faithful clusters may also emerge. Phrase-finally, (26c), stops and affricates are mapped faithfully, although epenthesis may marginally occur.

Côté (2000) successfully analyzes the mappings in (26) using positional markedness constraints relativized to prosodic boundaries (see (1), Section 1). However, since positional markedness proves problematic for the mappings outlined in Sections 3.1 and 3.2, the question remains whether the analysis advocated in the current paper yields satisfactory results when it comes to epenthesis. Consider the evaluation of the mappings in (26) given in (27) under the constraint set proposed in the preceding sections. The constraint Dep-a militates against a-insertion (after Côté 2000). Wiggly lines denote optional reranking; AnchR_P and AnchR_W are abbreviated to AnchR_P/W to save space.

The ranking argument in (27) generates all and only attested outputs. Max[–con] may be optionally ranked below AnchR_P/W in order to generate affricate spirantization instead of vowel epenthesis in (27iii). The ranking T → V >> AnchR[–con]_W assures the avoidance of non-prevocalic stops phrase-internally; the reverse ranking yields the fully faithful outputs in (27iii) and (27iv). Finally, phrase-final faithfulness is guaranteed by a high-ranked AnchR[–con]_P. Marginal epenthesis is possible after the promotion of T → V to an undominated position.

The final processes to be considered are voiced stop lenition and voice assimilation. These processes may interact with stop deletion as well as affricate spirantization and hence must be included in the uniform model. It is shown that such interaction is not problematic for the proposed constraint set in parallel OT. The analysis presented in this section includes the most frequent mappings and hence does not constitute an exhaustive account of voicing effects in Basque.

In Basque, voiced stops are in complementary distribution with their non-strident continuant counterparts, as shown in (28). The former surface before non-continuants or a pause, (28)b; the latter surface when preceded by a [+continuant] segment, (28)a.15

Voiced stop lenition is active across word boundaries (Hualde 2003: §2.1.4.6). Accordingly, word-initial stops undergo lenition when preceded by vowels or continuant consonants, e.g., ze[r d]a → [r ð] ‘what is’.

Another relevant process affecting consonant clusters in Basque is voice assimilation, exemplified in (29).

Based on the data in (29) we can formulate a generalization that voiceless fricatives become voiced before voiced obstruents. An alternative strategy of assimilation is progressive devoicing; however, it is far less common and it is predominantly found with the negative particle [es], e.g., e[s d]ago → [s t] ‘there is no’. Voice assimilation effects are also observed for stops, e.g., du[t b]este → [ð β] ‘have another’; however, since deletion of preconsonantal stops is a prevalent phenomenon in Basque, stop voicing may be less frequent than fricative voicing (Hualde et al. 2019: §6.2). Progressive devoicing of stops, on the other hand, is predominantly found word-internally, e.g., bat+batean > ba[p]atean ‘suddenly’.

Lenition and voice assimilation interact with stop deletion and affricate spirantization discussed in the preceding sections, as documented by the data in (30).

The fact that voiced stops become [+continuant] when preceded by [+continuant] segments in (30) may be attributed to a markedness constraint Lenition (Len), banning sequences of continuants followed by voiced oral stops (cf. Broś 2018 for Spanish lenition in OT). The voicing effects in (29) and (30b), on the other hand, may be analyzed in terms of an additional output requirement that adjacent consonants must agree in voicing, Agree[±vc] (Lombardi 1999). The competing faithfulness constraint is Max[±vc]. The asymmetry between word-internal and phrase-internal assimilation effects is captured by anchoring constraints targeting edge [±voice] features. Finally, we need to include an inventory constraint militating against non-strident fricatives in order to distinguish between stop lenition and stop deletion. Consider the evaluation in (31).

Simplification across morpheme boundaries in (31i) preserves the feature [–voice] on the surface. Effectively, candidate (31i-c) does not violate Max[–vc] as this feature is realized in the output. As established in Section 3.1, the choice between cluster simplification and the fully faithful outputs in (31ii) and (31iii) is carried out by the probabilistic ranking of T → V and AnchR[–con]_W. The candidates preserving the quality of the underlying voiced stops after simplification, (31ii-b) and (31iii-c), are eliminated by Len in favor of candidates (31ii-d)/(31ii-e) and (31iii-d), which exhibit lenition. The remaining candidates, (31ii-c) as well as (31iii-e) and (31iii-f), are excluded by AnchL[+vc]_W or Agree[±vc]. Regressive voice assimilation cum lenition in fricative-stop clusters, (31iv-b), is due to the interaction of AnchL[+vc]_W, Len and Agree[±vc] with the lower-ranked faithfulness constraints.

It is evident that Basque exhibits voice assimilation, but not at all cost. Such assimilation does not yield voiced affricates (due to inventory restrictions; Hualde 1991) and tends to preserve the feature [+voice] rather than [–voice] (at least in fricatives; see Hualde & Bilbao 1992). Furthermore, Agree[±vc] does not trigger cluster simplification, but only the voicing effects. This is a desired result in regards to the too-many-solutions problem (Steriade 2009): from the perspective of language typology, deletion is not a strategy of satisfying the requirements of voice assimilation.

The voicing effects as well as lenition found in Basque entail some additional mappings (e.g., Hualde et al. 2019). However, the intricacies of this topic merit a separate in-depth analysis. Further research is needed to account for the complexity of these data in OT.