Isomorphism between orthography and underlying forms in the syllabification of the Armenian schwa

Orthographic representations are often derived from phonological analyses or representations, and can even lead to claims about phonological representations (Sproat 2000). In Armenian, many strings of orthographic consonants are broken up by schwas in pronunciation. As a grammatical process, this spelling-pronunciation mismatch is sensitive to a host of phonological, morphological, and morphophonological factors. I systematically catalog these factors, and this systematicity reinforces previous generative arguments that the orthographic form (without schwas) matches the underlying form (without schwas) (Vaux 1998). As for these factors, I argue that, phonologically, the epenthesis is triggered by directional syllab-ification and other syllabification-based constraints, including constraints on sibilant-stop contiguity (Itô 1989). Morphologically, epenthesis respects morpheme boundaries even when the boundary is semantically opaque, whether from prefixation, compounding, reduplication, or pseudo-reduplication. And from the morphophonology, there is evidence that epenthesis is simultaneously a phonological rule. It is an early lexical rule and it interacts opaquely with allomorphy and strata. Thus, this paper argues for a tight integration of orthographic, phonological, and morphological structures (cf. Boersma 2011; Hamann & Colombo 2017).


Introduction
Cross-linguistically, there is ample documentation of vowel epenthesis processes that are sensitive to phonological and morphological factors (Itô 1986;Kenstowicz 1994;Côté 2000;Hall 2011;Silverman 2011).This paper discusses schwa epenthesis in Armenian.We first treat epenthesis as a process that transforms orthographic representations to surface pronunciation forms.The near-perfect systematicity of this process is then a strong argument that schwa epenthesis likewise operates from underlying forms to surface forms, such that the orthographic and underlying representations are isomorphic.The orthography is thus a proxy for the underlying form.
Armenian is an Indo-European language with two standard varieties: Western and Eastern.Eastern Armenian is the official language of the Republic of Armenia, and it's spoken in Georgia, Iran, and Russia (essentially in areas that were part of the Russian and Persian empires), as well as diaspora communities elsewhere in the globe.In contrast, Western Armenian developed as a koine in the Ottoman Empire.After the Armenian Genocide, Western Armenian became a diasporic language spoken in the Middle East, Europe, Americas, and elsewhere.I focus on Western but the generalizations extend to Eastern.
Armenian is primarily a CjVCC language, but the orthography contains large sequences of consonant clusters (Table 1).Many of these orthographic clusters are pronounced with one or more unwritten schwas: <mnal>→ [mə.nɑl] 'to stay'.We argue that orthographic forms (<mnal>) by default mirror the underlying forms (/mnɑl/) by lacking these schwas.This paper provides a near-complete catalog of the different contexts for schwa epenthesis as a productive and psychologically real rule at the interface of orthography, phonology, and morphology (Hamann & Colombo 2017).A sample is shown in Table 1.The overarching analysis is that schwa epenthesis is governed by right-to-left directional syllabification.Mismatches between schwa location and directionality arise from the interaction of phonotactic and morphological constraints whether word-initially or word-medially.These constraints include contiguity constraints in sibilant-stop clusters, syllable alignment to morpheme left-boundaries, and alignment in (pseudo-)reduplicative structures.Schwa epenthesis is a productive and psychologically real rule that applies from the orthography to surface forms.Based on this systematicity, we argue that this rule simultaneously applies from underlying forms to surface forms.The generated schwas are phonologically visible, inserted early in the derivation, feed allomorphy, and interact opaquely with late post-lexical processes.Thus, for words with epenthetic schwas, the orthographic forms are essentially the same as the underlying forms (Vaux 1998:ch3).
The take-away from this paper is thus to document the existence of a systematic process at the interface between orthography, lexical representations, and morphophonology.This process interacts with directional syllabification, phonotactics, morphological structure, and derivational strata.This paper is organized as follows.§2 provides background on Armenian orthography and syllable structure, with a focus on the role of schwas.§3 catalogs the distribution of orthographic clusters and epenthetic schwas in small and large consonant clusters.This distribution is predicted from right-to-left syllabification.§4 catalogs deviations in the placement of schwas, i.e., when a schwa is placed in a location that's not predicted by directional syllabification.These deviations are caused by phonotactic constraints on sibilant-stop clusters, and on left-alignment to morpheme boundaries.§5 discuses the psychological reality of schwa epenthesis as an orthographic and phonological rule, both diachronically and synchronically.§6 teases apart schwa epenthesis from vowel reduction.Conclusions are in §7.The appendix discusses an alternative analysis of schwa epenthesis by Vaux (1998) (Appendix A), and an alternative complete analysis for vowel reduction (Appendix B).
Data is collected from published grammars and dictionaries, published both in English and in Armenian.Data is verified against my native Western intuitions and double-checked on English/Armenian Wiktionary.The supplementary materials provide transliteration software and complete OT tableaux with URs and all our constraints, verified with OT-Help (Staubs, Becker, Potts, Pratt, McCarthy & Pater 2010).1Such tableaux may be useful as pedagogical problem sets.

Background on orthography and syllable structure
This section provides background information on Armenian orthography, specifically how schwas show mismatches in their written vs. pronounced forms ( §2.1).§2.2 clarifies my methodology and the relationship between this paper and the wide body of work on Armenian schwas.§2.3 then discusses Armenian syllable structure, and how Armenian is largely a strict C(j)VCC language with constraints on licit complex codas.

Orthography of schwas
Armenian utilizes its own script, the Armenian script.The script was invented in around the 5 th century for Classical Armenian, the oldest attested written variety of Armenian (Sanjian 1996).The script has continued to be used for the modern standard lects: Western and Eastern Armenian.In the 20 th century, there was a series of spelling reforms in Soviet Armenia (Dum-Tragut 2009).These reforms affected the spelling of words for Eastern Armenian in Armenia, but not for Eastern communities in Iran nor for Western Armenian in the diaspora.For this paper, we utilize the reformed spelling of words because it is closer to the surface pronunciation of words.Transcriptions however are in Western Armenian.We use the transliteration scheme in Table 2. Some pairs of graphemes are homophonous in Western Armenian.To maintain a one-to-one correspondence, we disambiguate these homophonous graphemes via diacritics.Some graphemes have multiple surface pronunciations based on context, including devoicing and post-fricative deaspiration processes in consonant clusters: <aɣp>→/ɑχpʰ/→[ɑχp] 'trash'.These complications do not affect our analysis and we don't discuss them.The nasal /n/ becomes [ŋ] before velars.
We place transliterations in angles <>, underlying forms in slashes //, intermediate representations in double slashes // //, and surface pronunciations are either unbracketed or get [].We place morpheme boundaries in our representations, when useful.
The orthography captures all phonemic contrasts.It is relatively predictable to go from the orthographic representations (ORs) to the pronunciation or surface representations (SRs).The vowel inventory consists of /ɑ,u,i,e,o,ə/.Each vowel has its own grapheme, set of graphemes, or digraphs.The schwa has its own grapheme ը <ə>.But, not all surface schwas are written down in the orthography.Based on the correspondence between the ORs and SRs of schwas, we arrive at three categories, shown in  [mənɑĺ] [kʰíɾ] [kʰəɾél] 'friend' 'to stay' 'writing' 'to write' Some schwas are present in both ORs and SRs: <ənger> ∼ [əŋɡeɾ] 'friend'.We call these 'present' or underlying schwas.Most schwas in the SR are absent in the OR.We call these 'inserted' schwas: <mnal> ∼ [mənɑl] 'to stay'.A special case of schwas corresponds to deleted or reduced high vowels.These high vowels are pronounced when stressed: <kir> ∼ [kʰíɾ] 'writing'.But they are deleted or reduced to a schwa in certain morphological contexts when they are unstressed: <krel> ∼ [kʰəɾ-él] 'to write'.
Cross-linguistically, there is little work on formalizing the derivational relationship between orthography and linguistic structure (Aronoff 1985;Dresher 1994).For the derivational relationship between ORs and SRs, the largest work to my knowledge is Sproat (2000), who argues that ORs represent a level of representation that, depending on the language, can match either the UR, SR, an intermediate representation, or even a much deeper representation than the UR.As for how to derive SRs from ORs, there is some pre-existing work (Giegerich 1999;Wiese 2004;Song & Wiese 2010;Neef 2012;Hamann & Colombo 2017;Hamann 2020), mostly for Germanic languages.
For Armenian, it is commonly argued that the orthography reflects the presence or absence of schwas in the underlying representation (UR) (Vaux 1998).I ultimately argue for this position as well, but I first establish the systematic correspondence of ORs and SRs.I utilize constraints from Hamann & Colombo (2017) for the OR-to-SR mapping, couched within Bidirectional Phonetics and Phonology (BiPhon: Boersma 2011).Given that ORs and SRs are systematically related, I then argue that this systematicity means that URs essentially match ORs ( §5.2; Nevins & Vaux 2008).We focus on the case of 'inserted' or epenthetic schwas because they are the most complex, and later briefly discuss the other categories of schwas ( §5.2.1, §6, Appendix B).We ultimately argue that the three categories of schwa display different phonological behavior.

Methodological problems in Armenian
For Armenian schwa insertion, there have been different proposed analyses to predict the location of these epenthetic schwas, given either the UR or OR as input.Traditional grammars simply list different contexts and locations where these schwas appear without any overarching prediction (Thomson 1989;Bardakjian & Thomson 1977;Bardakjian & Vaux 2001;Սուքիասյան 2004;Dum-Tragut 2009).There is some Soviet Armenian work on predicting the location of schwa epenthesis (Гулакян 1965; Ղարագյուլյան 1974, 1979; Թոխմախյան 1988).Khachaturian (1985) provides a summary of Soviet work.In generative phonology, Armenian schwa epenthesis has been analyzed with left-to-right syllabification (Levin 1985), simultaneous rules (Schwink 1994), left-alignment (Orgun 2000), markedness constraints based on vowel-consonant adjacency (Khanjian n.d.), and right-alignment (Vu 2014).These various analyses are however incomplete and look at only small subsets of the Armenian data.Vaux (1998) is the most developed algorithm or analysis.His system is inspired by sonority-based syllabification of syllabic consonants in Berber (Dell & Elmedlaoui 1985).
Because the data is quite complicated and relatively unknown (outside of Armenology), this paper starts from the data first and incrementally builds up an analysis.This paper acts as an typologically-oriented metaanalysis, a standalone theoretical analysis, and as an empirical catalog of the individual variables involved in schwa epenthesis.We emphasize the empirical nature because there are methodological issues in past work, which we try to address.These issues are UR CONSTRUCTION, REDUCTION-EPENTHESIS CONFLATION, and FACTORIZATION.
For UR CONSTRUCTION, most generative work on schwa epenthesis aims to directly model schwa epenthesis as a process that occurs from URs to SRs.To do so, these analyses implicitly adopt the following premises: a) given the OR, it is predictable to know how to pronounce schwas in the SR, b) the UR should lack predictable information, thus c) the UR should match the OR.Although we ultimately argue that (c) is true, that there is near-identity between the UR and OR (= the OR is a proxy for the UR ), this argumentation has to first establish the truth of premise (a).
Furthermore, URs are an abstract representation that one has to explicitly argue for, and not take for granted.Depending on the language, the UR can either match or mis-match the OR.In contrast, the ORs and SRs are given non-abstract representations.Thus, it is a methodologically simpler task to first establish that there is a systematic relationship between ORs and SRs.After doing so, we argue that the OR is a proxy for the UR such that UR matches the OR.By doing so, we reduce the problem of abstraction, and we increase the replicability or psychological reality of our analysis.
As a reviewer points out, the analysis opens the issue of how URs and schwa epenthesis would be acquired by literate vs. non-literate speakers.The mental grammar of literate speakers would necessarily have a transformation/mapping of ORs and URs to SRs.Such speakers would get the most obvious evidence for needing a schwa epenthesis rule; for them, the OR acts as a proxy to the UR by default.In contrast, the mental grammar of illiterate or pre-literate speakers would only have a UR-to-SR mapping.For both literate and illiterate speakers to acquire schwa epenthesis from URs to SRs, they would need to exploit the predictability of this rule ( §3-4) and its morphophonological behavior ( §5) in order to treat these epenthetic schwas as absent from the UR.
For REDUCTION-EPENTHESIS CONFLATION, most work on schwa epenthesis tends to argue for a single monolithic analysis of the distribution of inserted and reduced schwas.The problem with this approach is that it presumes that the exact same generalizations apply to both types of schwas.In this paper, we tease apart this conflation.We at first focus strictly on the behavior of inserted schwas.By establishing the generalizations that govern inserted schwas, we then find that reduced schwas display subtle differences that merit a separate analysis, which we explore briefly in §6 and in-depth in Appendix B.
For FACTORIZATION, schwa epenthesis is a quite convoluted process that affects a wide range of the Armenian lexicon.It can thus be difficult to isolate the individual generalizations that govern schwa epenthesis.Such individual generalizations are tantamount to epiphenomenal generalizations that arise from our formal grammar.For example, most grammars or analyses of schwa epenthesis provide catalogs of inserted schwas based on the size of the consonant cluster.However, these catalogs don't disentangle or isolate a) the presence of morpheme boundaries, b) the possible presence of reduced schwas, c) and the historic role of vowel syncope.This paper resolves this issue by methodologically cataloging the different contexts for schwas, and by focusing on one individual parameter at a time.The end result is a relatively clearer sense of what are the basic patterns in schwa epenthesis: right-to-left syllabification (avoiding medial Cə syllables), morpheme alignment, and other factors such as the effects of syncope and vowel reduction.
When we do the above three methodological finesses, we end up with two positive outcomes.One is a clearer sense of what is schwa epenthesis in its own terms.The second is a yardstick to evaluate the empirical coverage and theoretical elegance of an analysis.As we discuss later, we find that Vaux (1998)'s analysis has impressive empirical coverage, but that there are two corners of the grammar where it has problems: #CCCCV clusters (Appendix A) and disentangling vowel reduction ( §6).This doesn't mean that we reject Vaux's analysis, but it encourages future work to see how pre-existing analyses can be augmented.

Syllable structure
Armenian has relatively simple syllable structure (Table 4).The maximal syllable is CjVCC.Complex onsets are generally avoided except for C+j sequences, which are arguably complex segments (Vaux 1998:81).Complex codas generally need falling sonority.Some falling-sonority consonant clusters are not licit complex codas, and they can trigger schwa epenthesis.For example, the lateral /l/ can be the second member of a complex coda in native words (Table 5), but it cannot be the first *kʰɑlt.Word-finally, /ɾm/ is a licit complex coda, but /ɾn/ can variably trigger schwa epenthesis for some speakers.Word-medially, /ɾm/ can form a complex coda in native words, while /ɾn/ cannot.Word-medially, /ns/ triggers epenthesis for some speakers.2For /mC/ complex codas, the C has to be a homorganic stop (Vaux 1998:103); otherwise we get epenthesis.For most types of licit complex codas, these codas can be licensed either by full vowels or schwas.Both full vowels and schwas can host a complex coda (Table 6) that is a homorganic nasal-stop cluster, homorganic nasal-affricate, fricative-stop cluster, rhotic-stop cluster, or a rhotic-affricate cluster.Word-medially-medially, neither can host [ɾn].However in native words, there are some clusters that can be licensed by full vowels, but generally not schwas.Full vowels can host [ɾm] and rhotic-sibilants, while schwas generally don't (Table 7).I do not provide an analysis for these asymmetries but take them as given.'to snore' 'to wrinkle' 'to sneeze' 'to sprinkle' Besides complex codas, Armenian licenses a limited set of extrasyllabic appendix consonants.These appendixes are added at the edges of syllables or words, while violating sonority laws.For example the nominalizer suffix -kʰ can be added after complex codas: bɑɾtkʰ 'debt'.This appendix -kʰ is limited to the end of stems (Vaux 1998;Vaux & Wolfe 2009;Dolatian 2020).Other final appendixes exist but are rare, including /s,χ,ʃ/.These appendixes do not interact with schwa epenthesis, so we do not discuss them in depth. 4hroughout this paper, I assume a shorthand constraint *CC σ which encompasses all the above restrictions on licit onsets and codas (1).The exact formulation of this constraint is tangential.A fruitful approach would be to use nuanced projection constraints (van Oostendorp 2011).But for this paper, what matters is knowing what counts as a well-formed syllable.
(1) General constraint on syllabification *CC σ Do not have a syllable margin more complex than what is allowed by the phonotactics of the language.

Distribution of schwas in clusters
Despite the relatively restricted syllable structure, the orthography displays many cases of consonants clusters in ORs that are broken up by schwas in the SR.These clusters are especially common root-initially.I go through a catalog of these clusters as they are found root/word-initially ( §3.1), root/word-medially ( §3.2), and root/word-finally ( §3.3).A common analysis across of all these cases is right-to-left syllabification.
Note that as explained in §2.2, we at first treat the OR as the input to syllabification/phonology. But since the OR is a proxy for the UR (= the OR and UR match), the schwa-less OR inputs in subsequent tableaux can be equivalently replaced with schwa-less URs.For convenience, I also provide the OR-based URs in the tableaux.

Root-initial and word-initial clusters
Root-initially, there are ORs that show long clusters of consonants.But most of these clusters are broken up by one or more schwas in pronunciation.These root-initial clusters range in size from 2 to 6 (Table 1).In this section, the words that we focus on are either monomorphemic roots like <krban>→[kʰəɾbɑn] 'pocket', or consist of a bound root and some category-forming suffix like <mnal>→ [mən-ɑl] 'to stay'.These words are not derived from roots which surface with a high vowel.That is, it is not the case that the schwa in mən-ɑl is derived from the high vowel of a morpheme like *min.The schwas in forms thus are non-alternating.
By focusing on these words, we remove any confounds from vowel reduction. 5We argue that in the transformation of ORs to SRs, these schwas are inserted via directional syllabification.

2C and 3C clusters
The simplest types of root-initial clusters are 2C or 3C clusters.For both types of clusters, a schwa is pronounced after the first consonant.We provide a sample of 2C clusters in Table 8.The data is organized to illustrate how pervasive these clusters are.The first consonant ranges over all possible consonants, while the second consonant is either a plosive, fricative, or sonorant.Complications arise when cluster is sibilant-stop, discussed in §4.1.Similarly for 3C clusters (Table 9), a schwa is placed after the first consonant, forming a closed CəC syllable.Again, these 3C clusters can involve a diverse permutation of possible consonants.In pronunciation, these initial clusters undergo epenthesis via very simple constraints (2).We assume a general constraint *CC σ that is violated by illicit complex onsets or complex codas.As a constraint on the correspondence between ORs and SRs, we utilize the constraint <>/ə/ that is violated whenever a schwa in the SR does not have a corresponding grapheme in the OR (Hamann & Colombo 2017).This constraint is analogous to a DEP constraint between URs and SRs.This constraint is outranked by *CC σ , thus licensing schwa epenthesis.For a 2C cluster like <mnal> 'to stay', a schwa is epenthesized after the first consonant mənɑl.The constraint ONSET prevents epenthesis before the consonant: *əm.nɑl.Similarly for a 3C cluster like <krban>, a single schwa is placed after the first consonant thanks to these constraints: kʰəɾ.bɑn'pocket'.
(2 In the above tableaux, the actual input for the derivation is just the OR.We include the UR for illustrative convenience (and it almost always matches the OR). 6The constraint <>/ə/ only works over ORs and SRs.Interested readers can equivalently replace this constraint with DEP, and turn the OR-SR mapping to a UR-SR one.

4C clusters and larger
For 2C and 3C clusters, syllabifying the ORs is straightforward.But for larger clusters, we see variation in where schwas are pronounced.We argue that the placement is based on sonority and directionality.
In 4C clusters, there are largely two classes of words based on the location and number of epenthesized schwas.In one class of words (Table 10), a single schwa is inserted after first consonant: <CCCCV> → [C 1 əC 2 C 3 .CV].Here, C 2 C 3 form a complex coda.The use of a single schwa after C 1 is predicable because a single schwa is all that's needed to syllable this cluster.Our current set of constraints can derive this (3).11).One schwa is after the first consonant, another after the second.Two schwas are needed because C 2 C 3 cannot form a licit complex coda.For the words in Table 11, C 2 C 3 are a rising-sonority cluster and thus cannot be parsed into a complex coda.A second schwa is thus needed.However, the current constraint system cannot explain the reason why the schwa appears after C 2 , and not after C 3 : <mgrdel> → mə.ɡəɾ.del and not *məɡ.ɾə.del 'to baptize'.Notice that the correct output generates a medial closed syllable ɡəɾ, while the incorrect output generates a medial open syllable ɾə. 7o motivate the location of the schwa, I argue that the relevant factor is directional syllabification (ter Mors 1985;Noske 1985;Itô 1986;Svantesson 1995;Farwaneh 1995).Specifically, syllables are parsed right-to-left.Syllabification aims to fit as many segments into a C(j)VCC template as possible. 8Schwas are epenthesized to syllabify clusters that are not adjacent to a vowel.
To illustrate, (4c) shows a derivation of right-to-left parsing via a rule-based system for [mə.ɡəɾ.del]'to baptize' (Itô 1986(Itô , 1989)).This can be equivalently derived via parallelist alignment constraints (McCarthy & Prince 1993;Mester & Padgett 1994;Crowhurst & Hewitt 1995).The alignment constraint ALIGN-σ-L aims to reduce the number of segments that separate a syllable from the word-initial edge.The analogous constraint ALIGN-R is not shown because it is inactive and low-ranked.9 (4) Deriving two schwas via directional syllabification for [mə.ɡəɾ.del]'to baptize' a. ALIGN-σ-L Assign a violation for every segment that separates a syllable from the left-edge.(Eisner 2000).
14 I have not been able to find larger root-medial clusters.Although word-medial 5C clusters (and larger) are attested, the ones that I have found always involve an intervening morpheme boundary, such as a compound boundary ( §4.2.2).
15 Some speakers pronounce this word as [ɑ.bəs.pə.ɾel].The extra unexpected schwa is due to diachronic syncope from an earlier form with a vowel: <absbarel>.Thus such speakers have a UR /ɑbspʰəɾəl/ with an underlying schwa (see footnote 13).Some authors list this word without any pronounced schwas ɑps.pɾel (Allen 1950:182).This is due to schwa elision in connected speech because of the presence of frication (Hovakimyan 2016)  The next section shows the same result for final clusters.

Root-final and word-final cluster
Root-finally and word-finally, 2C and 3C consonant clusters exist.Larger clusters do not exist.These restrictions are due to historical factors (Vaux 1998:26,83).We again find epenthesis in the expected locations due to the right-to-left syllabification. 16or 2C clusters, most cases involve a falling-sonority cluster that can form a complex coda (Table 17).Here, no schwa is inserted.But there are some words where the 2C cluster cannot form a licit complex coda (Table 18).These clusters are usually <Cn> or <Cr> clusters.Some are <Cɣ>.In standard speech, a schwa is added after the first consonant.In standard speech, a final rising-sonority cluster undergoes epenthesis.But in colloquial speech for some lexemes, the schwa is optional: tʰɑ.ɾən∼tʰɑɾn 'bitter'.It is unclear what are the conditions for this optionality.It is also unclear what is the prosodic status of the final rhotic in these schwa-less clusters.It could be an extrasyllabic appendix or a syllabic rhotic.I leave this variation as an open question.What matters is that these clusters can undergo epenthesis, and the schwa is always after C 1 .17 In 3C clusters, there is limited variation (Table 19).I have found only one case where a schwa is added after C 1 .Most other cases have epenthesis after C 2 .The use of epenthesis in 3C clusters is the norm in both dialects. 18  Table 19: Epenthesis in final 3C clusters:
18 Some 3C words like <asdɣ> 'star' undergo epenthesis in modern Istanbul Western Armenian ɑs.təʁ, but not in Lebanese Western Armenian ɑstχ∼ɑsχ.Further variation for this word is attested in Vaux (2003).Idiolectal variation is attested by Fairbanks (1948:22) who documents a Western speaker who regularly blocks epenthesis in final <VCCr> clusters.
the formation of codas over onsets (i.e., a rime or coda language : Broselow 1992;Farwaneh 1995;Kiparsky 2003).Such epiphenomenal generalizations aren't part of the mental grammar of Armenian reader-speakers.But, they are useful heuristics for linguists.If the location of some inserted schwa violates the above heuristics, then this violation indicates that some other linguistic factor is affecting the schwa's location.In this section, we go through corners of the Armenian grammar which violate the above epiphenomenal properties.These deviations are due to complications from the phonology ( §4.1) or the morphology ( §4.2), specifically involving sibilant-stop clusters or morpheme-syllable alignment.These complications constitute their own constraints on schwa epenthesis.

Phonologically-induced deviations: Sibilant-stop clusters
For root-initial 2C clusters, the generalization was that the schwa was epenthesized after the first consonant.This creates a root-initial open syllable.This generalization is robust with one core exception.Specifically, root-initial sibilant-stop-vowel (STV) clusters undergo schwa prothesis.That is, the schwa is placed before the sibilant.This behavior is cross-linguistically unsurprising (Fleischhacker 2001;Goad 2011).
The sibilant can be s,z and the stop can have any place of articulation (Table 20).If either the stop or sibilant is voiceless in the orthography, then they both surface as voiceless.  1I could not find any native words where C 2 is an affricate.The closest case I could find was a nativized borrowing <sd͡ ʒux> → [səd͡ ʒuχ] 'sujuk'.Most initial zm clusters are pronounced with medial epenthesis: <zmurs> → zə.muɾs 'myrrh'.One class of exceptions is represented by the word <zmaylil> which is pronounced with schwa prothesis: əz.mɑj.lil'to admire'.This is because the initial z here is a reflex of the Classical Armenian prefix əz (Աճառյան 1971:239-40).Synchronically, this morphological segmentation is lost for some cases like əz.mɑj.lil, and the word's pronunciation is simply an exception.

8!
Thus, initial sibilant-stop clusters operate slightly differently from other clusters due to contiguity.Otherwise, the phonology of schwa epenthesis respects right-to-left syllabification.

Morphologically-induced deviations
Previous examples of epenthesis consisted of free-standing roots, or bound roots with suffixes.In more complex types of morphology, we find that schwa epenthesis respects morphological left-edges in prefixation ( §4.2.1,) compounding ( §4.2.2), and reduplication ( §4.2.3).Specifically, we find that the left-edges of roots stay aligned with syllable boundaries at the expense of optimal right-to-left parsing, with limited deviation.We formalize this deviation with alignment constraints.
Note that throughout this section, we often place a morpheme boundary (-) in the orthographic form for illustration, but this boundary is not actually represented in the orthography.

Prefixation
Armenian is primarily suffixing but it does have some productive prefixes.The most productive derivational prefix is the negative prefix ɑn-(Table 26).It can be added to adjectives as a category-preserving prefix, or to nouns and verb roots as a category-changing prefix.The prefix is parsed as a separate syllable before C-initial roots, while its consonant is an onset before V-initial roots.Schwa epenthesis applies in consonant clusters in roots and ignores the prefix (Table 27).If a schwa appears in the root in isolation, then the schwa likewise appears under prefixation.A schwa is epenthesized even if the consonant cluster could've been syllabified with the prefix.22To capture the above dependencies, we can use either cyclicity or alignment.I utilize alignment.The alignment constraint ( 14) requires that the root-initial segment starts a syllable.It is outranked by a constraint against heterosyllabic /C.V/ sequences, thus allowing V-initial roots to syllabify with prefixes (14d). 23he alignment constraint outranks ALIGN-σ-L, thus maintaining the schwas of root-initial schwas in clusters (14e).Vowel reduction ( §6) later provides evidence for the role of alignment that cannot be handled via cyclicity alone.24We likewise find epenthesis in prefixes themselves (Table 28).The verbal negation prefix is t͡ ʃ-.It is prefixed faithfully before a V-initial verb.A schwa is epenthesized before C-initial verbs.For roots with initial consonant clusters, we see schwas both in the prefix and in the root, possibly leading to a sequence of open syllables that would generally be avoided by right-to-left parsing.The positive indicative prefix is ɡand it behaves the same.25The two open-syllable schwas in t͡ ʃə-hə.ɾemfollow straightforwardly from right-to-left parsing, rootsyllable alignment, and onset requirements (Table 15).Such a sequence of two open syllables would be blocked inside roots, but is permitted across the prefix-root boundary.A few other consonantal prefixes exist and they show the expected locations of schwa epenthesis, e.g. the negative adjectivalizers <d->, <tʒ->, and <dʒ-> (Table 29).A minor complication arises from root-initial sibilant-stop clusters (Table 30).These undergo prothesis in isolation.The schwa survives after C prefixes and after the negative ɑnprefix.Note that nz and ns complex codas are generally avoided word-medially for native words ( §2.3), thus they are illicit.The current constraint system captures the data ( 16).The constraint CONT-STV is high-ranked enough to always require that sibilant-stop cluster stays contiguous.This necessitates a violation of ALIGN(M,σ)-L.I show the derivation of a C-prefixed word below.For prefixed ST-clusters, it seems that the sibilant can never become part of a preceding complex coda.For the prefix ɑn-, this is avoided because [ns] is a marked complex coda word-medially.But there is a prefix veɾwhich can attach before consonant clusters (Table 31).When it attached to an ST-cluster, we see epenthesis even though [ɾs] is a perfectly licit complex coda.*veɾs.kɑl'to push' 'to push upward' 'to receive' 'to receive again' 'to feel' 'to feel again' To capture the above margin of the Armenian grammar, I utilize a constraint that is violated if the sibilant of a root-initial ST-cluster forms a complex coda without its stop (17). 26Further complications with sibilants arise in compounding, which we discuss in the next section.
(17) a. *VC-S.T Assign a violation if a root-initial sibilant is part of a sibilant-stop cluster, but forms a complex coda without its stop.

Compounding
We have seen so far that prefix boundaries create syllable structures which violate maximal right-to-left parsing.This is due to a high-ranked constraint that prefers keeping root left-edges aligned with syllables.We see further instances of this in compounding.
In the prefixal data, the use of schwa epenthesis was necessary even if the schwa-less forms would have syllabifiable clusters: <an-tram> → [ɑn-tʰəɾɑm, *ɑntʰ.ɾɑm]'moneyless'.The apparent redundant use of schwas is more visible in compounding because compounds involve linking vowels.
A compound is generally formed by concatenating two stems with a linking vowel -ɑ-(18 ).If the second stem is V-initial, we usually find that the linking vowel is absent, and that the two stems are syllabified together.
(18) a.When the second stem has an orthographic cluster, we find schwa epenthesis in the expected location (Table 32).If the stem or root has a schwa in isolation, then it keeps the schwa in compounding.The schwa is not dropped even though the output would be a syllabifiable cluster: d͡ ʒɑ.ʃɑ.sə.ɾɑh not *d͡ ʒɑ.ʃɑs.rɑh'dining hall'.In fact, there are many near-minimal pairs where the presence of a schwa distinguishes a compound from a root-suffix construction (Աճառյան 1971:279; Ղարագյուլյան 1974:170): ʃɑɾʒ-ɑ-nəɡɑɾ 'cinema' vs. vɑŋɡ-eɾ 'syllables' (Table 33).The location of the schwa in the second root is unmotivated with respect to syllabification.A simple parse wouldn't have used a schwa at all: *d͡ ʒɑ.ʃɑs.ɾɑh.The schwa is necessary because of root-syllable alignment.This is illustrated in ( 19).d͡ ʒɑ.ʃɑ.sə.ɾɑh

*
Furthermore, there is a small handful of high-frequency compounds which lack a word-internal schwa (Ղարագյուլյան 1974:171).For such words, the presence of an internal schwa likely existed in early stages of the language: <ham-a-lsaran>→[hɑm-ɑ-ləsɑɾɑn] 'university', literally 'all-auditorium'.But in the modern language, the schwa is absent: [hɑmɑlsɑɾɑn].The absence suggests that such compounds have become grammaticalized or fossilized from root-root compounds into root-suffix constructions.
When the second stem starts with a sibilant-stop cluster, we see striking behaviors (Table 34).Most cases of such compounds involve exocentric compounds.Here, we do not find epenthesis (Ղարագյուլյան 1979, Սուքիասյան 2004:27-28,61).The second stem's sibilant is resyllabified with the linking vowel: se.vɑs.kest'dressed in black'.27The absence of the above schwa follows from our current constraint system (20).Root-syllable alignment has to be violated in order to maintain contiguity.Epenthesis is unneeded because the sibilant can resyllabify with the linking vowel, without creating a complex coda.( 20 In sum, compounding creates morphological boundaries which are respected by schwa epenthesis at the expense of directional syllabification.Sibilant-stop clusters must stay contiguous, at the expense of morpheme-syllable alignment.

Reduplication and pseudo-reduplication
The last morphological construction we consider is reduplication.In brief, Armenian has prefixing reduplicative morphology.As prefixation, the reduplicant boundary can trigger schwa epenthesis at the left edge of roots.
In citation form, simple verbs consist of a root, theme vowel, and infinitival suffix -l.The canonical type of verbal reduplication involves reduplicating a CVC(C) root on its own (Table 35).The meaning of the derived verb is sometimes compositionally predicted from that of the un-reduplicated word, but not always.The two copies tend to stay identical even if it creates clusters of heterogenous voicing or post-fricative aspriation: tʰoʁ-tʰoʁ-ɑl 'to tremble'.This type of reduplication is 'canonical' in that it is not accompanied by any segmental modifications to either copy of the root.However, it's relatively rare to have canonical reduplication where the root has a non-schwa vowel.What's more common is to reduplicate a verb where the root's only vowel is a schwa (Table 36).The schwa surfaces in both copies, but it is orthographically absent in both copies of the root.Although not as common as monosyllabic CVC roots, there are some rare cases of CVCC roots in verbal reduplication as well as bisyllabic roots (Table 37).The root syllables surface with schwas.For all these varieties of reduplication, the location of the two schwas follows from the interaction of schwa epenthesis and the left-syllable-alignment of roots ( 21).The reduplicant boundary causes a schwa to be inserted at the left-edge of the root.This creates a word-medial open syllable at the expense of directional syllabification.The above data concerns canonical verbal reduplication whereby the the two copies are identical.Another common construction is change the vowel of the second copy into a schwa (Table 38).I call this 'syncopated reduplication'.Diachronically, the missing vowel was present at an earlier stage of the language, but is now exceptionally absent (Աբրահամյան 1959).Orthographically, the schwa is unwritten.[d͡ zepʰ-d͡ zəpʰ-ɑl] 'to throw' 'to throw around' 'to deceive' (intense) 'to coat' (intense) Regardless of the above variation, given the OR and the morphological structure, the presence of the medial schwa is predicted thanks to left-alignment of roots to syllables ( 22).In the above cases of canonical and syncopated reduplication, the reduplicated verb has a lexically-related un-reduplicated verb, i.e., a base verb.The lexicon thus provides evidence that the reduplicated verb is derived from the un-reduplicated verb.However, there are cases of pseudo-reduplication (Table 39), whereby a reduplicated verb lacks a lexically-related un-reduplicated form.[kʰəɾ-kʰəɾ-ɑl] [dən-dən-ɑl] [tʰəpʰ-tʰəpʰ-ɑl] [χəʃ-χəʃ-ɑl] [ɡəm-ɡəm-ɑl] [pʰəs-pʰəs-ɑl] 'to croak' 'to loiter' 'to wing' 'to be stirred' 'to lisp' 'to whisper' The above words are not related to any unreduplicated word.They are thus synchronically underived but reduplicative stems (Stolz, Levkovych & Dewein 2009).I call them pseudo-reduplicative.We can analyse the morphological structure of these words in two ways.One is to treat the pseudo-reduplication as morphologically-triggered, meaning that a verb like kʰəɾ.kʰə.ɾ-el'to croak' contains two semi-identical bound morphs: kʰəɾ-kʰəɾ-el.A two-morph analysis is a common analysis for early work on pseudo-reduplication in Austronesian languages (Clynes 1995), and is sometimes called internal or double reduplication (Buckley 1997).The use of two morphs would trigger schwa epenthesis in order to make the left-edge of the second morph aligned with a syllable (23).In contrast, another approach is to analyze the pseudo-reduplication as phonologically-triggered, i.e., aggressive reduplication (Zuraw 2002;Yu 2004;Bat-El 2006;Stanton 2020;Repetti-Ludlow 2021).In this approach, a word like kʰəɾ.kʰəɾ-ɑl would consist of a single bound root morph /kʰɾkʰɾ/ or [kʰəɾkʰəɾ-].In order to trigger epenthesis in an open syllable kʰəɾ.kʰə.ɾɑl,we would need a constraint like REDUP which would require that sequences of (nearly-)identical substrings become more identical: /kʰɾkʰɾ/ becomes [kʰəɾ.kʰə.ɾ-] and not *kʰəɾkʰ.ɾ-sothat we have two identical substrings [kʰəɾ].For simplicity, I go for the morphologically-based two-morph analysis for the pseudo-reduplication of verbs. 28 Baseless verbs show the need for morph boundaries to trigger schwa epenthesis.Similar data comes from 'corrupted reduplication' whereby the two copies differ in either their onset or coda.For onset corruption (Table 40), the second copy's onset onset has undergone diachronic lenition to a bilabial segment.'to draw' 'to scribble' 'to scribble' 'to scribble' For coda corruption (Table 41), the two codas share a place of articulation but differ in their manner of articulation.Corrupted reduplication involves syncopation in either copy, both copies, or neither copy.A corrupted reduplicated verb can have a base verb or not.
'to coil up' 'to murmur' 'to shine' 'to tremble' 'to squat' 'to scrutinize' The two copies are no longer segmentally identical, but the reduplicative structure is recoverable because of the near-identity between the two copies.As before, the location of the schwa is predictable due to the morph boundary.

Psychological reality of schwa epenthesis
The previous sections went through a variety of phonological and morphological contexts where we see schwa epenthesis.These factors are summarized in Figure 1 with undominated constraints on syllable structure and contiguity.These constraints dominate morpheme alignment.There are various low-ranked constraints on directional syllabification, epenthesis, and onsets. 28Although there isn't evidence for phonologically-based pseudo-reduplication from verbal reduplication, there is some evidence for it from vowel reduction.For example in Western Armenian, for the root word <grgin> [ɡəɾɡín] 'double', its derivatives generally delete the high vowel: <grgn-el> [ɡəɾɡ.nél]'to repeat'.But for Eastern Armenian, Մարգարյան (1997:97) reports that the cognate can be optionally pronounced with a schwa: [kəɾk.nel∼ kəɾ.kə.nel] 'to repeat'.For these derivatives, the root is still a single morph kəɾkn-∼ kəɾkən-.I hypothesize that the optional presence of the schwa is due to phonologically-triggered pseudo-reduplication.More data is however needed.The data focused on the role of schwa epenthesis as a spelling-pronunciation rule, i.e., as a transformation from orthographic forms to surface pronunciation forms.This section discusses the significance of the data in terms of psychological reality, both for the orthography ( §5.1) and for phonology proper ( §5.2).

Diachronic and synchronic status of orthographic rules
The previous sections argued that schwa epenthesis is a systematic process that relates orthographic forms to surface pronunciation forms.This section argues that such a spelling-pronunciation rule has been active since Classical Armenian, and it is acquired subconsciously and productively by native readers.
Modern Armenian is written in the Armenian script, a script that has been used since its invention for Classical Armenian (CA).Many modern words with unwritten schwas are diachronic reflexes of Classical Armenian forms.One could hypothesize that the modern pronunciation forms with schwas descend from classical forms that were not pronounced with schwas.But this is false.
Classical Armenian was also a CVCC language that banned complex onsets and had strict restrictions on complex codas.As an ancient language, we don't know how CA was exactly spoken 2000 years ago.However, at some time between the 8 th and 12 th centuries, Armenians developed a 'traditional pronunciation' for Classical Armenian words (Godel 1975:24;Macak 2017Macak :1039)).Part of this convention is to pronounce unwritten schwas in exactly the same places as their modern forms (Godel 1975:16;Thomson 1989:116;Macak 2017Macak :1043)).Some of these unwritten schwas are diachronic reflexes of Proto-Indo-European full vowels that got reduced in Proto-Armenian (Vaux 1998:26).For Classical and Modern Armenian (24), there is no synchronic evidence of an unreduced vowel in the underlying form for these unwritten schwas.
Trans.CA (Godel 1975:116) Modern Western Armenian Thus, these epenthetic schwas are diachronically stable.They have been pronounced but unwritten for at least 1,000 years.There have been a few attempts at formalizing the rules for pronouncing these unwritten schwas for Classical Armenian (Hammalian 1984;Schwink 1994;Pierce 2007).Pierce (2007) has noted that as a spelling-pronunciation rule, essentially the same schwa epenthesis rules are active for Classical Armenian and for Modern Armenian.
For the Modern Armenian grammar, all the previous sections showed that the spelling-pronunciation rule for schwa epenthesis is a systematic rule for converting orthographic forms to pronunciation forms.As a rule, the question is then how this rule is taught or transferred by native speakers.In my experience, L1 textbooks written in Armenian don't provide rules for schwa epenthesis, but at most just list out different cluster sizes and their schwas. 29The Armenian teachers who I have consulted told me they don't explicitly teach any schwa epenthesis rules.Thus, school children pick up the rules for schwa epenthesis without ever getting explicit instruction on how to apply schwa epenthesis in small or large clusters.It thus seems that schwa epenthesis, as a spelling-pronunciation rule, is passively acquired by literate Armenian speakers.This rule is not taught.
Given that this spelling-pronunciation rule is acquired and not taught, the next question is whether this rule is synchronically active in the mental grammar of a literate speaker.Evidence suggests this rule truly is productive.Essentially, if an adept reader is given the OR of some unknown form, then the reader will pronounce the SR with schwas placed in the expected locations.For example, Vaux (1998:66) gave a native speaker a list of almost 100 attested Armenian words that had long consonant clusters.The speaker only knew 20% of these words, but he correctly pronounced all but 7 words.Of those 7 'mistakes', his mistakes were because he gave the wrong morphological parse for these unknown words.Similar results were obtained elsewhere (Հովհաննիսյան 2014:ch6).
Thus, as a rule for relating orthography to phonology, schwa epenthesis is a stable and psychologically real grammatical process.As we explore in the next section, all of this data is a strong argument that the underlying representations match the orthography by lacking these epenthetic schwas.

Phonological status of schwa epenthesis
As a spelling-pronunciation rule, schwa epenthesis is productive.The next question is whether schwa epenthesis is just restricted to being a spelling-pronunciation rule, or whether it is also a phonological rule.I argue that schwa epenthesis likewise resides in the transformation of URs to SRs.The rule is an early lexical rule that creates phonologically visible schwas and that creates opacity.

Epenthesis is a phonological rule
Within the linguistic literature on Armenian, there is debate on whether these unwritten schwas are present in the lexical representation of words (Խաչատրյան 1966, 1988;Khachaturian 1985) or not (Hamp 1963;Гулакян 1965;Vaux 1998).If present, then the phonological grammar would not need a schwa epenthesis rule.Like Vaux (1998), I argue that for ORs that have an unwritten schwa, the underlying representation (UR) likewise lacks this schwa.I provide two pieces of evidence for this.The first is that schwa epenthesis applies in nativized loanwords to create maximally CVCC syllables (25a), cf. to native words (25b) (Vaux 1998:67).The lack of epenthesis is seen as typical of non-nativized loanwords (Ղարագյուլյան 1974:64-5).
(25) a. <t'rak 't'or> [tʰə.ɾɑk.toɾSecond, the pronunciation and placement of these unwritten schwas is predictable and governed by morphological parses. 30If schwa epenthesis is sensitive to such morphology, then this highly suggests that the OR has a close connection to morphophonological representations (= URs).
Because of these factors, I argue that schwa epenthesis applies both from ORs to SRs, and from URs to SRs.This means that ORs and URs are largely identical, putting aside some voicing mismatches in clusters.
epenthesis for different types of clusters.But for clusters of size 3 and above, these textbooks treat epenthetic schwas as random and as something to be memorized (Bardakjian & Thomson 1977).This is further evidence that schwa epenthesis is an acquired rule, not a taught rule.Furthermore, the only time I see any sense of instruction is when children are taught that they need to hyphenate words that are broken across lines: <tram>∼[tʰəɾam] 'money' is hyphenated as <tə-ram>.But this hyphenation is then just teaching readers to write words as they are pronounced.
30 Kiparsky (2018) distinguishes three levels of representation (UR, lexical representation, SR).If we apply his system, these unwritten schwas can be considered as absent from morphophonemic representation (URs) and as not M-phonemes.These schwas are added in the derivation during the stem-level stratum.The outputted schwas are thus phonemes in the lexical representation (L-phonemes).
The same morphological and phonological constraints from the previous sections control the distribution of epenthetic schwas, both from ORs and from URs.One would only need to add a DEP constraint that controls the correspondence between URs to SRs.This DEP constraint would be equally ranked as our <>/ə/ constraint that governs between ORs and SRs.
In contrast to unwritten schwas, a schwa is written in the OR when its pronunciation is not predicted by morphology or by directional syllabification.Diachronically, these written schwas are often a reflex of sporadic unpredictable vowel reduction, e.g., the Modern Western word <ənel>→ [ənel] 'to do' descends from an earlier form <anel>. 31 Synchronically, these written schwa must likewise be present in the UR: /ənel/.
To illustrate, consider the near-minimal pair in (26).These derivations show how the SR can be derived from either the OR or UR, mitigated by UR-to-SR constraints (MAX, DEP) and their analogous OR-to-SR constraints (<X>//, <>/ə/).The MAX-like constraint for ORs is taken from (Hamann & Colombo 2017:690) nə.ɡɑɾ * For the phonology, these epenthetic schwas are not added late in the phonological derivation, but must be inserted early, as a lexical rule.For example, the Armenian plural suffix displays syllable-conditioned allomorphy (27a): -eɾ after monosyllabic bases, -neɾ after polysyllabic ones (Vaux 2003).Epenthetic schwas are counted in the plural (27b 31 There are a handful of exceptions where an unpredictable schwa is present in the UR but not in the OR: footnotes 13, 15. 32 For CVCC with final rising-clusters, schwas appear in isolation: <vakr>→[vɑ.kʰəɾ]'tiger'.These roots are treated as monosyllabic in the plural, take the -er suffix, and don't use a schwa: [vɑkʰ.ɾ-eɾ]'tiger-PL'.For space, I cannot discuss how schwa epenthesis behaves in these words.One could either argue that schwa epenthesis in final clusters applies late due to root-final extraprosodicity (Vaux 1998(Vaux , 2003)), because of a cyclic rule of schwa deletion (Svantesson 1995;Burzio 2005), or due to the interaction of syllabic and trapped consonants (Scheer 2008).This issue does not affect my analysis.
Besides allomorphy, other evidence for the relatively early timing of phonological insertion of schwas includes the involvement of schwas in language games (Հովհաննիսյան 2014;Hovhannisyan 2011) and perception (Hovakimyan 2016).In sum, schwa epenthesis is a phonological rule that applies early in the derivation, early enough to interact with morphological structure and license allomorphy.Further evidence of its derivational timing is presented next.

Opacity of schwa epenthesis
Schwa epenthesis is a phonological rule that applies early in the derivation.This section shows that variable articulatory weakness of the schwa has led to the emergence of a phonological process of schwa elision in connected speech.This post-lexical process has created opacity with schwa epenthesis and allomorphy.The opacity is further evidence of the early and phonological nature of schwa epenthesis.
This process of schwa elision applies late in the phonological derivation as an optional post-lexical rule.Evidence for the lateness of deletion comes from allomorphy and opacity (Table 42).An inserted schwa can license the polysyllabic-selecting plural suffix -neɾ, and then end up deleting due to schwa elision.[tʰə.ɾɑm] [tʰ(ə).ɾɑm-neɾ] [kʰə.ʃoʁ] [kʰ(ə).ʃoʁ-neɾ]'money' 'monies' 'driver' 'drivers' The interaction of epenthesis, allomorphy, and elision leads to inter-stratal opacity (Kiparsky 2000;Bermúdez-Otero 2011) caused by input-driven phonologically-conditioned allomorphy (Paster 2006(Paster , 2009)).I illustrate a simple cyclic derivation in Table 43.Epenthesis and allomorphy apply in the lexical stratum.The presence of the epenthetic schwa licenses the plural allomorph -neɾ.Post-lexically, this same schwa is optionally deleted, making the choice of allomorphy opaque.The above patterns show that the epenthetic schwa is epenthetic, not intrusive (Vaux 2003;Hall 2003, For the Direct Reduction approach, the location of the schwa is predictable simply because the schwa is the output correspondent of the high vowel.But in the Deletion+Epenthesis approach (30), the clearest way to force the location of the reduced schwa is to argue that there is an anti-contiguity constraint (Alderete 2001).This anti-contiguity requires that if some segment u (the high vowel) was deleted between x, y, then some element (the schwa) must get epenthesized in the same location as u.The end result is that this constraint favors the reduced schwa surfacing in the same location as the deleted high vowel.
Within a rule-based system, Vaux captures the above unexpected reduced schwas by positing a difference between schwas epenthesized via cyclic syllabification (what we call inserted schwas), vs. schwas inserted via post-cyclic syllabification (Vaux 1998:ch3.3.4).He posits that vowel reduction is high vowel deletion, and feeds post-cyclic syllabification, and post-cyclic syllabification prefers creating new syllables instead of modifying previous syllables: //.ɑz.nǐ.vu.tʰjún// → //.ɑz.{n}.vu.tʰjún// → [.ɑz.nə.vu.tʰjún]where {} marks unsyllabified material.Thus, it is technically possible to derive reduced schwas via epenthesis.But to do so, we need additional constraints that conspire to make the location of the reduced schwa match the location of the original vowel, even at the expense of right-to-left syllabification.This causes us to question why the child learner should posit that that these reduced schwas are epenthetic rather than just the output correspondents of the high vowel.
There are additional cases where the reduced schwa is used in locations where schwa epenthesis does not predict, but in the location of the original high vowel.For words that have a /VCvČC-V/ context, the high vowel is replaced by a schwa even if deletion would've created a syllabifiable consonant cluster (Table 48).Consider the near-minimal pair below between a /VCvČC-V/ context and a /VCCvČ-V/ context.In both words, deletion would create a syllabifiable fricative-stop cluster.Deletion is found in (31a) but not (31b).For Vaux's system, we would require that the post-cyclic syllabification algorithm can create new simplex codas, but not new complex codas; similar band-aids would need to be used for our Deletion+Epenthesis system.
(32) a. súk 'grief' As is clear, for the Deletion+Epenthesis approach, additional mechanisms are needed to force schwa epenthesis to place the reduced schwas in the same location of the high vowel.33I do not formulate such constraints for space.But the need for such constraints reinforce the fact that a single monolithic analysis for inserted schwas and reduced schwas obfuscates the fact that the two schwas display subtle empirical differences.In Appendix B, I flesh out a Direct Reduction approach which handles the unexpected schwas.

Cyclicity vs. morpheme alignment
By distinguishing vowel reduction from schwa epenthesis, we likewise find evidence that Armenian phonology requires both cyclicity and morpheme-syllable alignment.
The schwa epenthesis data did not allow us to distinguish between a cyclic vs. alignment account for these post-prefix schwas.However vowel reduction provides evidence that alignment is the main factor.Consider the CVC roots in (33).These roots get prefixed to form a VC-CVC derivative.These prefixed forms then get suffixed.Suffixation triggers vowel reduction.In the suffixed forms, the root's high vowel is reduced to a schwa.Such a schwa is predicted by morphemealignment (34a).But within a simple cyclic framework where, the input to suffixation is prefixed word [VC-CVC] because of the semantics.If we assume bracket erasure (Orgun 2002), then this prefix boundary should be invisible, meaning there wouldn't be a morpheme-alignment constraint (34b The above prefix data act as a bracketing paradox because a) vowel reduction is triggered by suffixation, but b) the output of reduction is sensitive to the internal prefix-root boundary (Pesetsky 1979(Pesetsky , 1985;;Newell 2019).In terms of semantic scope, the suffix has scope over the prefix.The vowel reduction data thus cannot work with an analysis that uses bracket erasure to remove prefix boundaries.34See Elkins (2020) for crosslinguistically similar data on prefixes.
Thus, by disentangling schwa epenthesis from vowel reduction, we find that vowel reduction explicitly requires morpheme-alignment in order to predict the location of reduced schwas in prefix-root-suffix constructions.

Conclusion
Orthographic representations can often encourage phonological analyses, or inversely be motivated by a phonological analysis.This paper documented schwa epenthesis in Armenian, as a spelling-pronunciation rule or as a transformation from orthographic forms to surface forms.The rule targets consonant clusters that can range in size and location.The location of the epenthetic schwa is motivated primarily from directional syllabification, with deviations caused by sibilant-stop contiguity and morpheme-syllable alignment.
As a process, schwa epenthessis is passively acquired by literate speakers, is productive, and is psychologically real.Because of its systematicity, schwa epenthesis is simultaneously a part of the phonology proper.It applies not only from orthographic forms to surface pronunciations, but also from underlying forms to surface forms.The orthography is thus a proxy for the underlying representations.Schwa epenthesis is integrated tightly into the language's morphophonology, can interact with allomorphy, and takes part in opacity.
Because schwa epenthesis is also a phonological rule, a reviewer wonders if we can abandon schwa epenthesis as a OR-SR transformation and just use it as a UR-SR transformation.Such a stance however is psychologically unreal.If an Armenian speaker is literate, then they must have a mental grammar that can read existing and nonce words (do a OR-SR transformation), otherwise they wouldn't be literate.Furthermore, having both an OR-SR and UR-SR transformation does create a duplication of generalizations for the two processes, but such duplication is evidence that such generalizations are extremely robust language-internally.As for pre-literate speakers, the fact that the URs lack these epenthetic schwas likely helps these speakers learn how to read the likewise schwa-less ORs.
In sum, the Armenian schwa is a case of nearly isomorphic interaction between orthography, phonology, and morphology.explores an alternative analysis whereby vowel reduction does not involve schwa epenthesis at all: the reduced schwa is the output correspondent of the original high vowel. 35he data sets from §6.1 and §6.2 require a high-ranking constraint *ǐ,*ǔ against destressed high-vowels (see Dolatian 2021a for an exact formulation of this constraint).This constraint competes against reduction to zero (MAX) and reduction to schwa (ID[F]).Note the correspondence indexes disambiguate epenthetic schwas from reduced schwas (36).The main setback for the Deletion+Epenthesis approach is that it cannot explain why the high vowel is reduced to zero for words like (37a), but reduced to schwa for words like (37b), repeated from (31).Deletion would create syllabifiable clusters in both.In contrast, the Direct Reduction approach suggests that this difference is because deletion can only happen for word-medial open syllables but not for closed syllables.52).Such syllables are found when a) the schwa is epenthetic and there's a preceding morpheme boundary, b) the schwa is reduced and its deletion would create an unsyllabifiable cluster, and c) the schwa is underlying and unpredictably located because of diachronic syncope.Other logically possible contexts for open syllables are blocked, both for epenthesis and for reduction.We can summarize the above contexts in (38).There is a low-ranked constraint against word-medial open [Cə] syllables.This constraint is low-ranked because such syllables are attested.There is however a high-ranked constraint conjunction with a generic faithfulness constraint F. This constraint conjunction is violated whenever such a syllable is created via faithfulness violations (epenthesis and reduction), but not by underlying structures (underlying schwas).This constraint essentially distinguishes underlying schwas (which don't violate this constraint) from phonologically derived schwas (which violate this constraint) as a type of Derived Environment Effect (Kiparsky 1973(Kiparsky , 1993;;Łubowicz 2002;Hall 2006;Burzio 2011) bɑ.ɾə 1 s.pél * For derivatives of words like [bɑɾísp], the Direct Reduction requires that reduction to schwa is the default behavior, while deletion is a last resort.37This is in contrast to the Deletion+Epenthesis approach, where deletion is treated as the default behavior, while using a schwa is a last resort.
Finally, for derivatives of words like [ɑznív] (Table 52:b), the high vowel is reduced to schwa.A medial [Cə] syllable is created because deletion would create an unsyllabifiable cluster.To predict why we prefer reduction to a [Cə] syllable in the same location, rather than deletion and epenthesis of a [CəC] syllable, we need DEP to outrank our constraint conjunction.This means that we prefer creating reduced schwas in [Cə] syllables rather than epenthesizing [CəC] syllables.To summarize, we need a separate constraint system to handle vowel reduction within the Direct Reduction approach.Such a ranking can be united with our main constraint system for schwa epenthesis.The union of these rankings is in  The constraint system is quite complicated.The supplementary materials contain the various OT tableaux that we used to determine the ranking of the union of the schwa epenthesis and direct reduction.The tableau are complete with violation counts for all the constraints.The data was run against OT-Help (Staubs et al. 2010) to validate the correctness of our analysis.

( 10 )
Epiphenomenal generalizations from directional syllabification a. Schwa epenthesis must not create word-medial open Cə syllables.b.Schwa epenthesis minimizes the number of syllables.
To build up this generalization, consider what contexts allow or disallow medial [Cə] syllables (Table

Figure 2 :
Figure 2: Hasse diagram for schwa epenthesis constraints from sections §3-4 and for vowel reduction from Appendix B

Table 3 :
Categories of schwas in orthography vs. pronunciation

Table 4 :
Syllable structure of Armenian 3

Table 5 :
Some falling-sonority clusters that are not syllabifiable

Table 6 :
Complex codas licensed by both full vowels and schwas

Table 21 :
Medial schwa epenthesis in SCV clusters where C is not a stop . This contiguity constraint requires that root-initial orthographic STV clusters stay contiguous in the SR.It outranks ONSET.I demonstrate with the word <sdor> [əs.toɾ] 'below' and <sxal> [sə.χɑl] 'wrong'.I set aside the devoicing.

Table 23 :
No schwa prothesis for other <ʃC> clusters or for <ʒC> More refined versions of this contiguity constraint are needed to capture variable prothesis for ʃ C clusters.I put this issue aside.

Table 29 :
Schwa epenthesis in other consonantal prefixes

Table 33 :
Near-minimal pairs from epenthesis in compounds

Table 34 :
Variation in schwa prothesis for sibilant-stop clusters in compounds

Table 35 :
Canonical CVC verb root reduplication from existing bases

Table 39 :
Pseudo-reduplication of verbs without a base

Table 47 :
/V(C).Cv.C-V/ contexts where the reduced schwa is in a medial open-syllable Deletion+Epenthesis predicts that the surface schwa should be before the original vowel's location so that we get a closed syllable.But this is an incorrect prediction (29), and the schwa is in the same location as the high vowel: [ɑz.nív] derives ɑz.nə.vu.tʰjun instead of *ɑ.zən.vu.tʰjun.

Table 48 :
Deletion is blocked if the schwa is in a closed syllable /V.CvČ.C-V/ ).But without this constraint, then we incorrectly predict [VCC.CV...] without any schwa.