2.2.1 The tripartite system of number marking

As is characteristic of Nilotic languages, Sengwer and other Kalenjin languages have a system of tripartite of number marking (Dimmendaal 2000; Di Garbo 2014; Kouneli 2021). This means that nouns are classified into three patterns of morphological marking for number: (a) inherently singular nouns which are unmarked in the singular and marked in the plural; (b) inherently plural nouns which are marked in the singular and unmarked in the plural; and (c) numberless nouns which are marked in both the singular and the plural. Table 2 shows examples of nouns from all three number-marking patterns.

Morphemes marking the singular number in Kalenjin languages, such as -jà (b) and its allomorph -cà (c) have often been called singulatives in the literature (first used by Rottland 1981a; 1981b). Following Kouneli’s (2021) interpretation of the phenomenon, however, these are best thought of as singular suffixes instead. This is because, while singulars are simply allomorphs of singular number, singulatives are classifier-like individuating suffixes which modify collective nouns only (Greenberg 1978; Grimm 2012; 2018). Since there is no evidence that inherently plural nouns in Kalenjin are collectives, it is best to call these singular rather than singulative suffixes (cf. Kouneli 2021).

Within the dataset (Falletti 2023a), nouns are not evenly distributed across these three patterns: (a) 53.3% (653 items) of nouns are inherently singular, (b) 24.5% (300 items) are inherently plural and (c) 22.2% (272 items) are numberless. These morphological classes can also be described in terms of the semantic categories of nouns they contain (Dimmendaal 2000; Kouneli 2021):

1. The inherently singular class mostly contains count nouns;

2. The inherently plural class mostly contains mass nouns (e.g., flour, water, seeds) and other referents which are usually found in groups (e.g., animals, plants and people).

3. The numberless class contains referents found in groups (e.g. trees, plants, people, pairs of objects) but does not contain any mass nouns. A larger proportion of nouns in this class are derived compared to the other two classes.

As in other Nilotic languages (Dimmendaal 2000; Moodie 2016; 2019), the semantic characterisation of the three patterns of number marking represents a reliable generalisation rather than a consistent and regular pattern. For instance, against what one would expect from its semantics, a mass-like noun such as mèn ‘clay’ is found in the inherently singular class rather than in the inherently plural class. Moreover, there is often speaker variation regarding which number category a noun belongs to, in particular when the word is not common.⁴ For instance, the word for ‘cricket’ is inherently singular for some speakers (19) and numberless for others (20).

2.2.2 Inflection for definiteness

Sengwer nouns take two suffixes for definiteness: one for the singular and one for the plural. In turn, each of these has two allomorphs. The singular definite suffix has the allomorphs -ɪ́t and -tâ, which are almost equally common. Since their vowel is underlyingly -ATR, it harmonizes with +ATR vowel(s) in the preceding root—as in (22b).

The plural definite has the allomorphs -ɪ́k (23) and -kâ (24). However, the latter has a very limited distribution, with only four nouns in the whole dataset, all of which have irregular stems.

Therefore, expanding on Table 2 and taking into account the allomorphy of the definiteness marker, Table 3 shows that every noun can have up to four basic forms⁵: singular indefinite, singular definite, plural indefinite and plural definite.⁶

As shown in this table, definite suffixes often interact with the preceding stem, which may consist of a root or a root plus a number-marking suffix, depending on their number class. As a result the underlying forms -ít and -ík are rarely realised as such. Instead, these suffixes merge with preceding vowels in sandhi processes, surfacing with different vowel quality and length– see for instance the addition of the definite suffix in kɔ́rɔ̀ɔr-jɛ́ɛt (b) and kɛ̀pɛ́p-cɛ́ɛt (c). The rules governing these sandhi processes are discussed in more depth in the next section.

In Endo and Nandi, just as in Sengwer, these suffixes express specificity and definiteness (Zwarts 2001; Hollis 1909). In other Kalenjin languages, such as Kipsigis (Kouneli 2020) and Kony Ogiek, this system has partially collapsed, with the definite form becoming the only form for many nouns. Having no clear semantic role, the label “secondary” has often been used to refer to such lexicalised suffixes across this language family (Toweett 1979; Creider & Creider 1989).

3.3.1 Evaluating the thematic suffix analysis

In her 2021 paper on number-based noun classification, Kouneli proposes that the unexpected vocalic segments which appear in the noun paradigms of Kipsikis are akin to thematic vowels in languages such as Latin, Spanish and Ancient Greek. While admittedly tentative and left as a topic for further research, this analysis presents interesting points to our discussion. In this view, all ghost segments are declension class markers: morphologically active suffixes which determine the declension paradigm of a noun (Aronoff 1994). In Indo-European languages, the presence of these thematic suffixes depends on the noun stem: while some stems require them, others do not. The quality of a thematic suffix is also lexically specified, falling into a limited number of categories (i.e., declension classes).

While at first glance this is a compelling analysis for Sengwer as well, when applying our diagnostic tests, treating these segments as thematic—i.e., belonging to the stem—becomes problematic. Though it is true that the segments in question often appear either after roots or after suffixes—as they do in Latin for instance—they can also appear simultaneously after both. As shown previously, the word pèet (62a) has a root-final vowel /u/ that only surfaces in its inflected forms. One such form is the plural pèetúus (62b). Here, sandhi between the underlying stem-final /u/ and the suffix-initial /i/ of the plural marker -íis yields the allomorph -úus in the surface form.

However, the nominative plural pèetúusjʌ̀ (62c) and plural definite pèetùusjêk (62d) of the same noun show that the suffix -íis itself has a ghost sequence /jʌ/. This means that both /u/ and /jʌ/ are simultaneously present ghost material in the inflection of the noun pèet. According to the model used by Kouneli (Oltra-Massuet & Arregi 2005), declension class suffixes should only occur once for every stem—whether that is after the root or after the suffix—but not after both. In other words, while this analysis predicts only one thematic suffix per stem, the data commonly includes nouns with two thematic suffixes per stem: one after the root and one after a suffix.

Moreover, since the ghost sequence /jʌ/ is always present after the suffix -íis, even in nouns such as côok (63) which do not have any root-final latent element in their inflected forms—i.e., are not thematic—we cannot attribute its appearance to a stem-dependent process. Therefore, the fact that Sengwer ghost segments appear after functional bound morphemes independently of whether the stem has a latent segment itself argues against the interpretation that they are thematic suffixes.

Even if the ghost sequence /jʌ/ were a declension class marker introduced by the suffix -íis, in order to be considered the marker of a declension class, it would have to appear in more than just a single environment. However, in plural nouns, the ghost sequence /jʌ/ can only be found after -íis.

Another argument that challenges the analysis of these segments as thematic suffixes is the fact that it does not provide a descriptive distinction between root-final segments which appear in the unmarked surface form and those which do not. Nouns with a root-final /u/ which surfaces in the unmarked form such as léŋkû (64) take the same singular definite allomorph and the same singular proximal demonstrative allomorph as nouns with a ghost root-final /u/ such as pèet(u) (65). This contrasts with nouns with no root-final /u/ such as côok (66).

Both /u/ vowels in (64) and (65) are root-final elements which do not contribute any extra meaning to the noun while influencing the choice of inflection: the only difference between them is whether or not they surface in the unmarked form. Therefore, if we were to label the /u/ of pèet(u) as a thematic suffix, we should use the same label for the /u/ of lénkû—despite their obvious differences in behaviour. Labelling root-final segments as thematic would not only make every root-final short vowel a thematic suffix but also every root-final /p/, /s/, /c/ and /j/—all of which are possible consonantal ghost segments. This would mean that as well as a declension class for each of the ten potential stem-final short vowels, more declension classes would have to be posited for several stem-final consonants. For instance, in this view, éemʌ̂(s) (67) and kɔ̀mɔ̀s (68) would both have thematic suffix -s and belong to the same declension class.

However, as shown in more depth in Section 3.3.3, positing declension classes in Sengwer is not necessary when analysing these segments as cases of deletion rather than insertion, as the whole nominal inflectional system becomes predictable. Declension classes such as those in Latin and Ancient Greek are nominal inflectional patterns which are not predictable by the shape of the noun and where synthetic processes make it impossible to accurately separate the root from its suffixes. While the current stage of Sengwer does have an abundance of sandhi processes which can make the boundary opaque, the root and its suffixes are still distinct.

For these reasons, we believe that there is no evidence that ghost segments are suffixes, as these latent vocalic and consonantal segments occur in all inflected forms of their noun without contributing any semantic content. Calling them “suffixes” in light of this would counter all definitions of the term (Crystal 1980; Hartmann & Stork 1972).

To summarise, the evidence presented so far suggests that it is not possible to analyse the phenomenon of latent segments in Sengwer as an instance of thematic suffixes which mark declension class based on their behaviour and distribution. First, we have shown that their presence is not always dependent on the stem, as these segments can be found (a) after suffixes, (b) after roots and (c) after both at the same time. Therefore, they cannot be called thematic. Second, the behaviour of these latent root-final segments patterns with that of counterpart segments that surface consistently in their inflected forms. Therefore, analysing them as suffixes would mean that all root-final elements should also be analysed as suffixes, which is not tenable given that neither segments carry any meaning. Third, these segments do not all form natural declension classes with segments in other nouns, since certain elements only occur after a single suffix or a single noun root.

3.3.2 Lexical and affixal ghost segments

While ghost material such as that seen in pèet(u) (65) and éemʌ̂(s) (67) discredit a suffix analysis, there are other cases in which the label suffix is warranted by the distribution of certain segments and their contribution to word meaning. A large subset of nouns in the data shows alternations in the presence of ghost segments between their singular and plural forms. The noun múrèn ‘man’, for instance, has a ghost vowel /ʌ/ in the singular (69a) but lacks a ghost segment in the plural (70a). This is evidenced by the fact that the ghost segment triggers regular sandhi in the singular definite inflection (69b) but does not in the plural definite inflection in (70b). If this ghost vowel were part of the stem, the plural definite inflection should be the ungrammatical form in (70c) instead.

The absence of a stem-final vowel in the plural forms is further confirmed in the nominative inflection; compare the nominative singular mùrènʌ́ (71b) and the nominative plural mùrén (72b). Once again, while we find a stem-final /ʌ/ in the singular, there is no such segment in the plural.

The inverse situation is also possible; stem-final ghost vowels can appear in the plural inflection of some nouns but not in the singular. For instance, the singular noun kɛ̂ɛt ‘tree’ (73a) only shows a change in tone in the nominative kɛ̀ɛt (73b), receiving no additional segment. In contrast, its plural form, while appearing with no stem-final vocalic segment in the unmarked form kɛ̀ɛt (74a), “unexpectedly” receives a high-toned /ɪ/ in the nominative plural kɛ̀ɛtɪ́ (75b).

This leads to the conclusion that these segments do not depend on the root but on the number specification of the noun. Therefore, while segments that surface in all marked forms of the noun must be part of the root and segments that surface in all marked forms after suffixes must be part of those suffixes, segments which only surface in the singular or plural forms of the noun must be, in fact, number suffixes. The difference between these three categories can be observed in their distribution among the nominative forms in the singular and plural inflection illustrated in examples (75)–(77).

The ghost vowel /u/ is part of the root of the noun pèet (75), as it appears in all its plural and singular inflections and does not influence word meaning. The ghost sequence /jʌ/, on the other hand, is part of a suffix as it invariably appears after the plural suffix -ɪ́ɪs in the inflection of nouns pèetúus(jʌ) (75) and cóok-íis(jʌ) (76). Finally, the ghost vowel in múrèn(-ʌ) (77) is a whole suffix, as it appears in all singular inflections but never occurs in the plural inflection and its presence marks the difference between the singular and the plural.

Following this analysis, 159 nouns which appear to be inherently singular and 23 nouns which have segmentally identical singular and plural forms can now be reanalysed as having one of the three ghost singular suffixes (-ja, -a and -tɔ) instead. Take the noun in múrèn(-ʌ) in (77) again, for instance; on the surface, it appears to only differ in tone from singular to plural, with no clear indication of which of the two forms is unmarked. However, after applying the diagnostic tests, the singular múrèn(-ʌ) is shown to have a singular ghost suffix. This makes the plural múrên the unmarked form and classifies this noun as inherently plural.

On the other hand, nouns such as rɛ́ɛrɛ̀ɛs(-ja) (78a) below appear to be unmarked in the singular in their surface form while, in fact, they have a ghost singular suffix -ja.

This suffix regularly appears in its singular inflected form, such as the nominative in (78b), but is completely absent from its plural form, where the plural suffix attaches directly to the root (78c). If this was not the case, and these syllables were root-final elements we would expect to find a glide /j/ surfacing in the plural forms. However, the form in (79) is ungrammatical:

Therefore, although appearing to be inherently singular, the noun root rɛ́ɛrɛ̀ɛs is, in fact, numberless in terms of its number class, as it takes suffixes both in the singular and in the plural.

Corroborating this analysis is the fact that some ghost singular suffixes such as -jà (and its +ATR counterpart -jʌ̀) are well-attested in the data as surface singular suffixes—as shown for sìkìr-jʌ̀ (80a).

Furthermore, the surface and ghost variants of -jà share a semantic domain distribution: both are particularly high in flora and fauna terms, while limited to a few items across other semantic domains.

By postulating ghost number suffixes, many of the suprasegmental differences in tone, length and ATR specification between surface singular and plural forms, which were previously labelled irregular, become predictable. For instance, the differences in tone and ATR between the singular and the plural of the noun ŋɛ́ljɛ̂p (81), can now be explained by the presence of a plural suffix ghost segment -i. Its presence predictably triggers three suprasegmental changes: it regularly lengthens the preceding syllable’s vowel, induces a (L.)H replacive tone pattern and changes the ATR specification from -ATR to +ATR. The existence of this suffix is evidenced by its presence in the nominative plural ŋèljèep-í (81d) and its absence in the nominative singular ŋɛ̀ljɛ́p (81b).

Evidence for this plural marker (-i) is found in 80 noun tokens in the data (see Table 4). This suffix is particularly productive in the derivation of deverbal agentive nouns, as shown in (82).

The verbal root al, which is an underlyingly toneless -ATR morpheme, become high-toned and +ATR with the addition of the ghost plural suffix -i. Since the vowel of the verb root al is short, it is lengthened in the agentive.

Another surface irregular change that can be explained by the presence of a ghost suffix is the alternation between palatal and velar consonants in the absence of a surface suffix.

In Sengwer, certain words—nouns, adjectives and verbs alike—have a root-final palatal which turns to velar when a suffix is added. For instance, the noun tjʌ̂ʌɲ (83) has a root-final /ɲ/ in its unmarked form but consistently surfaces with a root-final /ŋ/ in all its suffixed forms.

This same alternation can occasionally be seen even when no surface suffix is present. For example, the verbal root rwʌʌc has a root-final /c/ in its imperative (84a) but surfaces with a root-final /k/ in its deverbal derivation (84b).

However, applying diagnostic tests such as the addition of the singular definite suffix (85c), we can see that this alternation is, in fact, triggered by a singular suffix -ʌ.

Therefore, by postulating the presence of ghost suffixes, many seemingly irregular segmental and suprasegmental patterns found in nouns become predictable. These alternations are evidence of their presence in the underlying representation; even when absent from the surface representation, ghost segments still influence preceding surface material.

However, while alternations in ATR harmony and length can be easily predicted by the presence of specific ghost segments changes in tone are not always as straightforward. While there is evidence for regular changes of tone and length in relation to ghost segments and ghost suffixes in particular, a study of these patterns would require a thorough discussion of the tone processes in the language which falls outside the scope of this paper.

To summarise, then, although it seems that ghost segments are never thematic suffixes, for some of the nouns in our data, we did find evidence of ghost suffixes. These appear to be number suffixes rather than declension class suffixes. They are a small group of morphemes (listed in Table 4) which appear either only in the singular or only in the plural forms of nouns. The observations presented so far in this section mean that ghost segments in Sengwer can be classified into three categories: part of the root, part of a suffix and whole suffixes.

Though differing in behaviour, all these ghost segments have common features: (a) they are absent stem-finally in the base forms of the noun but present in all or most marked forms of the noun form they appear in, either as a result of suffixation or a suppletive change in tone pattern, (b) they are either a single coda consonant C#, a short-vowelled open syllable (C)V# or a combination of both in that order C.CV#. Therefore, rather than being suffixes added onto marked stems to determine their inflectional class, evidence suggests that these segments are stem-final material deleted for phonological reasons. In the next section, we present some arguments for the use of ghost segments rather than thematic suffixes to refer to this phenomenon.

3.3.3 Arguments for a ghost segment analysis

The fact that some of the Kalenjin ghost segments could be part of the morphemes they appear after, rather than being separate suffixes, is proposed twice in the literature: by Bennett (1974) and by Dimmendaal (2012). Both papers deal with features of the Nilotic family at large and, therefore, do not account for this phenomenon in detail. In his description of tone in relation to the Nilotic case system, Bennett (1974) compares the ghost segment phenomenon in Kalenjin to the so-called shadow vowels of Teso, a related East Nilotic language. In Teso, some short root-final vowels in an open syllable are elided in the unmarked form of the noun but occur in all suffixed forms as well as before any consonant-initial word. This is in line with Zimmermann’s (2019: 1) definition of the term ghost as:

“segments that (1) are idiosyncratically bound to specific morphemes and (2) alternate with zero in a way that the majority of segments within this language do not.”

Though Teso’s shadow vowels as reported by Bennett are similar to the Sengwer phenomenon at issue in this paper, there are some important differences. The surfacing of shadow vowels in a particular word in Teso is phonologically predictable by the make-up of the following word as well as the addition of a suffix. This means that the patterning of ghost segments in Teso is sensitive to contexts that cross word boundaries as well as those which are word-internal. For Sengwer ghost segments, on the other hand, the context across word boundaries is irrelevant. More importantly, Teso’s shadow vowels surface purely for phonological reasons; these ghost vowels appear to prevent consonant clusters, which are not allowed by Teso’s phonology. This means that, although the presence and quality of shadow vowels are lexically specified, they surface to satisfy a phonological constraint on the phonotactics of the language. Lindsey (2019), in her discussion of ghost phenomena, calls this kind of ghost phenomenon hero ghosts: the shadow vowels “come to the rescue” to avoid consonant clusters. Sengwer ghosts, on the other hand, appear to be purely lexically determined; ie. they do not appear to interact with any markedness constraint in the language.

Though the term ghost can often be found in the literature to refer to phenomena described by Zimmermann’s definition (Zoll 1993; Kiparsky 2003; Archangeli 1984; Szypra 1992), many other labels have been used depending on the element affected, its behaviour and the tradition of the field: floating feature (e.g., Remijsen & Ayoker 2020); latent segments (e.g., Tranel 1996a), phantom consonants (e.g., Schmidt 1994), liaison consonants (e.g., Adda-Decker et al. 1999) and epenthetic segments (e.g., Hyman 1972). This plethora of different terms has not always aided researchers in recognising the striking similarities between parallel linguistic features. In this section, we will show how the umbrella term of ghost segment allows us to draw interesting parallels between similar linguistic phenomena.

While authors such as Larsen (1991) have proposed that the Kalenjin ghost segments are also inserted for phonological reasons, there is no such evidence in our data. In his description of the nominal morphology of Sabaot, Larsen (1991: 7) states that: “the purpose of consonant insertion is in all cases to avert an unwanted vowel clash.” In his view, the ghost segment /nta/ at the end of the singular suffix -jaa would be a thematic suffix (glossed by the author as thm) inserted “to avoid vowel fusion”. Using Larsen’s analysis and notation, example (86) shows that the noun mʊr-jaa, for instance, would receive a thematic suffix -nta in order to avoid a hiatus between long /aa/ and /ɪ/.

This analysis has one obvious problem: while the hiatus between the long final vowel of the singular suffix and the initial vowel of the singular definite suffix is resolved, the addition of /nta/ gives rise to a new hiatus between short /a/ and /ɪ/. As is the case for all other instances of single adjacent vowels, this new hiatus is resolved by coalescence (i.e., sandhi) rather than the insertion of any “thematic” material. Therefore, we could restrict Larsen’s claim and state that thematic suffixes are added only to resolve hiatus when long vowels are involved. Yet, hiatus between two adjacent long vowels or a long and a short vowel is extremely common in our Sengwer data, occurring in at least 100 cases both root-internally and across morpheme boundaries. Examples (87) and (88) show these two possibilities (VV.V and VV.VV) with the same vowels that would result in hiatus in the ungrammatical form *mʊr-jaa-ɪt in (86).

However, inserting material that is not at all present is different from realising underlying “weak” material; that is, the language could be making use of underlying material when possible to avoid hiatus. This effect is called Emergence of the Unmarked (McCarthy & Prince 1994). Still, even taking this into the account, the addition of the ghost sequence /nta/ does not truly avoid hiatus in (85), as it introduces a final vowel /a/ which merges with /ɪ/ at the morpheme boundary, resulting in /ɛɛ/. This is true for the majority of our data: ghost segments in Sengwer normally contain at least one vocalic segment which would incur in hiatus with any following suffix but is instead resolved by coalescence (rather than consonant insertion). Therefore—at least for Sengwer—there is no reason to believe that ghost segments are inserted to avoid hiatus or any other phonological constraint. Instead, as argued in the previous section, it is more likely that the phenomenon of ghost segments is one of stem-final deletion. In his paper on metrical structure in the morphophonology of Nilotic, Dimmendaal (2012: 16) puts this interpretation forward in relation to the noun system of the Kalenjin language Nandi, stating that the phenomenon is one of “omitted (truncated or deleted) thematic vowels”.

Although deletion appears to be the best explanation for this phenomenon, the quality and distribution of ghost segments in our data do not point towards a single phonological triggering environment in the current stage of the Sengwer language. Instead, the fact that this phenomenon is common to all branches of the Kalenjin language group (Dimmendaal 2012) suggests that this is a process which has its roots in Proto-Kalenjin at the very latest. In fact, as we have seen in the example of Teso, similar phenomena occur all over Eastern Nilotic (Dimmendaal and Breedveld 1986, Dimmendaal 1983), one of the three main branches of Nilotic. Even within South Nilotic, Rottland & Creider (1996: 1–2) state that Datooga’s short vowels in final position are realised “very weakly and are generally voiceless”. Dimmendaal (2012: 17) comes to the same conclusion, stating that: “the Kalenjin system of “thematic vowel” truncation presents the end result of the kind of alternation still found synchronically in neighbouring Teso-Turkana languages”.

Moreover, it seems that different deletion triggers were active at different times in the evolutionary history of this language family. Comparative data between Kalenjin language varieties suggests that deletion from the unmarked form occurred both before and after these languages split apart. For instance, in the cognate nouns in (89), the Sabaot cognates have a ghost vowel whereas the Sengwer cognates do not.

Since these languages have a common ancestor, Sabaot must have deleted stem-final vowels such as ghost /e/ after it split from Sengwer. Conversely, cognates such as those in (90) show that Endo (Zwarts 2003), for instance, kept certain stem-final consonants, while Sengwer deleted them.

Therefore, the deletion of certain segments stem-finally is a diachronic phenomenon which must have occurred several times, being triggered by different phonological conditions and affecting different segments. The exact triggers for such deletion events cannot be fully understood without a comparative study of the lexicon of South Nilotic which goes beyond the most recent reconstruction by Rottland (1982; 1989). However, the data suggests that syllable structure and tone were involved, therefore some preliminary hypotheses can be made.

As explained in Section 3.2, ghost material is limited to four phonological shapes, all of which occur stem-finally: C#, .V#, .CV# and C.CV#. However, the disappearance of these four shapes from the unmarked stems can be reduced to two deletion events. First, ghost sequences of the (.CV#) kind only occur after a preceding consonant, meaning that the deletion of the consonant onset was to avoid a complex consonantal coda, a constraint found in all of South Nilotic. For instance, the deletion of (.V#) from (91a) to (91b), would have produced the unallowed coda /ŋk/. To avoid this, the syllable onset was likely deleted in tandem with the stem-final vowel (91c).

Therefore, the deletion of .CV# sequences must have been part of the same deletion process as single vowels (.V#). Second, as explored earlier in this chapter, the C.CV# ghost sequences are the result of two deletion events, first the deletion of .CV# and then the deletion of C#. Therefore, we can say that although four phonological shapes were deleted, there were only two deletion events: an earlier one which deleted .(C)V# sequences and one which deleted C# segments.

Single coda consonants are the least common kind of ghost material in the data. They are mostly (though not exclusively) found after long vowels with a falling tone and are only present in high-frequency nouns such tjêe(p) ‘girl’, cîi(c) ‘person’ and kwɛ̂ɛ(s) ‘buck’. As mentioned in Section 3.1, some coda consonant ghost segments are found in the surface form when changes in tone are applied to mark the nominative case. This fact supports the hypothesis that the tone of the preceding vowels triggered the deletion of ghost material. However, this explanation does not work for all ghost coda consonants; only the most common of these, /s/, is found to surface in the nominative. The other ghost consonants only surface when suffixes are added (e.g., the definite suffixes or demonstrative suffixes). This suggests that either the historical deletion of /s/ had a more complex route that lead to the pattern seen today⁸ or that this is an example of paradigmatic analogy, where speakers have generalised a tone rule that applies to vocalic ghost material to a more general rule that applies to the most common kind of consonant ghost material as well. Since nearly half (49%) of all the nouns in the data end in a consonant and only 8 high-frequency nouns and 1 high-frequency suffix contain a ghost consonant (see Table 4), we can assume that the deletion of final consonants was not a regular process.

On the other hand, all ghost material containing a vowel can be found in high-frequency as well as low-frequency nouns, after a variety of phonological contexts and after all three tones. However, although the preceding tone environment is not predictable itself, the tone sandhi interactions between the ghost segment and the following suffixes are partially predictable by the preceding context. If the preceding context is a low tone, the addition of a high vowel suffix to the ghost segment can yield either a level high tone or a falling tone on the sandhi syllable. Moreover, if the preceding context is a high tone or a falling tone, the addition of a high vowel suffix to the ghost segment always yields a falling tone on the sandhi syllable. In example (100), the singular definite suffix -ɪ́t is added to the low-toned monosyllabic noun sòt(o); the ghost vowel and the suffix undergo sandhi and yield a high-toned vowel. However, when the plural definite suffix -ɪ́k is added to another low-toned monosyllabic noun tʌ̀ʌk(-i) in (101), the sandhi between the ghost vowel and the suffix yields a falling tone instead. On the other hand, both the high-toned noun kwéen(u) in (102) and the falling-toned noun mbâr(a) in (103) when suffixed with the singular definite suffix result in a falling tone on the sandhi vowel.

This patterning suggests that there is tonal contrast in vocalic ghost segments, at least for those preceded by a low tone, such as (100) and (101). To explain this kind of variation, authors have suggested that ghost vowels are underlyingly specified for tone (Kouneli 2021; Creider & Creider 1989). Although the hypothesis that all ghost segments are specified for tone is compelling, a full analysis of tone and tone sandhi interactions in Sengwer would be necessary to prove it.

Nevertheless, this patterning suggests that certain tone sequences influenced the deletion of stem-final short-vowelled open syllables. For instance, based on their sandhi behaviour, the underlying tone specification expected for (100) would be sòt(ò)—with a low-toned ghost vowel—while that of (101) would be tʌ̀ʌk(-î)—with a falling-toned ghost vowel. However, these two underlying patterns for nouns (CV̀.CV̀ and CV̀.CV̂), though not common, can be found in the dataset and the nouns in which they occur cannot be proven to be loanwords. Therefore, we must assume that the lexical tone specifications have shifted enough since this Proto-Kalenjin deletion process took place that it is no longer possible to predict the specific environment in which it took place.

Still, it is possible to make some informative observations. First, three out of four tone patterns (100, 102, 103) involve a lowering of pitch in the ghost syllable while the pattern in (101) only occurs in a handful of nouns which have the plural suffix -î. Since the latter almost exclusively triggers a high tone in the preceding syllable, making its environment almost always that of (102), we could consider these few cases of CV̀.CV̂ as exceptions in which the replacive high tone was blocked by the stem. Second, the deletion event did not affect nouns in the nominative case, where all targeted segments receive a high tone. Therefore, it appears that the lowering of pitch is another contributing factor in the deletion of ghost segments. This is not surprising considering that a lowering of pitch corresponds to a lowering in saliency of a particular unit.

In summary, these observations mean that, although it is not possible to outline the exact phonological environments which triggered them, there were two separate deletion events: (a) a first more widespread and regular process which targeted stem-final open syllables with a short vowel and a lower pitch than the previous vowel and (b) a second more restricted and irregular process which targeted stem-final consonants (particularly those preceded with a falling tone and a long vowel) in high-frequency morphemes.

In light of this, we argue that Sengwer ghost segments are a type of ghost segments which is lexically determined and not markedness-determined. While the deletion of Sengwer ghost segments can be traced back to phonological processes in linguistic history of Sengwer which were likely markedness determined, these are no longer active synchronically and their alternations have become fossilised. Therefore, building on Zimmermann’s (2019), we can expand the definition as follows (the addition is highligted in italics):

“Ghost segments are segments that (1) are idiosyncratically bound to specific morphemes and (2) alternate with zero in a way that the majority of segments within this language do not. These alternations can either be determined lexically or be conditioned by phonological markedness constraints.”

This extension to the definition specifies the two scenarios in which ghost segments can be found, explicitly including both cases.