2.1 Background on Yorùbá

2.2 Background on language games

In this section, we discuss the typology of language games and particularly what is known about games similar to the one we will describe. Bagemihl (1995) notes four categories for how language games are created typologically. These are shown with examples in (5) (chart taken from Ozburn & Schellenberg 2019: 511).

The categories are not mutually exclusive. As an example, the English language game Pig Latin involves affixation of a suffix [-ej] and reversal of the word-initial consonant(s) (Barlow 2001), making it a combination of types (a) and (c).

All the categories identified by Bagemihl (1995) are attested in Yorùbá (Iṣọla 1982).⁴ Of these types, the one of interest here is infixation/affixation games (5a), in which nonsense material is infixed or affixed within the word, often multiple times. Yorùbá Ẹnà is quite similar to a language game described by Bagemihl (1987; 1988) in Tigrinya, in that –gV is inserted after every syllable, with V being a copy of the preceding vowel. For example, bɨ͡tʃa ‘yellow’ in natural language Tigrinya becomes bɨgɨ͡tʃaga in the game. Krämer and Vogt (2018) provide a useful background and typology of this type of language game with numerous examples.

Yu (2008) calls language games of this type iterative infixation ludlings (IIL) and also provides several examples from a wide variety of languages, including Lingua do Pê in Brazilian Portuguese, which uses a [-pV] infix, and hábàʔábà in Hausa, where the infix form is [-bV]. Iterative infixation ludlings are canonically monosyllabic across languages but may vary in size, segmental content, and inserted position. For example, the argot infix in Colombia Jerigonza is inserted after each syllable but the one in Peruvian Jerigonza is inserted before each syllable (Piñeros 1998), as shown in (6).

One of the factors that determines an inserted position is the size of adjacent syllables. For example, Colombia Jerigonza examples in (6) show that the ludling infixes are inserted after the syllable, regardless of whether the preceding syllable has a coda or not. However, this is not the case in Costa Rican Jerigonza (Piñeros 1998), Akan (Adomako 2018) and Tigrinya (Bagemihl 1988). As shown in (7), these ludling infixes separate the rest of the syllable from the coda when the preceding syllable has a coda. In this case, the insertion site or anchor point is each vowel, rather than the end of each syllable. As Yorùbá does not have codas, the options are essentially equivalent for this language.

Another property that can determine the position is stress. As shown in (8), the argot infix in Língua do Pê targets every syllable as an anchor point, except final unstressed syllables. Similarly, the argot infix in Faa di Vesu is inserted after a final syllable or an H tone syllable (in a privative tone system), as shown in (9).

Bagemihl (1988; 1995) and Vaux (2011), among many others, discuss how language games are frequently used as evidence for properties of natural languages, including insights about the structure of a specific language and about properties of human language more broadly. While phonologists have debated such uses of language games, Vaux (2011) notes that they are generally seen as valuable for providing insight into phonological systems and representations (e.g. syllable structure, phonotactics, tonal melodies, etc.).

Previous literature on language games, as well as on natural language infixation and reduplication, proposes a limited set of “anchor points” that insertion can target, namely “first syllable, first foot, first consonant, first vowel, stressed syllable, and final syllable” (Vaux 2011: 729; see Vaux 2011 for an overview of anchor points). In discussing affixation games, Vaux (2011) goes on to discuss the permutations of permissible precedence-altering instructions and anchor points as determining which outputs are possible in a game like Pig Latin. While not discussed in any detail, the same presumably applies to other iterative affixation games, in that the dummy affix should target valid anchor points and introduce new precedence relations in a way that corresponds with those anchors.

2.3 Background on VV sequences

Since the focus of this paper will be on the behaviour of VV sequences in Ẹnà, we conclude the background section by addressing the extensive literature on the behaviour of VV sequences cross-linguistically. Given that virtually all Yorùbá VV sequences contain identical vowels, we focus on the distinction between long vowels and heterosyllabic sequences of two identical short vowels.

As discussed by Kenstowicz (1971) on Lithuanian and Pyle (1970) on West Greenlandic, phonology has traditionally used two methods of representing identical VV sequences, a single segment marked [+long] versus a sequence of identical short vowels. These two representations make different predictions in how different phonological processes should affect long vowels, and both Lithuanian and West Greenlandic show evidence that both representations are necessary. For example, in West Greenlandic, lowering affects the entirety of long vowels, but not non-identical VV sequences, but long vowels can also block metathesis in the way that non-identical VV sequences do. As such, both representations seem necessary. While Pyle (1970) suggests that the representation switches at a point in the derivation, Kenstowicz (1971) instead suggests that different types of phonological processes (prosodic versus featural) use different types of representations. Pyle (1970) argues that both Lithuanian and West Greenlandic are consistent with both analyses.

More recently, it has been argued that in fact neither analysis works for other languages, such as Kamba (Roberts-Kohno 2000; Odden 2011). Instead, there has been a move towards an autosegmental analysis in which long vowels can be simultaneously a single vowel matrix and two timing positions (Odden 2011). It has also been noted that languages may contrast long vowels from heterosyllabic VV sequences, with languages such as Kamba, Gokana, and Matumbi showing phonological processes in which some VV sequences behave as long vowels, but others behave as a sequence of identical short vowels (Odden 2011). These languages show that the solutions proposed for Lithuanian and West Greenlandic (Pyle 1970; Kenstowicz 1971), where representation of VV sequences changes at a point during the derivation or is based on the process, cannot hold more generally. Instead, both types of representations must be possible within a language. We argue that Yorùbá is an example of a language where both representations occur.

It is also worth briefly examining some of the literature on the ambiguity of syllabification of identical VV sequences across a variety of languages. We will briefly discuss this question with respect to Fataluku (Heston 2014), some Kwa languages (Leben 2002), and Turkish (Kabak 2007).

Heston (2014) discusses a dialect of Fataluku that had previously been described as not having contrastive vowel length, but finds that it is in fact contrastive, just with very low functional load. Speakers in the study had some difficulty in certain cases saying whether vowels were long or short. Heston proposes that long vowels are sequences of vowels underlyingly, but that there is a rule that they can be realized in the same syllable when identical. This suggests that low functional load does not necessarily mean that there are no long vowels in a language.

Leben (2002) notes that in Kwa languages, if two vowels in hiatus are identical, they may sound like a single long vowel, and the syllabification is ambiguous according to speakers. For example, in chanting, such vowels can be mapped to one or two notes, which is not an option otherwise. Leben argues that these vowels therefore have ambiguous status, that they are phonologically two syllables but phonetically fused into one. He discusses some evidence from tonal patterns and language games about syllabification, including how language games can show VV sequences being treated as separate syllables even when ambiguous. This suggests the importance of language games for resolving ambiguity of syllabification.

Finally, Kabak (2007) discusses VV sequences in Turkish, which like in Yorùbá can come from loanwords and consonant deletion. Like Yorùbá, these VV sequences have been argued to be heterosyllabic and Turkish is said to not have contrastive vowel length. Kabak discusses vowel assimilation, which applies in certain contexts of VV sequences, and notes that it applies generally, but that speakers tend to maintain separate vowels in careful, formal speech. Kabak’s intuitions are that assimilated vowel sequences are heterosyllabic, particularly when there is a coda. However, he notes that syllabification judgements are difficult due to not knowing whether speakers might be using underlying or orthographic representations. He suggests that syllable-based language games might be used to test the status of VV sequences, and he discusses one in Turkish very similar to Yorùbá Ẹnà, in which –gV is added after every (C)V, where V is a copy. Informal data that Kabak collected on three assimilated words showed that the VV sequences were treated as separate syllables, confirming that they were heterosyllabic sequences. Kabak notes, however, that a more controlled study is necessary. The next section does so for Yorùbá and argues that not all are separate syllables.

3.1 Background on Ẹnà and methodology

Ẹnà is a general Yorùbá term for argot, ludling or secret language, encompassing both spoken language (Iṣọla 1982) and drum language⁵ (Villepastour 2010; Akinbo 2019; 2021). Our focus in this work is one version of Ẹnà in spoken language. While Yorùbá Ẹnà has been briefly described (Iṣọla 1982), it has not been documented in detail nor analyzed previously. Our data on Yorùbá Ẹnà were collected from one native speaker of Standard Yorùbá who was not involved in the analysis or writing of this paper. The consultant is educated to a Bachelor’s degree level, having studied linguistics and Yorùbá at the University of Lagos, and is also competent at reading and writing in Yorùbá, as he wrote his BA thesis in and on Yorùbá. He is also bilingual in English. The consultant was 25 years old at the time of data collection and learnt to use Ẹnà at the age of 7. According to the consultant, he learnt Ẹnà by observing his mother and her siblings using the ludling to mask their conversation. By the age of 12, the consultant was proficient and used the ludling with his own siblings, who learnt it from observing their mother as well. Since then, the consultant has been using Ẹnà to communicate with friends, classmates, and colleagues beyond his home. He is not proficient in the use of other language games besides the one on which data is reported here. Data was collected through written responses to wordlists, as the Yorùbá orthography effectively distinguishes relevant contrasts (e.g. tone, nasality). We met with the consultant via Zoom, and he was in Nigeria at the time.

3.2 General pattern

Our data show that, in general, in Ẹnà, gV is added after every syllable, where V is a copy of the vowel quality and tone of the preceding vowel. This is true for all vowels and tones, including nasal vowels, the language’s ‘default’ vowel [i], and the ‘default’ mid tone. Some examples are shown in (10). Notably, in (10), we observe that V and CV syllables are treated the same way, despite the fact that they are considered structurally different in many analyses of Yorùbá (see Orie 2000).

3.3 Syllabic nasals

As noted in Section 2.1.1, Yorùbá also has syllables that consist only of a syllabic nasal. When such syllables are copied in the game, again gV is added and the tone is a copy of the syllabic nasal’s tone, but there is no vowel to copy, and [g] cannot be an onset to a syllabic nasal according to Yorùbá phonotactics. In such cases, the consultant noted that the vowel in the Ẹnà syllable can be either [ũ] or [u]; the consultant’s description is consistent with either of these vowels being used in free variation with syllabic nasals in Ẹnà.⁶ Some examples are shown in (11). Note that the choice is the high back vowel, either oral or nasal, regardless of the specific nasal or on the surrounding vowels and consonants.

3.4 VV sequences

As noted in Section 2.1.3, VV sequences are quite common in Yorùbá: they can be underlying in loans, derived through consonant deletion, created with pronouns, or occur through compounding. Due to an assimilation process, adjacent vowels are always identical, except for the exceptional combination [ai] in certain lexical items. VV sequences, regardless of how they are derived, are traditionally analyzed in Yorùbá as belonging to separate syllables (Orie & Pulleyblank 2002).

In Ẹnà, however, there are two ways in which VV sequences behave. Some indeed behave as separate syllables, with VV --> VgVVgV, but others show the pattern VV --> VgV, with the [g] simply inserted between the two vowels. Pronouns, VV sequences derived from floating tones, and underlying VV sequences always behave in the latter way, and VV sequences in which the tones are identical can variably use either pattern, while other VV sequences use the former pattern. This is summarized in Table 3.

Table 3: Summary of VV patterns.

4.1 VV Sequences in Yorùbá

As discussed, previous literature argues that all VV syllables in Yorùbá are separate syllables, though arguments are primarily based on the behaviour of VV sequences derived through compounding and consonant deletion, as well as through a more theory-driven desire to avoid positing long vowels in the language. The data in Section 3 shows that VV sequences do not all act as separate syllables in the game. Instead, in Ẹnà, some VV sequences show separate syllable behaviour, VV --> VgVVgV, while others instead behave as VV --> VgV. The latter type of behaviour occurs with pronouns, underlying VV, and floating tone cases, as well as variably with VV sequences that share tone, while the former appears with compounding and consonant deletion (non-pronominal, non-underlying, non-floating-tone VV sequences).

There are two main possible directions that could be taken to account for these differences in behaviour: either previous analyses do not reflect differences in speakers’ representations of VV sequences, or the location of the dummy syllable in Ẹnà is not as simple as “after every syllable”. We argue here for the former. Indeed, the native speaker intuition of both the consultant and the second author is that there is no distinction in how types of VV sequences would be counted, but rather that they might be counting moras rather than syllables.⁷ For cases like pronominals, previous arguments in the literature for treating them as separate syllables are not based on strong arguments, as discussed previously, and they are inconsistent with the behaviour of the language game. We therefore argue that Yorùbá has long vowels, which behave as one in the game. Long vowels are clearly marked in the language, but as discussed in Section 2.3, low functional load does not mean that they do not exist. We suggest that the unmarked short vowel form emerges in the game context due to a reduplicative template and backcopying from the reduplicant onto the base. This process creates a case where it looks like the VV sequence is split by the game [g], when in fact there is copying and shortening. We will argue in Section 6 that this analysis is significantly simpler than any in which the language lacks long vowels.

4.2 Behaviour of same-tone sequences

While the basic Ẹnà pattern is that the Ẹnà syllable is inserted after every syllable, the situation is clearly more complex. It appears that there is no level of representation where this is true, particularly with respect to the non-pronominal VV sequences that share tone, which are analyzed in previous literature to be separate syllables with separate vowels both underlyingly and on the surface. According to our analysis of other VV sequences derived through compounding and consonant deletion, there is no reason to posit separate that the same-tone ones are long vowels, so it is odd that they can behave like the VV sequences that we do analyze as long vowels. Separate tones is also evidently not the appropriate analysis for where Ẹnà syllables are inserted, as pronominal forms have separate tones, yet behave as one for the purposes of Ẹnà.

To address this issue, we consider a crucial comment from our consultant, who says that the same-tone VV sequences treated like a single V behave as such because the VV becomes a single V in quick speech. The second author, who is a native speaker of Yorùbá, agrees that in colloquial, fast Yorùbá, same-tone VV sequences are typically produced with only V. The same is not true for distinct-tone VV sequences. A separate study of the nuances of Yorùbá colloquial speech would be necessary to fully understand this pattern, but for our purposes, we can assume that Ẹnà applies (or at least can apply) at a level of representation at which same-tone VV have been reduced to a single V. In this case, these examples behave in exactly the way that we would expect, with a single vowel. We suggest that the variability in the language game comes from the variability in natural language. Note that the other cases of VV that behave as one in the game, such as pronominals, cannot be reduced to V.

4.3 Distinct-tone VV sequences

What remains, then, is to understand the different types of distinct-tone VV sequences. As previously noted, previous analyses of Yorùbá posit that in derived (compound and consonant deletion) VV cases, one of the Vs deletes, followed by the other V joining to its mora. This structure was represented in (2), and it is identical to the structure that is assumed for underlying, pronominal, and floating tone VV sequences in previous literature, in which there is only a single underlying vowel but two moras for the vowel to attach to.

It is widely assumed that language games apply at intermediate levels of representation, not the UR or surface form (e.g. Vaux 2011). As noted, previous accounts assume that the surface forms of various VV sequences are the same, with one V associated to two moras, but the URs have a distinction between two separate vowels in compounds and consonant deletion cases, versus a single vowel in the cases of pronouns, underlying VV, and VV from floating tones. If Ẹnà happens at a level of representation where there are still two vowels in derived VV cases, before one vowel deletes and the other associates to a second mora, then we could claim that Ẹnà targets individual vowels. Other cases have only one V at all levels of representation. One potential issue with this analysis is that Ẹnà would then happen before vowel assimilation, as assimilation involves deletion. However, there remain two adjacent vowels that could undergo assimilation and force their base-argot correspondents to agree as well. For example, in ɔmɔ + ɛrã, if we added a gV to each syllable, we would get ɔgɔmɔgɔɛgɛrãgã, which still contains an ɔɛ sequence that could be forced to assimilate.

The larger issue for any proposal without long vowels is the paradox involving same-tone VV sequences, since we said previously that Ẹnà must apply at a level of representation after deletion of a mora occurs in same-tone VV sequences. Given that this mora deletion seems like a post-lexical phenomenon, as it occurs only in fast and colloquial speech, it would be strange for it to occur before the assimilation pattern that results in deletion of a vowel (but not a mora) in other VV sequences. Instead, we posit that Ẹnà actually applies on the surface, referencing colloquial surface forms. This requires a proposal for a new structure for some VV sequences. Specifically, we argue that some VV sequences are long vowels.

4.4 Proposed surface forms

We begin with the case of derived (compound and consonant deletion) VV sequences. Given that these vowels result from separate underlying vowels that assimilate, we posit that they remain separate syllables (but agreeing vowels) on the surface. Figure 1 illustrates our hypothesized normal (non-ludling) language surface forms for derived VV sequences. This analysis is the same as in previous literature.

Figure 1: Tone association in VV (bɛ̀ròòwò “start business” (<bɛ̀rɛ̀ “start” + òwò “business”).

For pronominal VV sequences, the traditional analysis of Yorùbá is that there is no underlying pronominal vowel, but rather the pronoun consists of a mora and tone. This mora then joins with the preceding vowel. As such, there is only one vowel, doubly-linked to two moras. We propose that these are in fact long vowels, within the same syllable. There is no clear evidence against this analysis for Yorùbá and it fits within the literature for other languages, both in terms of general discussion on the behaviour of VV sequences and in terms of what kinds of evidence a game can provide about representations. Specifically, the fact that the two types of VV sequences pattern differently in the game argues for different representations, and the fact that only one game consonant is inserted in these types of VV sequences suggests that they consist of a single syllable. Under this analysis, the game forms are derived through reduplicating the vowel as usual, but then a CV reduplicative template and backcopying onto the base results in short vowels in both the base and reduplicant. Figure 2 illustrates our proposed surface form of normal language pronominal VV sequences. Similar forms are posited for other structures that behave this way, such as floating tones.

Figure 2: Tone association in VV sequence (ràá “buy it” (rà “buy” + μ “3.sg obj)).

In the next section, we formalize this account, using a standard analysis of reduplication games based on that of Krämer & Vogt (2018), combined with a reduplicative template and back-copying. We show that these assumptions allow us to account effectively for the patterns in the game. We then argue in Section 6 that this game poses an issue for accounts that do not treat Yorùbá as having long vowels.

5.1 Basic pattern

We begin by illustrating our basic proposal about how Ẹnà works in the uncomplicated case, for (C)V syllables. As discussed, we assume that the syllable is the locus of insertion, although since the language does not have codas, this is equivalent to the vowel being the locus. We posit that Ẹnà involves the insertion of a single game morpheme, /-gV^red/, following Krämer & Vogt (2018). To fill out the dummy syllable, a vowel and tone need to be supplied, and reduplication occurs from the preceding syllable.

More formally, we adopt the constraints in (17) to analyze the basic pattern in Ẹnà.

Tableau (18) shows an example, with a VCV root meaning ‘plan’. Even though the game morpheme exists only once in the input, it appears after every syllable, since any other option is a fatal violation of MaxBR-V, following Krämer & Vogt (2018). The constraint IdentBR-T enforces tone copying. In (19), candidates (a) and (b) are ruled out by MaxBR-V, as they each only have one reduplicant instead of the iterativity required for the game. Candidates (c), (d), and (e) each violate both Contiguity and AlignL, but (d) and (e) are ruled out by violations of BR-Ident-T, as the tone associated with the reduplicant vowels differs from that on the base. Candidate (f) satisfies AlignL, but fatally violates AnchorL, as the edges of the source and game words do not correspond. The winning candidate is therefore (c).

5.2 Ẹnà VV sequences

For the VV sequences that we analyze as separate syllables, the exact same analysis as above will work, as these are simply two independent syllables and the constraints will derive the reduplicative dummy game affix being inserted after each one, in the same way as it did above with two syllables where the second had an onset. Since an identical tableau as above would work, with the word replaced by one with a separate-syllable VV sequence, we do not include the tableau for space reasons.

Turning now to the VV sequences that we analyze as long vowels in normal language Yorùbá, such as pronominals and floating tone, we posit that these forms show evidence for a reduplicative template and back-copying in Ẹnà. Specifically, while Yorùbá allows limited instances of long vowels, Ẹnà does not allow any long vowels; the long vowels are not copied into the reduplicative game morpheme, and they are further not kept in the base. Previous research on Yorùbá has posited a highly ranked constraint against long vowels. We diverge from previous literature in suggesting that this constraint is in fact violated in Yorùbá. However, we argue that a reduplicative template ensures that long vowels do not exist in reduplicants; all vowels in the game are short. Long vowels may not be copied as long into the game, and a requirement for length agreement between base and reduplicant forces the base to also have a short vowel in the game. We follow Gouskova (2007) and Kennedy (2009) on similar reduplication patterns that show long vowels becoming short in reduplication, both in the reduplicant and the base. Since all other vowels are short, this analysis does not affect anything in the other examples.

The long vowel examples are the only ones in which the tone differs on base and reduplicant. We suggest that this fact is due to the existence of two tones on the original base long vowels, which cannot be maintained on a single vowel once it is shortened. As such, the two tones spread out, one onto the base and the other onto the reduplicant, in order to keep both tones without creating contours. This is done at the expense of tonal identity between base and reduplicant. In other words, the tones in the Ẹnà form come from bans on deleting tones (MaxIO-T) or leaving tones floating (*Float) and on having two tones associated to a short vowel (*Contour), despite violating correspondence between the syllables in tone (IdentBR-T). These considerations are again not relevant for examples with short vowels. The additional necessary constraints are defined in (19).

The tableau in (20) illustrates. Candidates (a) and (c) are ruled out by having a long vowel in the reduplicant, which violates RED =σ_μ. Candidate (b), which keeps the long vowel in the base but shortens it in the reduplicant, is ruled out by MaxBR-µ, which requires the number of moras to be the same in base and reduplicant. In other words, this constraint enforces back-copying. The remaining candidates all violate MaxIO-µ, as they shorten the base vowel. Candidate (e) deletes one of the tones, violating MaxIO-T; candidate (f) leaves the second tone floating, violating *Float, and candidate (g) keeps both tones by creating contours, violating *Contour. The winning candidate is (d), which keeps both tones without creating contours, at the expense of violating IdentBR-T.

6.1 Syllabic nasals and tone changes

6.2 Existing analyses of similar games

As noted previously, there are reports of similar iterative affixation games in quite a few languages. Such reports have been primarily descriptive, but there are some existing analyses, both in rule-based frameworks and in OT. In this section, we consider the major components to existing analyses, aside from the one that we adopted, and consider them in light of the evidence from Yorùbá.

Several analyses of iterative affixation games, including Bagemihl (1988) and Piñeros (1998), motivate the game consonant and/or the game vowel through epenthesis. Bagemihl (1988: 247) has a rule that looks like ∅ --> X X / X __ to derive iterative affixation in a Tigrinya game. Similarly, Piñeros (1998) uses a markedness hierarchy of consonants to derive the choice of an epenthetic consonant in Spanish games. In contrast, in our analysis, the [g] from the game is part of the input for creating game forms.

Iterative affixation games do not typically insert the normal epenthetic consonant for the language (Bagemihl 1988; Alidou 1997; Yu 2008; Krämer & Vogt 2018). For example, the Yorùbá epenthetic consonant is normally [h] (Bamidele 2019), but the game inserts [g]. In fact, while coronals and glottals are the typical epenthetic consonants cross-linguistically, most of the iterative affixation games we have encountered in the literature use labials or velars as the affix onset. These facts suggest that the consonant belongs to the game, rather than being epenthesized, which makes it sensible to put it in the underlying form for the game. Since many other types of games are currently analyzed as having a game morpheme, either an overt one or one that serves as a trigger for a reversal or both (e.g. Borowsky & Avery 2009; Ozburn & Schellenberg 2019), it is also more congruent with other language games to have a game morpheme that accounts for the consonant, rather than epenthesis.

The choice of having a game consonant as part of the input, which then aligns with every syllable, matches the assumption of Krämer & Vogt (2018), the analysis that serves as the foundation of ours. The games discussed by Krämer & Vogt (2018) are in non-tonal languages, so they do not need to deal with the behaviour of tone. In cases of tone in non-game reduplication, it is sometimes copied (Akinlabi & Urua 2003; Pulleyblank 2008), but those are typically situations where entire syllables are copied, so there is no need to specify which parts of a syllable are reduplicated. In contrast, in games, it is typically only the vowel copied, with a constant consonant, and so for Krämer & Vogt’s (2018) analysis, they use V^red to indicate that only the vowel is reduplicated. In Yorùbá, then, we would appear to need separate reduplicative vowels and reduplicative tones, both V^red and T^red. Instead, we suggest that the tonal patterns are due to a base-reduplicant tone identity requirement. The violability of this constraint allows us to explain why, in certain VV sequences, the base and reduplicant do not have the same tone, by ranking this constraint below those that ban tone deletion, floating tones, and contour tones.

Using a reduplicative template and backcopying allows us to account for Yorùbá VV sequences in a way that would be quite difficult otherwise in Krämer & Vogt’s (2018) analysis. Specifically, in certain types of VV sequences in Yorùbá, as discussed above, the game [g] is inserted between the vowels. This means that, on the surface, it appears that neither vowel (or tone) is a reduplicated copy in these sequences. It therefore appears difficult to explain why it is allowed for that reduplicative vowel slot to be filled by a non-reduplicated vowel that is in fact part of the original Yorùbá word. We argued instead that there is reduplication, just combined with deletion of a mora due to a ban on long vowels in the game.

Accounts of iterative affixation games in both Piñeros (1998) and Yu (2008) make use of feet. Even if Yorùbá does use feet in its phonology, it would be difficult to motivate the different behaviours of VV sequences using that type of account: there is no direct match in Yorùbá between normal language syllables and the insertion of game syllables. We need to be able to explain why certain VV sequences are the locus of two game affix insertions while others are only one. Using feet does not help address this issue, so it has no apparent benefit over our analysis.

Yorùbá does not have codas or clusters, which means that it would be possible to motivate the presence of the vowels after every [g] through syllable structure constraints. However, our account does not require this, which is a benefit for extending the analysis to other languages with similar games that do allow for these more complex syllable types. While Yorùbá lacks codas, similar games in languages with codas often have the coda after the dummy CV from the game, as discussed in Section 2 in reference to Akan (Adomako 2018) and Tigrinya (Bagemihl 1988). For Yorùbá, after every syllable is equivalent to after every vowel, since the language does not have codas. This also means that we do not need to take a strong position on whether V on its own can be a syllable in the language (see Section 2.1.4).

6.3 Existing analyses of Yorùbá VV sequences

As noted, this paper argues against previous literature on Yorùbá by claiming that some VV sequences in the language are long vowels. We have already argued that the reasoning in previous literature for avoiding positing long vowels is primarily either theory-internal or based on compounding and consonant deletion, which do indeed seem to be separate syllables. There is no strong theory-external evidence related to the behaviour of other types of VV sequences, such as pronouns and floating tones. We argue that our language game data constitutes this evidence, because such VV sequences are treated as a single locus of dummy affix insertion, unlike other VV sequences. This suggests that they are single syllables.

In this section, we argue that an account of this game without positing that certain VV sequences are long vowels in Yorùbá would be difficult and stipulative. Specifically, any such account would need to determine a consistent location for dummy affix insertion, such that VV sequences from compounding and consonant deletion behave as two separate loci, while VV sequences from pronominals and floating tone behave as one locus. As noted, previous literature on the language assumes that all of these VV sequences are identically structured on the surface.

We argued in Section 4 that the Ẹnà game applies on the surface, because doing so provides the most straightforward account of the behaviour of same-tone VV sequences, which are variable in whether they behave as a single V or two in both regular language Yorùbá and in the game. As argued in Section 4, having the game apply at an early stage makes it difficult or impossible to account for why same-tone compounds behave differently from distinct-tone compounds.

The fact that the game seems to apply on the surface means that we cannot explain the different patternings of VV sequences in terms of their underlying structures. Previous literature on the language assumes that all VV sequences have identical surface structures, but they do not behave the same in the game. As such, we need to posit some sort of representational difference between these forms.

If we were to persist in claiming that there are no long vowels in Yorùbá, then the syllable could not be the locus of insertion of the game morpheme. We could potentially claim that the Ẹnà syllable is inserted after each unique vowel, but doing so would require some strange assumptions. Specifically, it would require compound and consonant deletion examples to have two unique surface vowels, instead of assimilation happening through deletion and re-association. Doing so violates many formulations of the OCP for vowels. It also remains difficult to formalize where the game syllable is inserted, in that the [g] would be occurring between two syllables, without being repeated for both. Further, we still would be unable to explain without extra work why we fail to get a reduplicated vowel, but instead split up the original two vowels. In other words, even if we could justify the presence of only a single game syllable, we would anticipate a game structure like VgVV, not VgV. This type of analysis, in order to insert the [g] between existing vowels, would also require internal inconsistency in where C/V tier segregation is used, in order to make the [g] appear between the vowels.

In contrast, our long vowel analysis is much simpler. Under this analysis, it is straightforward to describe where the game morpheme occurs: it is after every syllable. It is also straightforward to explain why the [g] appears to split up the vowels, in that it is an effect of a reduplicative template, back-copying, and requirements to maintain base tones in the game. All of these effects are well-established in other kinds of phonological processes in other languages.

Overall, long vowels clearly provide a more satisfactory account of the Ẹnà pattern. Without them, it would be complex and would require many unusual assumptions to describe where the game morpheme occurs and why reduplication fails to happen for the second V in some VV sequences. As such, we take Ẹnà as providing evidence that Yorùbá has long vowels. As previously noted, the evidence against long vowels in the language is primarily about compounds and consonant deletion, or theory-internal (not wanting to posit long vowels), or could be re-analyzed in terms of speakers counting moras rather than syllables. Thus, long vowels remain consistent with other facts about the language, but provide a more satisfactory account of the game.

6.4 Implications

The Ẹnà game discussed in this paper has multiple interesting implications to our understanding not only of Yorùbá, but also of language games and syllable structures more generally. This section considers these implications in more detail.

As noted in Section 2, there is an assumption in previous discussions of Yorùbá VV sequences that they are all separate syllables, regardless of provenance. However, there is also an assumption that their underlying structures differ, such as pronouns containing a mora without an underlying vowel. In this paper, we have seen that VV sequences pattern differently in a way that corresponds to the posited underlying structures. However, the game also operates on a level of representation where colloquial speech vowel deletion in same-tone VV sequences has applied. Language games are thought to operate on intermediate structures, but we end up with a paradox if we try to apply that to Ẹnà. Instead, we argue that VV sequences in Yorùbá that are derived from consonant deletion or compounding are indeed separate syllables, but other VV sequences are long vowels.

In Yorùbá, as noted in Section 2, there are multiple possible sources for VV sequences: underlying (e.g. in loanwords), pronouns, floating tone, consonant deletion, and compounding. In keeping with previous literature, we assume these sources have distinct underlying structures: respectively, underlying double-associated V, a V plus an unassociated mora, a V plus an unassociated tone, a VCV sequence with each V attached separately, and two separate V in separate words. Thus, all types have two underlying moras, but while the first three types have a single V underlyingly, the latter two have two underlying Vs. This difference corresponds to the difference in their patterning in Ẹnà: a single vowel underlyingly corresponds to a single Ẹnà insertion site, while two underlying vowels generally corresponds to the insertion of two Ẹnà dummy syllables, except when the two vowels share tone.

Previous work on Yorùbá suggests that, regardless of the provenance, all VV sequences end up as a single V associated with two moras in separate syllables. Moreover, the language game literature argues that language games operate on intermediate representations, rather than underlying or surface representations (e.g. Bagemihl 1988; Vaux 2011; Kitaoka & Mackenzie 2021). In Yorùbá, it is unclear why there would be an intermediate level of representation at which one of the moras in same-tone VV has deleted, but the deletion inherent in previous analyses of the assimilation pattern for other VV sequences has not yet occurred. As such, the combination of these assumptions causes a paradox for Ẹnà, where certain VV sequences behave as having a single syllable, while others behave as though they have two. In this paper, we have suggested that some VV sequences in Yorùbá are long vowels if they have only a single V underlyingly. In other words, if there are two underlying vowels, then they may assimilate on the surface, but they remain separate syllables, whereas other cases have underlying long vowels or morphological lengthening. The importance of game evidence for distinguishing long vowels from heterosyllabic VV sequences has been previously discussed by Leben (2002) and Kabak (2007), both of whom found that ambiguous VV sequences were separate syllables. Our paper is, to our knowledge, the first to show a case where games illuminate two distinct structures of apparently surface identical VV sequences, further emphasizing the importance of games to disambiguate. While traditionally phonology did not allow languages to have two different types of representations, modern work has shown that several languages do need to allow both long vowels and heterosyllabic identical VV sequences (see Odden 2011 for an overview), and Yorùbá appears to be such a case. As previously noted in other work on other languages, such as by Heston (2014), the low functional load of long vowels does not necessarily imply that they do not exist.

Overall, in terms of Yorùbá, this language game analysis has suggested new Yorùbá structural representations of VV. The analysis confirms the posited underlying distinctions, in that these distinctions are reflected in different patterning in the game. We have argued for different surface representations than previous literature on Yorùbá, particularly in reference to long vowels, which are necessary for a satisfactory account of the game patterns, despite claims in the literature that the language does not have long vowels. As has been previously argued by Leben (2002) and Kabak (2007), among others, games can illuminate intuitions of VV sequence syllabification in cases that are otherwise ambiguous, and Yorùbá is a clear example of this.

Finally, we note that the study of Ẹnà illuminates the importance of examining these sorts of language games in tone languages, in order to fully understand the requirements of a satisfactory analysis. While games like Ẹnà are known in the literature, there are few theoretical analyses, and many of the existing ones have difficulty dealing with the addition of tone to the equation and with VV in this game. The analysis of Ẹnà showed, for instance, that the simplest analysis is not reduplication of tone, but rather a base-reduplicant tone identity requirement. Examining language games in a wider variety of languages can only help in understanding the patterning of tone in language games, the behaviour of reduplication and games more generally, and what such patterns tell us about human language.