Voice Reversals and Syntactic Structure: Evidence from Hittite

We address the relationship between syntactic valency and voice morphology in Hittite ( Anatolian, Indo-European), focusing on cases where active syntax is expressed using non-active morphology, and vice versa . We argue that apparent “mismatches” between syntax and morphology are strictly a morphological rather than a syntactic phenomenon ( contra Alexiadou et al. 2015; Grestenberger 2018). Our study highlights voice “reversals” — i.e., cases in which the expected mismatch disappears and morphological and syntactic valency match. We determine that such reversals correlate with morphological locality, and cannot be derived by hierarchical factors. Our findings provide a novel argument for a uniform syntactic structure of voice (Wood 2015; Wood & Marantz 2018).


Introduction
The morphological expression of valency, i.e., voice, is of interest to syntacticians because it provides a window onto hierarchical structure. In general, the patterns seen in voice morphology can be directly mapped to syntactic trees. In such cases, it can be said that the morphology and the syntax are aligned. Verbs which are marked with "active" morphology have particular (perhaps language specific) syntactic properties, while verbs which are marked as "non-active" exhibit different syntactic behaviors (Embick 1998).
A phenomenon that has played an important role in the discussion of the relationship between voice morphology and syntax are cases of voice mismatches. These are contexts in which the morphology and syntax do not align. One such well-known case is deponency: non-active morphology appears on a transitive verb. Theories diverge on how to account for such misalignments. The crux of the debate hinges on whether such irregularities should be analyzed as truly exceptional properties of morphology (i.e., an idiosyncratic morphological quirk like English irregular past tense morphology), or whether a "strict" mapping between syntax and morphology should be maintained. The latter proposal essentially adopts the idea that the "mismatches" are not mismatches at all, but in fact directly reflect exceptional syntactic structure: deponents surface with non-active morphology because the structure is not that of a transitive verb.
In this paper, we offer an argument from Hittite (Anatolian, Indo-European) in support of the view that apparent mismatches between morphology and syntax are purely morphological in nature. Some patterns of voice morphology cannot be explained by appealing to an exceptional syntactic configuration; rather, they must be encoded as properties of the (post-syntactic) morphological component. Our argument comes from cases of voice reversal in Hittite. As we illustrate below, there are instances in Hittite where the voice Glossa general linguistics a journal of Yates, Anthony D. and John Gluckman. 2020. Voice Reversals and Syntactic Structure: Evidence from Hittite. Glossa: a journal of general linguistics 5(1): 120. 1-39. DOI: https://doi.org/10. 5334/gjgl.1164 morphology expressed on the verb "flips" from mismatch to match. We show that this flip is sensitive to morphological rather than to syntactic properties. That is, the flip cannot be explained by appealing to a hierarchical structure, and can only be explained as a product of how morphology spells-out the (features of the) head that is realized as voice morphology. Our conclusion is specifically an argument in favor of the proposal advanced by Wood (2015) and Wood & Marantz (2018), who argue that the head that introduces an external argument (here called Voice) is always projected, whether it introduces an external argument or not -i.e., so-called "expletive" Voice (Schäfer 2008;Alexiadou et al. 2015). We show below that voice reversal requires the presence of an expletive Voice, even in contexts where it cannot be syntactically or morphologically detected (contra Schäfer 2008; Alexiadou et al. 2015;.
Our paper has the following structure. We begin in §2 by providing the reader with some general background on Hittite. In §3, we turn to the morphological expression of voice in verbs of different syntactic types; we also describe the form and function of the aspectual morphology that in certain verbal classes affects the distribution of voice morphology. These interactions between aspect and voice are the subject of §4, where we show that only mismatch verbs undergo voice reversal. The next two sections are concerned with the analysis of this pattern. We discuss previous accounts of the distribution of voice morphology in §5, then in §6 proceed to lay out our own proposal. §7 concludes.

Background on Hittite
Hittite is the major representative of the extinct Anatolian branch of the Indo-European (IE) language family. Hittite was the official language of the kingdom of Ḫatti, and as such is attested continuously from the 16th-13th centuries BCE in multi-genre administrative texts, the majority on clay tablets excavated from the royal archives at Ḫattuša near modern Boğazkale in central Turkey.
Hittite was written in a cuneiform mixed syllabic-logographic script. All examples below are presented in so-called "broad transcription:" syllabic spellings are rendered in lowercase italics (long vowels marked with a macron); Akkadian logograms in uppercase italics; and Sumerian logograms in uppercase Roman letters (superscripted when functioning as determiners). See Hoffner & Melchert (2008: 9-50) for a more detailed description of the writing system, and Yates (2017: 38-40) for a concise overview of the phonological system that it represents.
The Hittite language is chronologically stratified into three stages, conventionally referred to as Old Hittite (OH), Middle Hittite (MH), and New Hittite (NH). The younger periods are distinguished from the older by linguistic innovations at all levels of the grammar (phonological, morphological, syntactic, lexical, etc.). This paper is concerned primarily with Old Hittite (ca. 1650(ca. -1450, where the voice alternations described and analyzed below are consistently maintained (Melchert 2017: 479-80; see further §4 below). Where these alternations are directly relevant, we therefore focus our study on original compositions produced in the Old Hittite period. However, due to the limited size of the Hittite corpus, our general discussion of voice and aspectual morphology in §3 draws on Hittite texts of all periods.

Voice and aspect in Hittite
This section is organized as follows. §3.1 presents an overview of Hittite voice morphology, describing its formal, functional, and general distributional properties. §3.2 discusses some morphosyntactic diagnostics for unaccusative syntax in Hittite, which are then applied in §3.3 to identify a class of unaccusative verbs that exhibit mismatch voice morphology. In §3.4, we describe a second class of Hittite verbs that exhibit voice mismatch, deponents. Yates and Gluckman: Voice Reversals and Syntactic Structure Art. 120,page 3 of 39 Finally, §3.5 describes the aspectual morphology that -as will be shown in §4 -interacts with the realization of Voice exclusively in these two exceptional mismatch verb classes.

Active vs. non-active voice
Hittite has a bivalent voice system with an opposition between active (act) and non-active (nact) voice, the latter traditionally called "middle" or "mediopassive" (cf. Hoffner & Melchert 2008: 180). Voice is encoded on the verbal stem by a set of fusional inflectional suffixes, which also mark person (1st, 2nd, 3rd), number (sg, pl), mood (ind, imp), and -for indicative verb forms -tense (npst, pst); 1 these suffixes are often referred to as inflectional "endings" because they always attach last to the verb and thus coincide with the right edge of the prosodic word.
The Hittite indicative active endings are given in Table 1 and the non-active endings in Table 2. 2 Note that while all Hittite verbs select the same set of non-active endings, verbal stems with active voice forms belong to one of two conjugational classes, the mi-conjugation (class i) or the ḫi-conjugation (class ii), which are characterized by partially distinct inflectional endings. A verb's class membership is a synchronically arbitrary lexical property. Class-specific endings are indicated as such in Table 1 with superscripts.
It can be observed in Table 2 that non-past tense non-active endings are optionally characterized by a particle -ri. Past tense non-active endings are distinguished from non-past by the presence of a final t, to which a particle -i may be optionally added.
The distribution of active and non-active voice morphology in Hittite is broadly similar to its distribution in other ancient IE languages, such as Ancient Greek, Vedic Sanskrit, Latin, and Tocharian, or in modern IE languages, such as Modern Greek or Albanian (cf. Grestenberger 2014a: 19-62, 102-5;2018: 489-91). 3 Active morphology is found in a wide range of syntactic contexts, whereas the distribution of non-active morphology is considerably more limited. Many verbs with active forms also surface with non-active morphology; for these verbs, the non-active marked verb forms are typically associated with a set of distinct functions, which include reflexives, reciprocals, anticausatives, and passives. These "oppositional" functions of Hittite non-active morphology vis-à-vis active morphology are illustrated in (1-4) below: 4,5 ( The examples in (1-3) show that non-active verb verb inflection is sufficient to mark these oppositional functions in Hittite. However, it should be noted that non-active marking often cooccurs in these functions (other than passive) with the reflexive particle =z(a). See Hoffner & Melchert (2008: 357-66) for a description of this particle, and for more detailed discussion of its relationship with non-active inflection Boley (1993); Garrett (1996);Luraghi (2010;; Cotticelli-Kurras & Rizza (2011;; Inglese (2020: 148-9, 222-8); Melchert (to appear). 5 (1-4) and subsequent examples are generally glossed according to Leipzig conventions. One exception is the use of c for "common gender," which is the standard term in Hittitological scholarship for animate grammatical gender (in contrast to n = neuter). Clause-connecting particles are glossed conn.  In addition, Hittite has a class of verbs that exhibit only non-active voice morphology in their basic stem forms. Members of this class -referred to here as media tantum ("middle only") verbs -largely belong to semantic classes that tend to appear with non-active voice morphology in bivalent voice systems cross-linguistically (Kemmer 1993;Zombolou & Alexiadou 2014;2019, i.a). Such canonical media tantum include the Hittite verbs in (5); the behavior of this class is illustrated in (6) with the stative verb ki-'lie' and the change-of-state verb kiš-'become'.

Unaccusativity in Hittite
Unaccusative verbs in Hittite show special morphosyntactic behavior by which they can be distinguished from transitive and even unergative verbs in certain syntactic contexts (cf. Hoffner & Melchert 2008: 280-3, 310 n. 7). The availability of such diagnostics of unaccusativity is an important feature that sets Hittite apart from the other ancient IE languages as well as from many non-IE languages, where the distinction among intransitive verbs between unaccusative and unergative is difficult to verify beyond lexical semantic classification. As will become clear in §4 below, this distinction plays a crucial role in the realization of Hittite voice morphology: (a subset of) unaccusative verbs show voice alternations between their basic stem form and their "imperfective" forms, whereas unergative verbs never show such alternations. This pattern can be observed precisely because the uniform syntax of these intransitive verbs can be reliably determined using independent criteria. One diagnostic of unaccusativity is the distribution of enclitic pronouns. A unique feature of Anatolian languages like Hittite among the ancient IE languages is the existence of a set of 3rd person subject-marking enclitic pronouns (with gender-specific forms), which are given in (7): The distribution of these enclitic subject pronouns serves as the most important diagnostic of unaccusativity in Hittite. Building on Watkins (1968-9: 93), Garrett (1990b;1996) has demonstrated that only unaccusative verbs co-occur with these pronouns, which are required when a 3rd person subject is not overtly realized by a DP. 6 In other contexts, subject pro-drop is obligatory; the enclitic subject pronouns in (7) thus never occur with agentive verbs, either transitive or intransitive (i.e., unergative), as illustrated in (8-9), where the subject is null: In contrast, unaccusative verbs like the canonical media tantum in (5) above require the 3rd person enclitic subject pronouns in (7) (StBoT 25.25 obv. i 6;Neu 1980: 63) Hittite lacks 1st/2nd person enclitic subject pronouns, and so this diagnostic only applies with 3rd person arguments of unaccusatives verbs. Unaccusative verbs also pattern differently from agentive verbs with respect to auxiliary selection in the formation of the periphrastic (or "analytic") perfect construction. 8 The periphrastic perfect in Hittite is functionally and formally similar to 'have'-perfect constructions in Romance, Germanic, and elsewhere. In Hittite, the lexical verb is realized as a participial form and selects an inflected auxiliary verb, either ḫar(k)-'have' or eš/aš-'be'. The choice of auxiliary verb is syntactically determined: as in (e.g.) Italian or Dutch, agentive verbs employ ḫar(k)-'have', but unaccusative verbs instead use eš/aš-'be'. 9 This morphosyntactic contrast is exemplified in (11)  Due in large part to the relative infrequency of the periphrastic perfect, most unaccusative verbs -including all of the media tantum discussed above -are not attested in this construction. However, in all attested cases the two diagnostics for unaccusativity pattern together: verbs that cooccur with subject clitics also select eš/aš-in the periphrastic perfect.

activa tantum verbs in Hittite
While many unaccusative verbs in Hittite are media tantum, it also has a substantial class of diagnostically unaccusative verbs that surface with only active voice morphology in their basic stem forms. We refer to this class as activa tantum (cf. Grestenberger 2018: 501 for the term).
Syntactically, Hittite activa tantum behave just like canonical media tantum: they obligatorily co-occur with enclitic subject pronouns in the absence of an overt DP subject, as in (17). Once again, this behavior contrasts with that of transitive and unergative verbs like (12-13) above. e. ḫatk-ešš-'become narrow' ⟹ ḫatk-ešš-nu-'cause to become narrow' (e.g., 2sg.npst.act ḫatkišnuši) As evident in (19), causatives derived from activa tantum take active morphology in syntactically active contexts. However, they can also be passivized with non-active morphology; this pattern is attested for ašnu-'take care of' in (19a), e.g., in (20) Thus insofar as Hittite activa tantum pattern with unaccusative media tantum like (5) with respect to the morphosyntactic diagnostics discussed above, they appear to be prototypical unaccusative verbs not just semantically but also syntactically. That is, there is no reason to believe that activa tantum and media tantum differ syntactically. (We will postpone until §5 explicit discussion of what that structure might be.) Morphologically, however, the activa tantum contrast with unaccusative media tantum, the former surfacing only with active morphology in their basic stems forms and the latter only with non-active morphology. As pointed out by Grestenberger (2018: 501), "[f]ormally active unaccusative verbs" of this kind are "a major problem for understanding the canonical distribution of voice morphology" in languages with bivalent voice systems: if it is the case that the class of unaccusative verbs which have identical syntactic characteristics and thus presumably have an identical structural representation are realized with non-active morphology (a widely adopted idea since Embick 1998), then the active morphology seen in Hittite activa tantum would, descriptively, "mismatch" its syntactic context (cf. Weisser 2014). One solution to this problem, proposed by Alexiadou & Anagnostopoulou (2004), Schäfer (2008), and Alexiadou et al. (2006; and adopted by , is that the activa tantum differ syntactically from unaccusative media tantum in a way that is not reflected in the unaccusativity diagnostics discussed previously -specifically, that the former lack the Voice layer present in the latter, which is responsible for the realization of non-active morphology. For these authors, the term "mismatch" applied to activa tantum is a misnomer, because in fact the morphology does reflect the syntax. The details of this proposal are discussed further in §6, where we argue explicitly against this kind of analysis; instead, we contend that Hittite supports the view that unaccusative activa tantum are "mismatch" verbs in much the same way as Hittite deponents, which we discuss now in §3.4. 15

Deponent verbs in Hittite
Although most media tantum in Hittite are syntactically unaccusative (like (5) above), Hittite also has a small class of transitive verbs that surface with only non-active morphology in their basic stem forms. Following Grestenberger (2014a;b;, we refer to verbs with these properties as deponent verbs. Deponents constitute a closed morphological class in Hittite; a complete list of securely attested deponent verbs is provided in (21): 16   (21) Hittite deponent verbs: While deponent verbs differ morphologically from active-marked transitive verbs, syntactically they behave in the same way: deponents take direct objects marked with acc case and never co-occur with enclitic subject pronouns. The examples in (22a-b) illustrate accusative object marking, while (22c) with null subject further demonstrates that deponents do not occur with subject pronouns. c. kinun=a 1 UDU LU-naš kāššaš=(š)aš ḫuittiyanta now=top 1 sheep man.c.gen.sg in.place.of=his draw.3pl.npst.nact 'But now in place of the man they shall drag in one sheep.' (KBo 6.26 i 41) In addition, there is some evidence (albeit limited) to suggest that deponents select 'have' in the periphrastic perfect construction just like active-marked transitive verbs; this pattern can be observed in (23): 17 tantum are approximately as numerous as unaccusative media tantum by type. Whatever its ultimate explanation, this distribution is clearly inherited from Proto-Indo-European (PIE): nearly all cognates of Hittite unaccusative activa tantum in the other IE languages also show active inflection (e.g., 3sg.npst.act PIE *h i és-ti 'is' > Hitt. ēšzi, AGk. esti, Ved. ásti), and similarly, cognates of media tantum show non-active inflection (e.g., 3sg.npst.nact PIE *ḱéy-(t)or 'lies' > Hitt. kitta, AGk. keîtai, Ved. śáye) (see Rix & Kümmel 2001, s.vv). Fundamentally, our point is one of class membership: among the unaccusatives, some are marked with non-active, some are marked with active. We further believe that characterizing the active morphology as "mismatched" in this context offers an explanation for why this class -and only this class -of unaccusatives exhibits voice reversal. 16 See Grestenberger (2014a: 265-76) for an assessment of the evidence. From her list we exclude only ḫanna-, in this respect following Puhvel (1991a: 77-84), who argues that the verb is fundamentally intransitive with the meaning 'litigate; pass judgement' and that its marginal transitive usage (only with internal cognate object ḫaneššar/n-'judgement') is a calque on an Akkadian legal formula (cf. Inglese 2020: 371-4). 17 The interpretation of (23) is disputed. For discussion see Inglese & Luraghi (2020), who argue that (23) is instead an example of the "stative construction" (Hoffner & Melchert 2008: 311-2), which is formally similar to but grammatically distinct from the periphrastic perfect. However, one respect in which deponent verbs differ from active-marked transitive verbs is that they appear to lack corresponding finite passive forms. 18 Grestenberger (2018: 496-7) suggests that this fact should be understood as a morphological blocking effect: passivization is unavailable for deponents because the non-active morphology regularly used to form passives of transitive verbs (see §3.1 above) is already employed in their syntactically active forms. 19 Thus, the only clear difference between deponents and active-marked transitive verbs is in their morphology: deponents exhibit non-active inflection, which in this syntactic context is anomalous within Hittite and cross-linguistically.

"Aspect" in Hittite
The preceding sections identified two classes of Hittite verbs which exhibit a mismatch between syntax and morphology: activa tantum, which are syntactically unaccusative but take active inflectional endings (3.3); and deponents, which are syntactically transitive but take non-active inflectional endings (3.4). These two mismatch classes are bolded in Table 3, which summarizes the relationship between syntax and voice morphology in the Hittite verbal classes discussed above.
In §4 below, we will show that the two mismatch verbal classes in Table 3 also interact in an idiosyncratic way with what are traditionally referred to as "imperfective" suffixes (Hoffner & Melchert 2008: 318). This section provides a concise description of the formal and functional properties of these suffixes. The labels "imperfective" and "imperfective stem" are used descriptively to refer to these suffixes and verbal stems containing them; we return to the issue of their exact grammatical function at the conclusion of the section.
Hittite has four verb-forming suffixes that attach to a root or verbal stem and determine its conjugational class (i or ii). 20 One is the productive causative suffix -nu-(class i), which was discussed already in §3.2 and §3.3 above. The other three are the imperfective suffixes -ške-(class i), -anna/i-(ii), and -šša-(ii) (Hoffner & Melchert 2008: 322-3). The latter two suffixes have a restricted distribution: just a few verbs use -šša-to form their imperfective stems, while -anna/i-is associated especially (although not exclusively) with the deponent verbs listed in (21) above. 21 The suffix -ške-, in contrast, is highly productive: it is regularly employed by the vast majority of Hittite verbs to form their imperfective stems and in this function spreads diachronically at the expense of the other two suffixes. 22 The usage of these deverbal suffixes is illustrated in (24-26): Imperfective verbal stems of the type in (24-26)  in.Luwian as.follows recite.ipfv.3sg.npst.act "While the ritual client is pulling the hair, lashes, and brows from his own head, the Old Woman is reciting in Luwian as follows.' (KUB 32.8 iii 6-10) 21 The distribution of the imperfective markers appears to just be lexically determined. There is some correlation between --anna/i--and deponents, but there are exceptions in both directions, and there is no widely accepted explanation of the correlation among specialists. The suffix -šsă-is likewise hard to explain: it occurs on just four verbs which have nothing obvious in common, semantically ('call', 'make, do', 'get angry', 'help') or grammatically (different conjugational classes, inflectional patterns, etc.). 22 Thus in New Hittite, e.g., walḫ-anna/i-in (25c) is replaced by walḫi-ške-'id.', and beside older ḫalzi-šša-is also attested ḫalzi-ške-'id.' (cf. Hoffner & Melchert 2008: 323). 23 There is a large body of literature on the distribution and semantics of Hittite imperfective verbal stems, especially those formed with the suffix -ške-; see Bechtel (1936); Dressler (1968); Puhvel (1991b); Melchert (1998);Hoffner & Melchert (2002;2008: 317-29); Cambi & Bertinetto (2006); Cambi (2007); Daues (2009;, and Inglese & Mattiola (2020). 24 We use the term "frequentative" in (30) to refer to cases in which the event named by the verb is repeated on multiple distinct occasions, and contrast this with "iterative" in (31), which we use to refer to cases in which the event consists of multiple subevents repeated on the same occasion (cf. Mattiola 2019: 23-4). This usage corresponds approximately to Schultze-Berndt's (2012) distinction between "event-external" and "event-internal iterativity," or to Cusic's (1981) "occasion-external" vs. "occasion-internal pluractionality." Inglese & Mattiola (2020) also make a distinction between "habitual" and "generic imperfective;" under their definitions, (28) would be classified as the latter (the former would include, e.g., pišker in (33) below). Two additional facts about Hittite imperfectives can be observed in these examples. The first is that any of the meanings in (27-31) can also be expressed by a verb's basic stem (i.e., without an imperfective suffix). 25 Thus, for instance, in (30) both penneškanzi ('they drive') in the main clause and turiyanzi ('they hitch up') in the preceding adjoined relative clause function as frequentatives, but only the former contains an imperfective suffix. 26 Further support for this view comes from cases in which a verb's imperfective stem stands in correspondence with its basic stem in different copies of the same Hittite text, e.g., tūriezzi ('he harnesses'; KBo 6.2 iii 60): tūriškezzi ('id.'; KBo 6.3 iii 65) (cf. Inglese & Mattiola 2020); this apparent interchangeability suggests that these forms could be perceived by Hittite scribes as functionally equivalent.
The second is that all three Hittite imperfective suffixes are grammatically equivalent ("suppletive allomorphs of a single morpheme" per Melchert 1998: 414). Their functional identity is supported, in particular, by examples like those in (31), where iterativity is expressed by a -ške-marked verbal stem in (31a) and by -šša-and -anna/i-marked stems in (31b). Their equivalence is confirmed, moreover, by the "supine construction" (Hoffner & Melchert 2008: 322, 338), a type of periphrastic ingressive/inchoative construction meaning roughly 'begin to X' in which a finite form of the verb d(a)i-'place' or tiye/a-'step; take one's place' combines with a verbal noun in its lexical value (i.e., X is the lexical meaning of the verbal noun). This construction requires that the verbal noun be formed from a marked imperfective stem (by adding a nominalizer suffix -wan (sup)); as evident in (32) The range of meanings available to verbal stems marked with these suffixes have led some scholars to analyze them as markers of imperfective aspect (Cambi & Bertinetto 2006;Cambi 2007). Alternatively, Inglese & Mattiola (2020) argue (building on Dressler 1968) that the suffix -ške-in particular is a pluractional marker, which could naturally extend to the other imperfective suffixes. Which of these analyses is correct will have implications for the morphosyntax of Hittite imperfectives. Whereas an imperfective marker would be expected to head the functional projection associated with grammatical aspect ("viewpoint aspect" in the sense of Tenny 1987; Smith 1991), van Geenhoven (2004; analyzes pluractional markers as modifiers of lexical (or "situation") aspect, which occupy a lower structural position (cf. Laca 2002;2004a;b;MacDonald 2008;Travis 2010). We take no strong stance on this issue here. Under the proposal that we develop in section §6 below, the crucial fact about these suffixes is not their hierarchical structure but their morphological linear position, intervening between the verbal root and voice morphology. For the present, then, we retain the labels "imperfective" for these suffixes and "Aspect" for their functional projection, and leave it to future research to determine the precise syntactic status of these suffixes and, in turn, how they come to stand in this position.
Finally, a reviewer asks whether the aspectual suffixes could not be realizations of the verbalizing head v. This is unlikely. The verbalizer v has distinct morphological realizations (the causative -nu-and the inchoative -ešš-, discussed in §3.3). As these are not in complementary distribution with any of the aspectual suffixes, we conclude that they are not in competition to realize the same head.

Interactions between aspect and voice
Having established the formal and functional properties of Hittite voice and aspectual morphology in the preceding section, we now turn to the issue of how aspect affects the realization of voice morphology. As we will show in §4.1, the selection of active or non-active inflectional endings for imperfective stems is in general governed by the same syntactic factors as their corresponding basic stem forms. Yet there are two verbal classes in Hittite that flout this generalization: deponents and activa tantum. In §4.2 and §4.3, we will demonstrate that these two classes, which were shown in §3.3 and §3.4 respectively to show "mismatch" voice morphology in their basic stem forms, exceptionally exhibit an apparent interaction between aspect and voice. Specifically, imperfective forms of these verbs instantiate a pattern that we refer to as voice reversal: the imperfective stem shows the opposite voice morphology as the basic stem in syntactically equivalent contexts. 27 Having presented the data, we then summarize the distribution of voice morphology in Hittite and offer a preliminary interpretation of this distribution in §4.4.

Non-interactions between aspect and voice
We begin with the general case. Transitive verbs whose basic stem forms show active morphology in syntactically active contexts have imperfective stems that also show active marking in these contexts. This generalization is exemplified with the verb p(a)i-'give' in (33), where the basic stem and imperfective stem occur in effectively identical contexts and both receive active endings: give.3sg.npst.act 'In the past they would give one mina of silver, but now one gives twenty shekels of silver.' (KUB 6.2 i 9-10) Similarly, the imperfective stems of the transitive verbs in (34-36b) show the same active marking as their corresponding basic stems in (34-36a). Note, too, that these imperfective stems are also compatible with non-active marking, in which case they exhibit the same set of functions as their basic stems do in non-active contexts -e.g., passive in (34c)  The situation with imperfective stems of unaccusative media tantum verbs is less certain due to the poverty of evidence for this type. 29 What little evidence is available, however, 29 The dearth of evidence for imperfective stems of unaccusative media tantum is partially due to the fact that imperfective stems of stative verbs are rare or unattested in Hittite (Bechtel 1936;cf. Hoffner & Melchert 2008: 318;Inglese & Mattiola 2020). This semantic restriction appears to be characteristic of pluractional markers cross-linguistically (Mattiola 2019: 144). Note, also, that we exclude here iyanna/i-'set out', which is sometimes analyzed as the -anna/i-suffixed imperfective of the media tantum iye/a-'walk' (e.g., Hoffner & Melchert 2008: 318). We follow Puhvel (1984: 328); HW 2 I: 1-4 and Kloekhorst (2008: 375-6) in analyzing this verb instead as the lexicalized, historical imperfective of the PIE root *h i ei-'go', whose Hittite reflex survives only marginally as an independent verb (opp.cit). For discussion of the possible historical implications of its active inflection, see Yates (2018).
suggests that these verbs show the same non-active marking in their imperfective stems as in their basic stem forms, as can be seen in (39) with the -ške-suffixed imperfective of the media tantum eš/aš-'sit down'. A second likely example is (40), where the imperfective suffix -ške-attaches to a reduplicated form of the media tantum verb ištu-'become evident'; the resulting stem išdušduške-exhibits non-active marking just like its base ištu-. Dempsey (2015) has shown that verbal reduplication in Hittite is associated with fundamentally the same set of aspectual functions as the imperfective suffixes, and these two marking strategies often co-occur as in (40) (with semantically "reinforcing" effect per Dempsey). We assume that the diagnostic value of this form is not compromised by the presence of reduplication, which on its own does not otherwise appear to have any effect on voice morphology: for instance, the transitive verb par(a)i- ([pr(a)i-]) 'blow' has a reduplicated stem paripr(a)i- ([pri-pr(a)i-]), both of which take active endings in syntactically active contexts (e.g., 3pl.npst.act priyanzi (KUB 2.3 ii 30): parippariyanzi (KBo 15.49 iv 9)), and similarly -a more direct parallel for ištu--the unaccusative media tantum verb kiš-'happen; become' has a reduplicated stem kikkiš-that likewise shows non-active marking (e.g., 3sg.npst.nact kišari (KUB 1.13 i 13): kikkištāri (KUB 14.5 rev. 10)). 30 In addition, the alternating verb nai/ne-'turn', which in the intransitive sense 'turn oneself' exhibits only non-active inflection, also shows non-active inflection in imperfective forms with this meaning, as in (41). The examples in (34-41) uniformly support the view that imperfective stems exhibit the same voice morphology as do their bases in equivalent syntactic contexts. This distribution suggests that voice is fundamentally independent of aspect in Hittite, the latter having no effect on its realization. To reiterate: aspectual morphology appears consistently with both active and non-active morphology, and the morphological realization of voice is contingent on and predictable from the syntactic context. Before proceeding to the two verbal classes that problematize this generalization, however, it is necessary to briefly discuss the chronology of the preceding examples. In §4.2 below, we demonstrate that in Old and Middle Hittite imperfectives of unaccusative activa tantum show voice reversal vis-à-vis their basic stem forms. It is therefore crucial that (37-38) and (39-40) are attested in Middle Hittite: these examples provide evidence for a synchronic contrast in the older language (i.e., prior to New Hittite) between unaccusative activa tantum and other intransitive verbs -on the one hand, active-inflecting unergative verbs, and on the other, non-active inflecting unaccusative media tantum -whereby only the unaccusative activa tantum exhibit voice reversal. 31 Similarly, §4.3 below shows that imperfectives of deponents show voice reversal in Old Hittite; it is important, then, that (33) also occurs in Old Hittite, since it shows that ordinary transitive verbs with activeinflecting basic stems do not show voice reversal in their imperfective stem during this period.

Interactions between aspect and voice in activa tantum
In contrast to the deponents examined in §4.3 below, the distribution of voice morphology in Hittite activa tantum has been previously studied. Concerning this class, the first important observation was made by Watkins (1969: 72), who pointed out that the very frequent unaccusative verbs pai-'go' and uwa-'come' take only active endings in their basic verbal stem, but almost exclusively non-active endings in their -ške-suffixed imperfectives. Neu (1968a: 86-9) noted the same behavior in the verb ak(k)-'die' and the productive class of deadjectival change-of-state verbs formed with the suffix -ešš-(discussed in §3.3 above).
This pattern of voice reversal has now been systematically treated by Melchert (2017), who established two crucial facts. The first is that it is restricted to unaccusative activa tantum. Unergative intransitive verbs with active inflection like palwai-'shout' and link-'swear' thus exhibit the same active inflection in their imperfective forms as in their basic stem forms, as discussed already in §4.1 above. 32 The other is that all unaccusative activa tantum consistently show voice reversal in their imperfective forms in Old Hittite. 33 Melchert's (2017) survey of the evidence reveals that exceptions to this pattern -i.e., diagnostically unaccusative verbs with active-inflected imperfective forms -are limited to New Hittite texts and copies of older texts produced during the New Hittite period. Table 4 lists imperfective forms of activa tantum attested in manuscripts produced during the Old and Middle Hittite periods. 34,35 31 As observed by Melchert (2017: 482), moreover, imperfective forms of the unergative verbs in (37)(38) show consistent active inflection even into New Hittite (18x total attestations for palwai-, 6x for link-). This behavior contrasts with comparably attested unaccusative activa tantum, most of which show some evidence for voice reversal even in New Hittite, where this pattern is no longer obligatory (cf. §4.2 below). 32 The fact that unaccusative media tantum do not switch to active inflection in their imperfective forms is not discussed by Melchert (2017); for the evidence and discussion see §4.1 above. 33 See Melchert (2017: 482) for arguments that šeš/šaš-'(go to) sleep', which shows both unergative and unaccusative behavior in Hittite, does not falsify this generalization. 34 For the reading in  Table 4 are in fact unaccusatives. It was demonstrated already in §3.3 that when activa tantum participate in the causative alternation they show additional derivational morphology -namely, the causative suffix -nu--which is present both when the derived verb takes active morphology (in active contexts) and non-active morphology (in non-active contexts). The absence of this morpheme in the forms in Table 4 thus argues against analyzing them as passives of transitive verbs. Moreover, if it were the case that the non-active morphology found in voice reversal on activa tantum was actually the expression of the passive, then we would expect to see instances of activa tantum with non-active morphology whether an aspectual suffix is present or not, since passives and imperfectives are, in principle, independent. The fact that activa tantum are systematically marked with active morphology is the defining characteristic of this class of verbs. More importantly, the non-active morphology with activa tantum is only found in the presence of imperfective aspect.
Yet while voice reversal is exceptionless only in the older language, there are also a number of imperfectives of activa tantum that are first attested in New Hittite texts or copies of older texts and show the same voice reversal. Examples of this type are provided in Table  5. 36 In New Hittite, however, these imperfectives which exhibit voice reversal coexist with imperfectives that show the same active inflection as their base, e.g.: ipfv.3pl.npst.act paiškanzi 'they go' (VBoT 74:5); ipfv.3sg.npst.act akkiškezzi 'he dies' (KUB 9.31 iv 45), ipfv.3pl.pst.act gallareške[r] 'they turned out unfavorably' (KUB 5.22:35). Given such variation, it is difficult to determine the grammatical status of voice reversal for activa tantum in New Hittite; however, the exceptionlessness of this pattern in the older language argues that voice reversal was obligatory for this type at least in Old Hittite (and probably beyond into Middle Hittite). 36 In Table 5 (j) and (k), the basic verbal stem (marked with a right asterisk) is not independently attested.

Interactions between aspect and voice in deponents
Like the activa tantum just discussed in §4.2, deponents exceptionally exhibit an interaction between aspect and voice: imperfectives of deponents select active endings in the same syntactic contexts as their base stems select non-active endings. While the evidence for deponents is much less robust (see further discussion below), voice reversal in this class is similarly exceptionless in Old Hittite; imperfectives of deponents attested in original manuscripts of Old Hittite texts are given in Table 6. 37 This voice reversal pattern in Hittite deponents has not, to our knowledge, been previously noticed. This is in all likelihood due primarily to the seriously limited nature of the usable evidence, which is restricted by (i) the fact that deponents are rarer by type and token than unaccusative activa tantum and (ii) the well-established diachronic tendency for deponents to adopt active inflection in their basic stem forms, 38 thus eliminating the mismatch between syntax and voice morphology that is characteristic of this class (cf. §3.4 above).
This tendency can be observed already in Middle Hittite, 39 including for all of the deponents that also have imperfective forms attested at this historical stage; thus while these imperfective forms almost consistently show active inflection just like the Old Hittite forms in Table 6, it is not clear that they are examples of voice reversal in the same sense, since they stand beside basic stem forms with active inflection, and are thus consistent with the principle established in §4.1 above that imperfectives in general exhibit the same voice morphology as their basic stem forms in equivalent syntactic contexts. Examples of such contemporaneous attestation in Middle Hittite include: ipfv.3pl.npst.act paršiannianzi (KUB 24.98:11) (KBo 20.39 l.col. 16') to tuḫš-'cut off'. In the case of iškalla-'pierce', the imperfective is attested in Middle Hittite with active inflection (3sg.ipfv.npst.act iškalliškizzi; KBo 3.21 i 6), whereas active-inflected forms of the basic stem occur first in New Hittite (e.g., 3sg.npst.act iškallai; KBo 6.4 i 39), but their absence during the earlier period may be due simply to accident of attestation.
In addition, a few apparent counter-examples to voice reversal begin to appear in the post-Old Hittite period. The imperfective of the verb šarra-'transgress' shows non-active inflection in its 3sg.npst form in five independent occurrences (e.g., šarrašketta; KUB 36.75 + Bo 4696 i 8), as well as once in its 2pl.npst form (šarriškettuma; KUB 34.75:5). 41 There is also a single example of an imperfective to ark-'mount' with non-active inflection, ipfv.3sg.npst.nact arkišketta (KUB 29.1 i 30). How these non-active forms should be explained is unclear, especially since they occur beside contemporaneous imperfective forms of the same verbs with active inflection, e.g., šarraš in Middle Hittite and ararkiškanzi in New Hittite (both cited above).
The evidence for voice reversal in deponents is thus somewhat less robust than for the unaccusative activa tantum treated in §4.2 above, but the overall picture is broadly similar. The few examples of deponents attested in Old Hittite all show voice reversal. 42 The later evidence mostly accords with Old Hittite, but its diagnostic value is compromised by the emergence of active inflection in the basic stem forms of historically deponent verbs during this period.

Exceptional interactions between aspect and voice
It was established in §4.2 that in Old and Middle Hittite unaccusative activa tantum consistently show voice reversal, their imperfective forms exhibiting non-active morphology in the same syntactic contexts as their basic stem forms unexpectedly show active morphology. 40 The form ararkiškanzi 'they mount' is characterized by both reduplication and the imperfective suffix -škejust like išdušduškettaru 'let it become evident' discussed in §4. have non-active šarrašketta, and that the same manuscript which has non-active šarriškatta (i.e., KUB 13.4 iii 61) elsewhere attests šarreškezzi (KUB 13.4 iii 5) with the same meaning. How this variation should be interpreted is at present unclear to us. 42 The fact that all imperfectives of Hittite deponents are formed with the suffix -anna/i-raises the possibility of an alternative analysis to the one we propose in §6 below -namely, that active inflection is an idiosyncratic property of this suffix (i.e., -anna/i-is +at, as is the case for the deadjectival verb-forming suffix -ešš-). This analysis, however, would predict that imperfectives in -anna/i-are incompatible with non-active morphology. While there are no non-active forms of verbs in -anna/i-attested in original Old Hittite manuscripts, there are two forms in later copies of Old Hittite texts which show that transitive verbs in -anna/i-can form non-active marked passives: (i) ipfv.3pl.npst.nact nannianta 'are driven' (KBo 21.75:10; cf. CHD L-N: 393), the (perhaps lexicalized) imperfective of nai-'turn' (Hoffner & Melchert 2008: 323;contra Kloekhorst 2008: 600); and (ii) ipfv.3sg.npst.nact tuḫšannatta (KBo 9.114 lc. 12) 'is cut off', the imperfective of deponent tuḫš-'cut off' (the context is broken, however, leaving the interpretation somewhat uncertain; see Neu 1968b: 117 n. 7). While it is in principle possible that these forms have been introduced by later copyists, it is simpler to assume that they were present already in the original texts, in which case they would demonstrate that verbs in -anna/i-are indeed compatible with non-active morphology. We therefore reject this analysis. A reviewer alternatively suggests that -anna/i-does not allow idiosyncratic features of the root to "transfer" across to the Voice head. This is, in fact, broadly what we propose below, though we frame it as a consequence of morphological locality. Finally, as we stated in §3.5, all imperfective suffixes are "suppletive allomorphs" of a single head (Melchert 1998: 414). It is therefore not a trivial task to make one allomorph bear idiosyncratic featural information, since, as allomorphs, they are merely distinct phonological realizations. This argues against a different reviewer's idea that there are two syncretic aspectual suffixes pronounced -anna/i-, one of which is associated with idiosyncratic information. This latter solution further raises the question as to why voice reversal can be stated in terms of verb classes, and is not an idiosyncratic property of particular, random, roots. §4.3 then demonstrated that in Old Hittite deponents show the converse voice reversal pattern, their imperfective forms exhibiting active morphology where their basic stems unexpectedly show non-active morphology. In Old Hittite, at least, there is thus a contrast between these two verbal classes and all other morphosyntactic types which, as shown in §4.1, do not undergo voice reversal. Table 7 provides a summary of the distribution of voice morphology in Old Hittite by morphosyntactic type.
Thus far a satisfactory synchronic explanation of voice reversal has proved elusive even for the better studied activa tantum; for a critique of earlier proposals see Melchert (2017: 478-9). Beyond their failure to take into account the chronological distribution of forms, the central problem that faces these accounts (especially Neu 1968a: 89) is, per Inglese (2020: 184), "the lack of synchronic functional motivation of the pattern in Hittite." This issue is starkly illustrated for the activa tantum by (42-43), where the active-inflected basic stem forms of pai-'go' and park-ešš-'become tall' occur in the same passages as their corresponding (bolded) non-active-inflected imperfective forms. For deponent verbs the same is evident in (44), where the non-active inflected basic stem form of parš(i)-'break' similarly cooccurs with its (bolded) active-inflected imperfective:  Examples like (42-44) are striking because there is no obvious contrast between the imperfective and basic stem forms of these verbs modulo the aspectual meaning associated with the imperfective suffix. From a functional perspective, then, the voice reversal observed in these forms appears to be essentially arbitrary. Recognizing this issue, Melchert (2017: 482-4) tentatively suggests a diachronic explanation for the activa tantum, suggesting that the voice alternations seen in Hittite activa tantum -i.e., basic stem forms with active inflection vs. derived -ske-forms with non-active inflection -was inherited from PIE. The evidence cited for this hypothesis, which comes exclusively from Tocharian, is both limited and mixed, and thus in our view insufficient to justify reconstructing the pattern. 43 Yet however Melchert's hypothesis is ultimately assessed, what is perhaps most relevant for the present analysis is that the Tocharian evidence does not support the possibility (suggested by a reviewer) that the Hittite suffix -ške-was at a diachronically earlier stage idiosyncratically associated with non-active morphology (i.e., +dep in terms of the analysis developed in §6 below). There is therefore no historical basis for assuming that Hittite -ške-has any such association, which would not in any case explain why the suffix triggers "voice reversal" only in Hittite activa tantum. Moreover, even if Melchert's hypothesis were correct for the unaccusative activa tantum, the converse voice reversal pattern in deponents would still require a separate explanation.
In the next two sections, we propose a unified account of voice reversal in Hittite. Under our analysis, the fact that only activa tantum and deponents show voice reversal in their imperfective forms is related directly to another unique property of these verbal classes -namely, that their basic stem forms exhibit voice morphology that mismatches their syntactic context. The rationale underlying our proposal is that these mismatches are "fixed" when additional verbal morphology intervenes between the basic verb stem and inflectional endings. In other words, what appears to be "voice reversal" in imperfectives vis-à-vis their basic stem is just the emergence of syntactically expected voice morphology. However, before presenting our own analysis in §6, we first discuss previous accounts of the morphosyntax of Voice in §5.

Previous accounts and issues
Previous work addressing the relationship between voice morphology and syntax has focused especially on similar alternations in Modern Greek. 44 Like Hittite, Modern Greek displays an active/non-active split. Non-active morphology appears with (some) unaccusatives, passives, (inherent) reflexives, and deponents (Mackridge 1987;Embick 1998; Melchert (2017: 482-4) identifies five Tocharian verbs that are consistent with this pattern and two that are exceptions. In Melchert's own assessment, this evidence is "less than compelling," and in any event, he concedes that he has no explanation for the pattern (which could in principle be explained by an analysis along the lines of ours in §6). For an alternative view of the Tocharian evidence see Yates (2018), who argues that non-active inflection in these unaccusative verbs is an innovation in Tocharian, the diachronic manifestation of a "dispreference for mismatches between (voice) morphology and syntax." 44 See also Kallulli (2013) and Trommer (2013) on Albanian, and Grestenberger (2014a;b; on deponency in ancient Indo-European languages (in particular, Ancient Greek, Sanskrit, Latin, and Hittite). 45 Non-active morphology is also found with dispositional middles, which we leave aside here. We will assume following Lekakou (2005) that dispositional middles are derived from passives (in Modern Greek and Hittite). Note that in Ancient Greek, middle and passive forms were morphologically distinct in some tense/aspect combinations (cf. n. 3 above). O Janis metahirizete to leksiko the John.nom use.nact the dictionary.acc 'John uses the dictionary.' In the influential studies of Embick (1998;, Embick proposed that properties of the syntax determine the featural content of the head that realizes voice morphology, observing that the constructions where non-active morphology appears can be stated as a syntactic natural class. The central idea, that non-active morphology appears when v/Voice lacks a specifier, has been adopted in almost all subsequent theoretical work on similar systems. (We use "voice" to mean the morphological exponence, and "Voice" to mean the syntactic projection.)

(46)
Voice ↔ Voice [NonAct] / __ No DP specifier : 102, adapted from Embick (1998 Active morphology is the elsewhere case, as it does not reflect a natural class. This is meant to capture the "normal" cases of active morphology on transitives and unergatives, but also the active-marked unaccusatives (our activa tantum). The proposal in Alexiadou & Anagnostopoulou (2004), Schäfer (2008), and Alexiadou et al. (2006; (followed by Grestenberger 2018) is that active-marked unaccusatives do not merge a Voice layer at all, and so cannot introduce a [NonAct] feature. Active morphology is therefore exponed as a default. Note that this relies on voice morphology being fusional -what is actually being exponed is the head that realizes tense/mood and inflectional agreement features, which may or may not include a [NonAct] feature. The situation is the opposite for deponents, because as syntactically transitive verbs, it is expected that the external argument is introduced in spec-VoiceP, and so the rule in (46) would not apply. In order to derive the non-active morphology on deponent verbs, something must be configured to allow (46) to apply.  accordingly proposes that deponents merge a "low" subject instead of the normal position for external arguments, spec-VoiceP. This leaves spec-VoiceP empty and so the rule in (46) applies. 46 The following structures exemplify when each exponent is realized. 46 The idea behind the proposal of Grestenberger (2018: 503-9) is that diachronically deponents are derived from self-benefactives, proposed to be introduced between vP and VoiceP in an applicative phrase. Over time, the self-benefactive meaning was lost, but the structural position of such arguments was retained. See also discussion and analysis in Spathas et al. (2015).  (47), v is a verbalizing head which introduces an event argument, and Voice's sole job is to mitigate the presence/absence of an external argument and its thematic role. Voice and v (and later, Asp) are related via c-selection. (Not all of these assumptions are necessary on the proposals described in this section or on our own proposal, a point we return to shortly.) Thus, structures which canonically lack an external argument in Modern Greek (i.e., unaccusatives, passives, reflexives, reciprocals) are realized using non-active voice morphology because these are structures in which Voice lacks a specifier. Additionally, deponents, argued to lack an external argument, are also subsumed by the rule in (46). 47 Unergatives and transitives will appear with active voice morphology because spec-VoiceP is filled. Deponents and activa tantum exhibit morphological "mismatches" because the syntax itself is configured so that (46) will or will not apply. In the former case, the transitive subject is merged in a low projection, "feeding" (46). In the latter case, Voice simply is not projected at all, "bleeding" (46). This line of analysis thus proposes to solve the issue of voice mismatch by putting the irregularity into the syntax. More precisely, there is no mismatch between syntax and morphology: the realization of Voice is always determined by the structural configuration. 48 It is easy to see how an account along these lines can prima facie cover Hittite's voice system (cf. Grestenberger 2014a: 103-5; 2018 for explicit discussion). Activa tantum are syntactically unaccusative and deponents are syntactically transitive; the former do not merge Voice while the latter merge a specifier-less Voice. However, it is precisely because the model proposes that the morphology of deponents and activa tantum is syntactically encoded that the proposal cannot extend to Hittite once the interaction with Aspect is considered. Put concisely, it is not possible to derive the voice reversals described in §4 if the deponent and activa tantum classes have unique structural representations.
Consider first deponents. Suppose that the subject is merged low, as in (47d). Although this class of verbs normally surfaces with non-active morphology, in the presence of Aspect active morphology appears. One possibility is that the subject stops in spec-VoiceP as it promotes to the subject position (an idea that Grestenberger 2018: 506 n. 22 considers and rejects). However, this movement must then be blocked in the absence of overt aspectual morphology, or else we expect a voice reversal in all contexts -that is, we expect there never to be deponent verbs since something will always pass through spec-VoiceP (with transitive verbs). This solution amounts to simply brute force stipulating the presence/absence of a specifier of VoiceP.
A reviewer suggests treating Asp as a "thematically raising" head (as in Ramchand 2008). Situated above VoiceP, AspP would then require the deponent subject to pass through spec-VoiceP on the way to AspP. However, this still requires massive stipulation: in the case of imperfectives of non-activa tantum and passives, the subject must be prevented from stopping in VoiceP on its way to AspP, since these verbs are realized with nonactive morphology.
Another possibility would be to try to suppress non-active voice morphology by deleting the [NonAct] feature in the presence of Aspect, i.e., impoverishment. Suppose the rule in (48) were added into the model.  (48) to a particular setting of aspect (e.g., perfective or imperfective). Again, this over-generates. The morphological exponence of voice does not correlate with a particular aspectual category (cf. §4.1). That is, it is not the case that verbs with imperfective suffixes consistently appear with either active or non-active morphology. What we find is that only with particular verbal stems, imperfective suffixes trigger a "flip" from active to non-active, or vice versa.
Activa tantum are likewise problematic for the analysis above. Recall that activa tantum display a reversal from active to non-active voice. Structurally, they are assumed to lack a Voice projection entirely in the analysis above (Alexiadou & Anagnostopoulou 2004, i.a.). Thus, to derive the appearance of non-active voice in the presence of Aspect, we would have to introduce a Voice projection into the structure after Aspect is merged. One way to do this is to have Voice c-select for Aspect. Voice would then be subject to the rule in (46). But for this to work, the appearance of Voice would not be subject to any thematic requirements of the root, but to the appearance of Aspect, whose distribution is independent of the selectional properties of the root.
The issue is not solved by switching the hierarchical order of Voice and Aspect (as discussed in §3.5). In this case, it would be necessary to counter-cyclically merge Voice whenever Aspect is introduced -but only for activa tantum, since in all other contexts, Voice is projected into the structure as a result of the lexical semantics or idiosyncratic properties of the verb. 49 That is, we would need an algorithm to determine when Voice can and cannot be counter-cyclically introduced that is sensitive to the presence/absence of Aspect and the lexical class. 50 In sum, under current models of the structural representation of valency, it is not possible to derive voice reversals in Hittite voice morphology by encoding the alternation into the syntax without massively stipulating a solution or violating foundational principles of syntactic theory. The issue for the approach outlined above can be summarized concisely: because the realization of voice is tied to lexical semantic/idiosyncratic properties of the root and the addition of Aspect does not change those properties, a voice reversal cannot be derived. Given this, we believe that the solution to the problem of voice reversal must be located outside of the syntactic component. We lay out our proposal in the next section.

Our proposal
We understand the central issue to be one of morphological locality: when Voice appears linearly adjacent to a trigger, voice morphology may be subject to contextual allomorphy; when Voice is phonologically separated from a trigger, voice morphology is no longer subject to contextual allomorphy because the environment is not met. For this reason, it must be possible to separate the head that is realized as voice from this idiosyncratic information. This therefore entails that voice morphology is always the exponence of (the features of) a Voice head, which is subject to rules of contextual allomorphy or not. It cannot be the case that voice morphology results from the lack of a Voice head.
Keeping the main architectural assumptions from Schäfer (2008); Alexiadou et al. (2015), and Grestenberger (2018), our core proposal is that Voice is always projected on verbal predicates. In this we follow other authors who argue for an "expletive Voice," such as Schäfer (2008); Alexiadou et al. (2015), and in particular Wood (2015) and Wood & Marantz (2018). Expletive Voice is a Voice head that does not introduce agentive 49 A reviewer thinks that this statement is on the wrong track; we respectfully disagree. On the analyses proposed in Schäfer (2008) and Alexiadou et al. (2015), whether Voice appears with unaccusatives is a function of the properties of the root. Internally caused verbs can (but need not) appear with Voice. Externally caused verbs must appear with Voice. Since these two classes can be defined in terms of the lexical semantics of the root, we believe our characterization is correct. However, even if it is not possible to independently determine which class a verb falls into, then the class information must be listed as an idiosyncratic feature of the root governing when Voice is and is not projected. This is essentially the view that Schäfer (2008: 256 n. 9) adopts. 50 Alternatively, a reviewer suggests that issues concerning counter-cyclically can be circumvented by stipulating that Aspect obligatorily c-selects for Voice, thereby requiring Voice to be present whenever Aspect is merged. However, as the reviewer concedes, we would need a way to make sure that only overt exponents of Aspect have such a c-selectional requirement, since we do not see voice reversal when Aspect is not overtly morphologically exponed. Moreover, we find this solution overly stipulative and ultimately unfalsifiable. There would be no way to independently verify that (overt) Aspect c-selects for Voice in Hittite.
semantics (and sometimes does not project a specifier). We return shortly to discuss why a syntactically and semantically vacuous Voice head should be merged at all. Following the authors cited above, we also assume a late insertion model of morphology, in which syntactic heads are associated with phonological form after structure building (i.e., Distributed Morphology; Halle & Marantz 1993). Because we assume that Voice is always projected, deponents and activa tantum are not associated with unique syntactic trees. All diagnostically transitive verbs have an external argument in Voice. All diagnostically unaccusative verbs lack an external argument in Voice. Voice's featural specifications may be determined by a rule such as (46), or by proximity to a morphological diacritic (i.e., a feature), triggering contextual allomorphy. We formalize this in the following way. If we assume with previous authors that active voice morphology is realized in the absence of a [NonAct] feature, then the following ordered rules derive the surface patterns. The features +dep and +at are class features, which stand for "is in the class of deponents/activa tantum," respectively. We assume that these class features are associated with the root. Our treatment of deponents follows that of Embick (1998 (49c). This is why activa tantum surface with active nonactive morphology. If the rules were "disjunctively" ordered, meaning that only one rule could apply, then (49c) would never apply, since the only way Voice NonAct can appear is by application of either of the two preceding rules. Similarly, if rules (49a) and (49c) swapped ordering, then (49a) could never apply because Voice NonAct will not have been "made" yet. (And moreover, it would be over-written by (49c) anyway, predicting that all activa tantum appear with nonactive morphology). The crucial prediction that this model makes is that in cases where Voice is not adjacent to the triggers +dep/+at, i.e., the environment is not met, "normal" morphology will appear. This is precisely what Hittite reveals, where the realization of an intervening aspectual head causes voice to revert to the morphological form expected from syntactic structure. We assume that this linear ordering comes about by displacement of Voice to T, since voice morphology is fused with tense/mood morphology and appears outside of Aspect. 51 51 We assume (without consequence) that the displacement is post-syntactic, formalized as a reordering rule, e.g., via Local Dislocation. Alternatively, we might attribute this to an effect of head-movement in the syntax, though it would require an explanation for why Asp is skipped. The apparent morphological displacement renders somewhat moot concerns about whether the aspectual heads are truly "aspect" or an event pluralizer, as discussed §3.5. We note, however, that it is not necessary that the linear ordering of the exponents be derived via movement. If in fact Voice were merged above Asp, our arguments still go through because they are based on morphological locality, not structural hierarchy. We have chosen the hierarchical ordering of Asp≫Voice≫ v because this is, in our opinion, the "standard" sequence for an Aspectual head which is not associated with lexical aspect, but we acknowledge that it is not the only possibility. Finally, Dave Embick (p.c.) points out, intriguingly, that it may be possible to segment a voice suffix -ri (see Table 2). If so, then the Voice head actually appears outside of tense morphology, but the intervening tense morphology does not give rise to similar locality effects.
We note that our proposal crucially relies on the assumption that null morphology does not count as a barrier for morphological locality. This is necessary because, on the assumption that an Aspect or v head is always projected into the structure, it does not inhibit the relationship between Voice and the root when Aspect/v is phonologically null for whatever category of aspect that is selected. Since perfective aspect is also null, it does not condition a voice reversal. This can be derived either by assuming that null morphemes are simply invisible for such operations (Embick 2010), or that such morphemes are "pruned" (Embick 2015). 54 It also renders moot whether we treat the functional heads in the verbal domain as related via c-selection, versus a functional sequence (Sundaresan & McFadden 2017). What matters is not whether a particular head is merged, nor where in the narrow syntax the head is merged. In the end, what affects the exponence of voice is whether there is morphological locality between the information associated on the root and Voice. If there is overt morphological material between Voice and the trigger for allomorphy, then the rules in (49b, 49c) cannot apply.
A reviewer understandably asks what the function of Voice is in this system: if Voice does not introduce a syntactic argument nor is overtly exponed, why is this not a violation of Full Interpretation (or equivalent)? A similar question arises for all other works that propose an "expletive" Voice, i.e., a Voice head that does not introduce an Agent thematic role, but is nonetheless morphologically present (Schäfer 2008;Alexiadou et al. 2015;Wood 2015;Schäfer 2017;Wood & Marantz 2018). 55 Wood (2015: 152-4) directly addresses this question. On the basis of Icelandic evidence, he proposes that Voice is always projected, whether or not it makes a thematic contribution or even has morphological exponence. He ultimately suggests that this reduces to selection: T/Asp selects for VoiceP. This proposal is meant to account for the fact that Voice is always projected in full clauses, but may be absent in reduced structures, like nominalizations. 56 In fact, the correlation between higher clausal heads and the presence of Voice is further consistent with  observation that voice mismatches disappear in some contexts. (She does not discuss the case of voice reversal in the presence of aspect observed here.) She argues that mismatches between syntax and morphology are seen only when Voice is projected into the structure. The evidence comes from non-finite forms 54 Note though that the analysis does not depend on whether we analyze the "aspectual" suffixes as true aspect, or some other category (cf. §3.5 above). The rules are sensitive to linearized morphology, not hierarchical structure. 55 It is important to keep in mind that Schäfer (2008) and Alexiadou et al. (2015) distinguish cases of expletive Voice (which may or may not project a specifier) from not merging Voice at all. Expletive voice is found in cases of marked anticausative, while the lack of Voice is found in unmarked anticausatives. Thus in their system, the difference between marked and unmarked anticausatives is again related to which heads have been merged. Though we do not address the difference between unmarked and marked forms in this paper, our proposal would extend to languages that mark such a difference in forms by analyzing the difference as a result of morphological spell-out, rather than which heads are merged. We believe that this improves on a system which idiosyncratically merges a vacuous Voice since, as noted in Alexiadou et al. (2015: 118, n. 32), "While marked anticausatives are lexically forced to appear in the context of Voice, unmarked anticausatives do not occur with expletive Voice presumably for reasons of economy." The problem here is that if it is possible, and in fact more economical, to not merge a semantically vacuous Voice, how can we ever force merge of such a head. (Schäfer 2008: 256 n. 9 suggests that it is a c-selectional property of roots.) This particular problem goes away in our proposal. Whether voice morphology is pronounced reduces instead to the morphological interaction of the head Voice (always present) and the idiosyncratic information available (and visible) on the root/categorizer. 56 In our framing of the discussion, we have adopted the idea from the cited authors that Voice's sole job is to "mitigate" an Agentive argument. But it is important to keep in mind that the external argument introducing head Voice/v has been proposed to provide a number of different functions, including (i) an Agent thematic relation, (ii) case, (iii) "eventivity" (i.e., issues related to inner aspect), and (iv) manner features (Alexiadou & Anagnostopoulou 2004: 119). See also Wood & Marantz (2018: 258f) for the idea that Voice (*i in their terminology) "closes off" the extended verbal projection. It is therefore possible that Voice is required in all verbal structures for one of these purposes, and is not truly a "vacuous" projection.
of deponent verbs -in particular, from cases in which it can be argued that no Voice projection is available, like nominalizations and participial derivations, both of which involve a truncated clausal structure. In such cases, the mismatch goes away, and the morphology of deponents looks like the morphology of a "normal" active verb. Thus in Hittite, for instance, both deponents like parš(i)-'break' and non-deponent transitive verbs like epp/app-'take' use the same suffix -ant-to form verbal adjectives with passive readings (parši-ant-'broken', app-ant-'taken'). 57 The system proposed above can also capture this kind of "leveling" as well: since (mismatched) voice morphology is dependent on a Voice head, and such contexts lack a Voice projection, mismatches will not occur in these contexts. If we further require that the presence of Voice can be the result of selection by a higher head, as suggested in Wood (2015), then these facts follow. 58

Conclusion
A close examination of Hittite's active/non-active voice system reveals that irregularities in voice morphology should be attributed to the morphological rather than syntactic component. In particular, we have shown that voice reversal in Hittite -that is, cases where voice morphology flips to a form that is expected based on the external syntax -must be due to morphological locality to a trigger, rather than to properties of hierarchical structure. Our analysis provides a natural and, we believe, intuitive approach to the relationship between voice morphology and syntax. To the extent that Hittite is representative of active/nonactive voice systems in general, the proposal can straightforwardly be extended to the data in Embick (1998;; Schäfer (2008); Alexiadou et al. (2015), and . However, our study raises a pressing question for future work: there is variation among the Indo-European languages as to whether intervening affixes give rise to similar voice reversals. For instance, Latin famously does not exhibit a voice reversal for deponents when voice morphology appears outside of intervening tense/aspect morphology. Thus, e.g., in (52) the deponent verb hortor 'urge, exhort' shows non-active morphology in the basic form of its imperfective stem in (52a) and also when the past tense-marking (traditionally, "imperfect") suffix -bā-intervenes in (52b) (cf. Embick 2000: 191, 197).