1 Introduction

Tone sandhi processes, broadly speaking, can be classified based on whether they preserve the underlying tone of the leftmost or rightmost item in a sandhi domain (Yue-Hashimoto 1987; Zhang 2007). These two classes have been correlated with a strong typological trend: “left-dominant” systems typically involve spread of the tone from the leftmost item across the domain, while “right-dominant” systems typically involve paradigmatic substitution of tones on all but the rightmost item. Shanghai Wu is a typical example of a left-dominant system (1a): The tone on the first syllable spreads across the disyllabic word, neutralizing the tone on the second syllable (Zee & Maddieson 1980; Zhu 1999; 2006). By contrast, Mandarin Tone 3 sandhi is an example of a right-dominant system (1b): The dipping tone 213 is substituted with a rising tone 35 exactly when followed by another 213; the rightmost 213 is preserved.

(1) a. Left-dominant: Shanghai Wu
    51-X 55-31
  b. Right-dominant: Mandarin
    213 35/___ 213

The focus of the present study is Khoekhoegowab, an understudied Khoisan language from Namibia with about 200,000 speakers (Lewis, Simons & Fennig 2016). Khoekhoegowab (also called Khoekhoe) has a sandhi process of opaque melodic substitution (Haacke 1999; Brugman 2009). As typically described, this process is typologically unusual in that it is left-dominant but involves paradigmatic substitution: The leftmost word in each domain keeps its underlying tonal melody, while all other words have their melody replaced. For example, (2) shows that only the leftmost word in a DP retains its underlying tone, while all other words undergo sandhi. In this example, all of the words are underlying high-rising [45]; sandhi causes this melody to be replaced with a level low tone [2] whenever the word is not leftmost in the DP.1

    1. (2)
    1. Sandhi in DPs (citation forms highlighted):          Brugman (2009)
    1.  
    1. a.
    1. súűku
    2. pots
    1.  
    1. b.
    1. ǀápa̋
    2. red
    1. sùùku
    2. pots
    1.  
    1. c.
    1. ǃnáni̋
    2. six
    1. ǀàpa
    2. red
    1. sùùku
    2. pots
    1.  
    1. d.
    1. ǁnáa̋
    2. those
    1. ǃnàni
    2. six
    1. ǀàpa
    2. red
    1. sùùku
    2. pots

There is a wrinkle in the description of Khoekhoe sandhi as left-dominant, however: In the default SOV word order, verbs show anomalous behavior. Prior descriptions of Khoekhoe disagree on the distribution of verbal sandhi. Brugman (2009) finds that verbal sandhi is determined purely by the syntax: Verbs in matrix clauses undergo sandhi, while verbs in embedded clauses do not. Haacke (1999), by contrast, finds that verbal sandhi is purely determined by the linear order of elements in the clause: If the verb is preceded by a tense-marking auxiliary, it will undergo sandhi; if it is followed by such an auxiliary, it will not.

These two descriptions lead us to quite different conclusions about the nature of Khoekhoe sandhi. If Haacke is right, then Khoekhoe sandhi is post-syntactic and left-dominant: The relevant sandhi domain for the verb also includes the tense marker, and so the verb will undergo sandhi whenever it fails to be leftmost in that domain. By contrast, if Brugman is right then the relevant generalization is a purely syntactic one: Certain syntactic configurations (such as embedding) control whether the citation or sandhi form of a word is inserted, making Khoekhoe neither left- nor right-dominant as such.

This paper presents a prosodic production experiment designed to adjudicate between these two analyses. The results of this experiment support a hybrid generalization: tense marker position controls verbal sandhi in matrix clauses (as in Haacke 1999), but embedded verbs always resist sandhi (as in Brugman 2009). This complicates the issue of Khoekhoegowab’s relevance to the generalizations described in Zhang (2007) about left- and right-dominant systems.

The rest of this paper will proceed as follows. In Section 2, I will present the basic facts of Khoekhoegowab tone sandhi and discuss the generalizations proposed for verbal sandhi proposed by Brugman and Haacke. In 3, I will describe the design & methodology used for a prosodic production experiment aimed at deciding between the prior analyses of Khoekhoegowab verbal sandhi. Section 4 presents the results of this experiment, and Section 5 discusses some implications of Khoekhoegowab sandhi for our typology of tone sandhi and avenues for future research.

2 Background: Khoekhoegowab tone sandhi

All lexical items in Khoekhoegowab are associated with one of six tonal classes2; each tonal class is, in turn, associated with a particular tonal melody made up of a sequence of at most two out of the four contrastive tone levels. The word will be produced with this melody, called the “citation melody”, in isolation or in certain prosodically strong positions (defined in more detail below). The citation melodies are given in Table 1 along with a near-minimal sextuplet illustrating the contrast.

Table 1

Citation melodies (Brugman 2009).

Melody Description Example Gloss
[1] Superlow [!ȁas] ‘servant’
[2] Low [ǁàas] ‘tie’
[12] Low-rising [!nȁàs] ‘story’
[4] High [ǂáas] ‘plain’
[5] Superhigh [!na̋as] ‘tortoise’
[45] High-rising [ǂáa̋s] ‘spittle’

As noted, the citation melody only surfaces in certain prosodic contexts; in most contexts a process of tonal sandhi applies. Sandhi is an opaque tonal substitution process mapping each of the six citation melodies onto another, apparently arbitrary melody. Sandhi can broadly be characterized as a weakening process in the sense that it reduces the number of cross-linguistically marked tonal melodies: The inventory of sandhi melodies has lower register overall than the inventory of citation melodies and contains fewer rising contours (which are cross-linguistically marked, see e.g. Yip (2002). The six citation melodies and their sandhi counterparts are given in Table 2. Note that some citation tones (namely the low-rising and low-level tones) are unaffected by sandhi. Elsewhere, the effect of sandhi is unpredictable: Level tones become contours and vice versa; high-register tones sometimes become low-register ones and sometimes do not; some contrasts are neutralized while others are maintained. Sandhi has the effect of neutralizing the contrast between the superlow and high tone classes, and also between the low and high-rising. In at least one case, sandhi involves apparent underapplication opacity (‘counterfeeding’): Underlying high tone becomes low-falling; but underlying superhigh tone becomes high without continuing on to become low-falling.

Table 2

Sandhi forms.

Citation Sandhi
Low-rising [12] [12] Low-rising
Superlow [1] [21] Low-falling
High [4] [21] Low-falling
Low [2] [2] Low
High-rising [45] [2] Low
Superhigh [5] [4] High

2.1 Sandhi domains

I have said that the citation melodies appear in prosodically strong positions, while sandhi applies everywhere else. It’s time to make that more precise. Within the nominal domain, the generalization is clear: The leftmost item in a DP (or PP) receives citation form, while all other items undergo sandhi3. This is illustrated with a set of DPs in (3), repeated from example (2). In (a) the noun surfaces with its citation melody; in (b), only the adjective ‘red’ takes citation form, while the noun undergoes sandhi; in (c) only the numeral ‘six’ keeps its citation form while both ‘red’ and ‘pots’ undergo sandhi; and in (d) only determiner ‘those’ keeps citation form while all other words take sandhi.

    1. (3)
    1. Sandhi in DPs (citation forms highlighted):          Brugman (2009)
    1.  
    1. a.
    1. súűku
    2. pots
    1.  
    1. b.
    1. ǀápa̋
    2. red
    1. sùùku
    2. pots
    1.  
    1. c.
    1. ǃnáni̋
    2. six
    1. ǀàpa
    2. red
    1. sùùku
    2. pots
    1.  
    1. d.
    1. ǁnáa̋
    2. those
    1. ǃnàni
    2. six
    1. ǀàpa
    2. red
    1. sùùku
    2. pots

Put another way, each maximal DP (or PP) is mapped onto a single sandhi domain. Within a sandhi domain, the leftmost position is “strong” in the sense that it resists sandhi and retains its lexically-specified form; all words not in that strong position lose their citation form and take on their sandhi form.

The association between the left edge of phrases and citation melody is preserved when the verb is moved to the left perhiphery (and thus winds up at the left edge of the clause): In this context, the verb takes citation melody regardless of what occurs later in the clause. In (4a), the verb khomai ‘read’ takes its citation tone (superhigh [5]) when fronted; (4b) shows a context in which it takes its sandhi tone (high [4]) in its base, clause-final position. This shows that verbs are subject to the same sandhi process affecting the nominal domain, and that when there is no material which could possibly precede the verb in the sandhi domain, the verb resists sandhi just as expected.

    1. (4)
    1. a.
    1. Khőmai
    2. read
    1. go
    2. pst
    1. =b
    2. =3ms
    1. ge
    2. decl
    1. Dandagoba
    2. D.
    1. ǂkhanisa.
    2. book
    1. “Dandago read the book.”
    1.  
    1. b.
    1. Dandagob
    2. D.
    1. ge
    2. decl
    1. ǂkhanisa
    2. book
    1. go
    2. pst
    1. khómai.
    2. read.
    1. ‘Dandago read the book.’

The situation becomes more complex when we consider in situ verbs, however. Previous works on verbal sandhi give contradictory generalizations. Brugman (2009) states that all root-clause (in situ) verbs undergo sandhi4, while all embedded clause verbs retain their citation form. She does, however, note the fact mentioned above, that in sentences like (4a) where the verb has been fronted, it retains citation form regardless of the tense marker, indicating that verbs are still subject to the normal prosodic rules in root clauses. Brugman captures the difference between root and embedded clauses by proposing an Agree relation in the syntax between the complementizer and the verb which, in embedded contexts, marks the verb as “sandhi-resistant”, preventing it from undergoing sandhi even it is not leftmost in the sandhi domain. In sum, Brugman proposes that syntax plays a crucial role in determining the tonal melody of the verb, independent of other prosodic concerns.

By contrast, Haacke (1999) gives a generalization purely based on the linear order of elements. The determining factor, for Haacke, is the placement of tense-marking. Khoekhoegowab marks tense, aspect, and polarity with a set of auxiliaries. These auxiliaries come in two classes. One class of auxiliaries appears postverbally (and generally clause-finally when the verb is in situ); the other class appears before the verb, encliticizing to some XP appearing before the verb. In both cases, the tense marking and the verb may be separated by other elements in the clause. For example, (5) and (6) show two coordinated VPs. In (5), the tense marker tama ‘negative non-future’, which belongs to the postverbal class, appears clause-finally, and is thus separated from the first verb huni ‘stir’ by the entire second conjunct. In contrast, (6) shows that the tense marker go ‘past’, which belongs to the preverbal class, may freely encliticize to either the first or the second object, with no change in meaning. If it encliticizes to the second object as in (6a), it is separated from the first verb; if it encliticizes to the first object as in (6b), it is separated from the second verb.

    1. (5)
    1. Aob
    2. man
    1. ge
    2. decl
    1. mai-e
    2. pap
    1. húni̋
    2. stir
    1. tsi
    2. and
    1. ǁgan-e
    2. meat
    1. ám̋
    2. grill
    1. tama.
    2. neg.nf
    1. “The man didn’t stir the pap or grill the meat.”
    1. (6)
    1. a.
    1. Aob
    2. man
    1. ge
    2. decl
    1. mai-e
    2. pap
    1. húni̋
    2. stir
    1. tsi
    2. and
    1. ǁgan-e
    2. meat
    1. go
    2. pst
    1. àm.
    2. grill
    1. “The man stirred the pap and grilled the meat.”
    1.  
    1. b.
    1. Aob
    2. man
    1. ge
    2. decl
    1. mai-e
    2. pap
    1. go
    2. pst
    1. hùni
    2. stir
    1. tsi
    2. and
    1. ǁgan-e
    2. meat
    1. àm.
    2. grill
    1. ‘The man stirred the pap and grilled the meat.’

Kusmer (2019) argues that the position of the tense marker in Khoekhoegowab is a morphophonological fact rather than a syntactic one.5 This conclusion is based on the fact that there is no systematic change in meaning or syntactic structure between clauses bearing preverbal or postverbal tense markers (beyond that attributable to the denotation of the tense marker itself). Put another way, neither the preverbal nor the postverbal class of tense marker forms a morphosyntactic natural class. Tables 3 and 4 give a complete list of all tense markers, separated by class; it can easily be confirmed that there is no one tense, aspect, or polarity feature that defines either class. Instead, the determining factor is a phonological one: Preverbal tense markers are all monomoraic, while postverbal ones are all bimoraic. Kusmer (2019) argues that the preverbal tense markers are postsyntactically-displaced into that position for prosodic reasons, but syntactically originate in the same position as the postverbal markers.

Table 3

Preverbal tense markers.

IPA Gloss
a [ra] present stative
ra/ta [ra]/[ta] imperfect
ge [ke] remote past
go [ko] recent past
ni [ni] future
ta [ta] negative non-finite
ga [ka] irrealis
Table 4

Postverbal tense markers.

IPA Gloss
tama [tama] non-future negative
tide [tite] future negative
i [iː] non-present stative
hâ [hãː] perfect

Haacke (1999) states that the tonal melody of the verb is determined by whether tense marking is preverbal or postverbal. Because the position of tense marking is determined postsyntactically, Haacke’s analysis thus holds that sandhi is a purely post-syntactic process. His analysis also maintains the characterization of Khoekhoegowab sandhi as “left-dominant”: If the verb and the tense marker are assumed to form a sandhi domain together, then the verb can only be leftmost in that domain (and thus resist sandhi) if tense marking is postverbal.

Brugman (2009) and Haacke (1999) thus present very different generalizations for Khoekhoegowab sandhi, with implications for its analysis.6 These differences are summarized in Tables 5 and 6.

Table 5

Brugman’s generalization.

Matrix Embedded
Preverbal tense Sandhi Citation
Postverbal tense Sandhi Citation
Table 6

Haacke’s generalization.

Matrix Embedded
Preverbal tense Sandhi Sandhi
Postverbal tense Citation Citation

In order to resolve the conflict between these generalizations, I conducted a prosodic production experiment, to be described in the next two sections. To preview the results, the final generalization resulting from this experiment is as follows: Root clause verbs undergo sandhi whenever they are preceded by a tense marker; embedded clause verbs do not undergo sandhi except in quotative clauses (marked with a special complementizer), where they behave like root verbs. This generalization is summarized in Table 7.

Table 7

Results of experiment.

Matrix Embedded
Preverbal tense Sandhi Citation
Postverbal tense Citation Citation

3 Experimental design & methodology

3.1 Speakers

The experimental subjects were 4 native speakers of Khoekhoegowab (3f, 1m), between the ages of 18 & 30. All speakers resided in Windhoek. Two were current UNam graduate students studying Khoekhoegowab; the others were recruited from the author’s prior fieldwork consultants.

3.2 Stimuli

The primary experimental manipulation was the position of tense marking. 15 pairs of sentences differing only in the position of tense marking were constructed, yielding 30 total test items. All of the sentences used the verbs listed in Table 8; these verbs were selected to be mostly sonorant7 (to aid in F0 tracking) and to have either High or High-Rising citation melodies, which are the two melodies showing the most detectable change under sandhi. Sample pitch tracks for each verb, all taken from the same speaker, are presented in Figure 1. In addition to the test items, 12 filler pairs (24 items) were added, which differed only in whether the direct object of the verb had scrambled past another XP; fillers thus superficially resembled test items in showing only word-order differences. Between fillers and test items, there were 54 items in total.

Table 8

List of verbs in experimental items.

Verb Gloss Citation Sandhi
oa ‘return’ High [4] Low-falling [21]
ā ‘cry’
om ‘build’
mû ‘see’ High-rising [45] Low [2]
huni ‘stir’
am ‘grill’
Figure 1
Figure 1

Sample pitch tracks for all six verbs, taken from the same speaker. Solid lines are citation form; dotted lines are sandhi form. Verbs in the left column alternate between [4] & [21]; verbs in the right column alternate between [45] & [2].

The test items were further subdivided into 6 syntactic frames, 3 matrix and 3 embedded: Matrix declarative clauses (7); matrix constituent Question clauses (8); Relative clefts (9);8 Nominalized embedded SOV clauses (10); Quotative embedded SOV clauses (11); and matrix VP Coordination clauses (12).9

The VP coordination syntactic frame had one systematic difference from the others: Because there were two verbs, tense marking could be in three locations: Before both verbs (12)[a]; between the verbs (12)[b]; or after both verbs (12)[c]. Because of this, test items in this syntactic frame were constructed in triplets (rather than pairs as described above); in the final analysis, each verb was treated as a separte trial and coded as either preverbal or postverbal.

    1. (7)
    1. Matrix
    1.  
    1. a.
    1. Khoeb
    2. man
    1. ge
    2. decl
    1. oms
    2. home
    1. ǀkha
    2. to
    1. go
    2. pst
    1. oa.
    2. return
    1. ‘The man went home.’
    1.  
    1. b.
    1. Khoeb
    2. man
    1. ge
    2. decl
    1. oms
    2. home
    1. ǀkha
    2. to
    1. oa
    2. return
    1. tama.
    2. neg.nf
    1. ‘The man didn’t return home.’
    1. (8)
    1. Question
    1.  
    1. a.
    1. ǁNa
    2. that
    1. tarasa
    2. woman
    1. go
    2. pst
    1. tae-e
    2. what
    1. am?
    2. grill
    1. ‘What did that woman grill?’
    1.  
    1. b.
    1. ǁNa
    2. that
    1. tarasa
    2. woman
    1. tae-e
    2. what
    1. am
    2. grill
    1. tama?
    2. neg.nf
    1. ‘What didn’t that woman grill?’
    1. (9)
    1. Relative
    1.  
    1. a.
    1. Oms
    2. home
    1. ǀkha
    2. to
    1. go
    2. pst
    1. oa
    2. return
    1. khoeb
    2. man
    1. ge.
    2. decl
    1. ‘It was that man who returned home.’
    1.  
    1. b.
    1. Oms
    2. home
    1. ǀkha
    2. to
    1. oa
    2. return
    1. tama
    2. neg.nf
    1. khoeb
    2. man
    1. ge.
    2. decl
    1. ‘It was the man who didn’t return home.’
    1. (10)
    1. Nominalized
    1.  
    1. a.
    1. Mî
    2. say
    1. ta
    2. I
    1. ge
    2. decl
    1. ra
    2. imp
    1. Dandagob
    2. D.
    1. go
    2. pst
    1. oms
    2. home
    1. ǀkha
    2. to
    1. oa
    2. return
    1. sa.
    2. -comp
    1. ‘I am saying that Dandago returned home.’
    1.  
    1. b.
    1. Mî
    2. say
    1. ta
    2. I
    1. ge
    2. decl
    1. ra
    2. imp
    1. Dandagob
    2. D.
    1. oms
    2. home
    1. ǀkha
    2. to
    1. oa
    2. return
    1. tama
    2. neg.nf
    1. sa.
    2. comp
    1. ‘I am saying that Dandago didn’t return home.’
    1. (11)
    1. Quotative
    1.  
    1. a.
    1. Mî
    2. say
    1. ta
    2. I
    1. ge
    2. decl
    1. ra
    2. imp
    1. arib
    2. dog
    1. ge
    2. decl
    1. ǀhôasa
    2. cat
    1. go
    2. pst
    1. mû
    2. see
    1. ti.
    2. c.quot
    1. ‘I am saying that the dog saw the cat.’
    1.  
    1. b.
    1. Mî
    2. say
    1. ta
    2. I
    1. ge
    2. decl
    1. ra
    2. imp
    1. arib
    2. dog
    1. ge
    2. decl
    1. ǀhôasa
    2. cat
    1. mû
    2. see
    1. tama
    2. neg.nf
    1. ti.
    2. c.quot
    1. ‘I am saying that the dog didn’t see the cat.’
    1. (12)
    1. Coordination
    1.  
    1. a.
    1. Aob
    2. man
    1. ge
    2. decl
    1. mai-e
    2. pap
    1. go
    2. pst
    1. huni
    2. stir
    1. tsi
    2. and
    1. ǁgan-e
    2. meat
    1. am.
    2. grill
    1. ‘The man stirred the pap and grilled the meat.’
    1.  
    1. b.
    1. Aob
    2. man
    1. ge
    2. decl
    1. mai-e
    2. pap
    1. huni
    2. stir
    1. tsi
    2. and
    1. ǁgan-e
    2. meat
    1. go
    2. pst
    1. am.
    2. grill
    1. ‘The man stirred the pap and grilled the meat.’
    1.  
    1. c.
    1. Aob
    2. man
    1. ge
    2. decl
    1. mai-e
    2. pap
    1. huni
    2. stir
    1. tsi
    2. and
    1. ǁgan-e
    2. meat
    1. am
    2. grill
    1. tama.
    2. neg.nf
    1. ‘The man didn’t stir the pap or grill the meat.’

These 6 syntactic frames were selected to fully distinguish between the two prior analyses. Most embedded clauses in Khoekhoegowab are nominalized; the contrast between the matrix and nominalized frames is thus crucial. Under Brugman’s analysis, all items in the matrix frame should undergo sandhi, while no items in the nominalized frame should; under Haacke’s analysis the items with preverbal tense marking in both frames should show sandhi, while the items with postverbal tense marking should not.

The other syntactic frames are present in order to test variations on the two analyses. Matrix declaratives in Khoekhoegowab always have a second-position clitic marking the clause type (Hagman 1977); embedded clausess do not have such a marker. A possible variation on Brugman’s analysis is to hypothesize that it is the presence or absence of such a marker that correlates with verbal sandhi, not the clause type itself. Matrix questions in Khoekhoegowab typically lack a clause-type marker (and thus superficially resemble embedded clauses); by contrast, quotative embedded clauses, which take a special complementizer only available under verbs of reported speech, exceptionally do take a clause-type marker (and thus superficially resemble matrix clauses). If it is the clause-type marker that controls verbal sandhi, we predict the quotative frame to uniformly undergo sandhi and the question frame to uniformly fail to do so.

The VP coordination frame serves to disambiguate two interpretations of Haacke’s generalization. In one interpretation, the presence of a tense-marker from the preverbal class that triggers sandhi on the verb regardless of their actual relative positions. In the other interpretation, it is the linear order of tense marker and verb that matters, not the class to which the tense marker belongs. If the former analysis is correct, preverbal tense markers will trigger sandhi on the first verb even when they linearly follow it; if the latter analysis is correct, preverbal tense markers will only trigger sandhi on that verb when they linearly precede it.

Finally, the relative cleft frame serves to confirm that it is embedded clauses in general, rather than nominalized clauses in specific, that resist sandhi under Brugman’s analysis.

An unavoidable confound in this design must be noted at this point. Declination (see e.g. Connell 2001) is a phonetic process which tends to lower & constrict F0 range over the course of a sentence. In the case of Khoekhoegowab, this may perceptually neutralize the tonal contrasts relevant to sandhi towards the end of a long utterance. Due to unavoidable syntactic constraints on word order, all verbs in the relative clause condition and the first verb in the Coordination condition are closer to the start of the sentence and thus will be less affected by declination; on a smaller scale, all verbs preceded by tense markers will be later in the sentence than verbs with postverbal tense markers, and so may also be more effected by declination. Because sandhi also has the effect of lowering F0, this means that verbs in items with preverbal tense markers may be more likely to be transcribed as having undergone sandhi, while verbs relative clause items, in the first position in coordination items, and in postverbal tense-marker items may be less likely to be transcribed as such. This confound is sadly unavoidable given the constraints of the language, and as discussed in Section 4 this does seem to have a small effect on the comparison of the nominative & relative frames, where the difference in verb position is largest. However, if we limit ourselves to considering only the nominalized and quotative syntactic frames, we find some reassurance that declination is not wholly responsible for the results. Both of those frames have the target verb in an embedded clause close to the end of the utterance, and so declination should act similarly in both contexts. And yet in the nominalized case we find 36/127 observations transcribed as sandhi, while in the quotative case we find 72/128.10 This points to a difference between the two cases which cannot be explained by declination.

A full list of all stimuli, including fillers, is presented in the appendix.

3.3 Procedure

Sentences were presented on a laptop screen; only one sentence was on screen at time, and speakers could advance to the next sentence at their own pace. Each speaker saw all 54 sentences in a random order, and were then instructed to take a short break, after which this was repeated with a different randomized order such that each speaker saw each item twice. The entire procedure took between 15 and 30 minutes, depending on speaker.

Speakers were asked to read each sentence aloud as naturally as possible. The sentences were all recorded on a Zoom H5 recorder using a Shure SN10A-CN head-mounted microphone.

3.4 Analysis

After recording, individual items were segmented and then force-aligned using the Montreal Forced Aligner (McAuliffe et al. 2017), which was trained on a dataset of the author’s fieldwork elicitation encompssing roughly 4.5 hours of transcribed Khoekhoegowab speech from 8 speakers. After alignment, the TextGrid boundaries of each verb were hand-adjusted in Praat Praat (Boersma & Weenink 2001) and a script was used to extract the audio of each verb token into its own file; in this process, 5 tokens were rejected because the resulting recording was inaudible due to the speaker reducing the verb.11 The remaining 283 tokens were coded for tense position (preverbal or postverbal) and syntactic frame. Tokens from the VP coordination frame were coded based on whether the tense marker linearly preceded the verb in question, not whether the tense marker was drawn from the preverbal or postverbal class. For example, in (13) the first verb huni ‘stir’ was coded as having postverbal tense marking because go ‘past’ linearly follows it, even though go is from the preverbal class. (Am ‘grill’ was coded as preverbal, as normal.)

    1. (13)
    1. Aob
    2. man
    1. ge
    2. decl
    1. mai-e
    2. pap
    1. huni
    2. stir
    1. tsi
    2. and
    1. ǁgan-e
    2. meat
    1. go
    2. pst
    1. am.
    2. grill
    1. ‘The man stirred the pap and grilled the meat.’

To exclude the possibility of confirmation bias in my own transcriptions, I used the following procedure to code the results: Three phonetically-trained naive transcribers (all native English speakers with no prior experience Khoekhoegowab) were asked to sort the tokens into “high” (citation form) and “low” (sandhi form). Transcribers were given the tokens sorted by speaker and lexical item, with all information about syntactic frame and tense-marker position removed, so as to blind them to the experimental manipulation. Additionally, I hand selected two tokens of each surface tone contour used in the experiment (High-rising, Low, High, & Low-falling) that I felt were prototypical examples, to serve as reference points for the transcribers. To provide one additional datapoint, I performed the same blind transcription.

There was broad agreement between the transcribers; the transcriptions overall showed a Fleiss’ Kappa12 of 0.77, indicating substantial agreement. What disagreement exists is likely due to the effects of voice quality obscuring perceptions of tone; in particular, Speaker 3 spoke predominantly in breathy voice, while Speaker 4 spoke primarily in creak.

In order to confirm that the transcribers were attending to the intended phonetic differences, the smoothed mean pitch tracks in Figure 2 were created. A Praat script was used to extract F0 at 20 evenly-spaced points across each verb. For the purpose of constructing these graphs, individual recordings were treated as having undergone sandhi only if a majority of transcribers marked that item as “low”; all others were treated as having citation form. Loess smoothing was used to construct an average pitch track across all items. From this, it can be seen that transcribers are in fact distinguishing the citation and sandhi forms: For both tone classes the citation forms (HR and H) are distinctly higher than the sandhi forms (L and LF); HR does show a distinctive final rise, while H is level. Both the L and LF forms fall only slightly, but are still distinguishable by level.

Figure 2
Figure 2

Mean pitch tracks.

3.5 Results

Table 9 shows summary statistics for the transcribed data, broken down by frame & tense-marker position, including the total number of observations and the number and percentage of observations transcribed as having undergone sandhi. Note that two of the Frame conditions, Matrix & Coordinated, have more observations than the other three Frames. In the case of the Coordinated frame, this is the result of each item having two verbs and three possible tense-marker positions; see discussion in Section 3. In the case of the Matrix condition, because these are the “baseline” unmarked condition I initially included more of them in the experimental design. Post-hoc exclusion of half of the Matrix items did not change the effect sizes in the final analysis, so all further discussion here will include all observations. Finally, two observations were excluded — from Matrix Post and Nominalized Post — due to the speaker entirely devoicing the vowel in production, making it impossible to perceive F0.

Table 9

Summary statistics.

Frame Position Observations Sandhi %
MATRIX POST 127 16 12.6%
PRE 128 122 95.3%
QUESTION POST 64 5 7.8%
PRE 64 57 89.0%
QUOTATIVE POST 64 11 17.2%
PRE 64 61 95.3%
COORDINATED POST 192 13 6.8%
PRE 176 146 83.0%
NOMINALIZED POST 63 14 22.2%
PRE 64 22 34.4%
RELATIVE POST 64 1 1.5%
PRE 64 2 3.1%

4 Analysis

Having confirmed that transcribers were distinguishing the relevant tone classes, the hypotheses discussed above were tested against these blind transcriptions by fitting a mixed-effects logistic regression model. The dependent variable was whether a given observation was transcribed as “low” (i.e. “sandhi”); the model looked for fixed effects of syntactic frame (6 levels: Matrix, Question, Coordinated, Quotative, Nominalized, & Relative) and tense marker position (2 levels: Pre and Post), plus interactions between these. The model included random intercepts for Speaker, Item, & Transcriber.

(14) Model:
  Sandhi ~ Frame*Position + (1|Speaker) + (1|Transcriber) + (1|Item)

In order to distinguish the various alternatives to Brugman’s generalization, a custom contrast matrix (Bruin 2011) was used for the syntactic frame variable to make the following comparisons:

(15) a. Group mean of Matrix, Question, Coordinated, & Quotative (‘matrix-like’ clauses) vs. group mean of Nominalized & Relative
  b. Mean of Matrix vs. mean of Question
  c. Mean of Matrix vs. mean of Quotative
  d. Mean of Matrix vs. mean of Coordinated
  e. Mean of Nominalized vs. mean of Relative

This model allows us to distinguish between 3 competing hypotheses (and some subcases):

(16) a. Hypothesis A: Haacke’s generalization
    The verb undergoes sandhi iff…
    (i) …it is preceded by tense-marking.
      Prediction: Main effect of Position; no main effect of Frame[d].
    (ii) A: …it is associated with a tense-marker from the “preverbal” class.
      Prediction: Main effects of Position and Frame[d].
  b. Hypothesis B: Brugman’s generalization
    The verb undergoes sandhi iff…
    (i) …it is in a matrix-like clause.
      Prediction: Main effect of Frame[a]
    (ii) B: …it is in a clause with a second-position clause type marker.
      Prediction: Main effect of Frame[b]; no main effect of Frame[c].
  c. Hypothesis C: Hybrid
    The verb undergoes sandhi iff it is both preceded by tense marking and in a matrix-like clause.
    Prediction: Main effect of Position and interaction between Position & Frame[a].

The fixed effects of the model are presented in Table 10; a summary of the random effects is presented in Table 11.

Table 10

Significant coefficients.

Estimate Std. Error z value Pr (>|z|)
(Intercept) –2.53214 0.33554 –7.547 4.47e–14 ***
Pos[Pre] 3.56074 0.38175 9.327 <2e–16 ***
Frame[a] 0.51619 0.66648 0.775 0.4386
Frame[b] 0.56716 0.77031 0.736 0.4616
Frame[c] 0.75369 0.57108 1.320 0.1869
Frame[d] –0.19415 0.71939 –0.270 0.7873
Frame[e] 2.97891 1.22747 2.427 0.0152 *
Frame[a]:Pos[Pre] 4.28657 0.88816 4.826 1.39e–06 ***
Frame[b]:Pos[Pre] 0.43218 1.10939 0.390 0.6969
Frame[c]:Pos[Pre] 0.63640 0.84772 0.751 0.4528
Frame[d]:Pos[Pre] 0.16661 1.16809 0.143 0.8866
Frame[e]:Pos[Pre] –0.01353 1.57806 –0.009 0.9932
Table 11

Random effects.

Group Variance Std. Dev.
Item 0.40030 0.6327
Speaker 0.15611 0.3951
Transcriber 0.07762 0.2786

The significant main effect of position is compatible with Hypothesis A (Haacke’s generalization). The positive coefficient indicates that preverbal tense-marker position does correlate with higher rates of sandhi on the verb. That there is no main effect of Frame[d] supports Hypothesis A over Hypothesis A’ — it is the absolute position of the tense marker with respect to the verb that matters, not which positional class it belongs to.

The lack of significance for a main effect of Frame[a] (which compares matrix-like syntactic frames to embedded ones) is incompatible with Hypothesis B (Brugman’s generalization): If sandhi were conditioned by the embedded status of the clause, this coefficient should be significantly positive. Similarly, the lack of a main effect of Frame[b] is incompatible with Hypothesis B′.

However, there is also a significant interaction between Frame[a] (which compares “matrix-like” syntactic frames to embedded clauses) and tense marker position. The positive coefficient indicates that transcribers were significantly more likely to mark a verb as having undergone sandhi if it was in a matrix-like syntactic frame and had preverbal tense-marking. This is compatible with Hypothesis C, the hybrid model: preverbal tense markers trigger sandhi on the verb only in matrix-like clauses; embedded clauses systematically resist sandhi, regardless of tense marker position.

Frame[e] compares Nominalized and Relative clauses; this comparison was included in order to confirm whether or not it is embedded clauses in general (rather than nominalized clauses in specific) that resist sandhi. The significance of this comparison, however, is likely due to the confound mentioned in section 3. Examples of both these syntactic frames are repeated below, with the verb highlighted. Note that in the Relative case, the verb is significantly closer to the start of the utterance than in the Nominalized case. This means that declination (see e.g. Connell 2001) has had longer to apply in the Nominalized case; in other words, the overall F0 range of verbs will be both smaller and lower in the Nominalized case than the Relative one. This likely lead to more verbs being transcribed as low (i.e. having undergone sandhi) regardless of ground truth.

    1. (17)
    1. a.
    1. Relative:
    1. Oms
    2. home
    1. ǀkha
    2. to
    1. go
    2. pst
    1. oa
    2. return
    1. khoeb
    2. man
    1. ge.
    2. decl
    1. ‘It was that man who returned home.’
    1.  
    1. b.
    1. Nominalized:
    1. Mî
    2. say
    1. ta
    2. I
    1. ge
    2. decl
    1. ra
    2. imp
    1. Dandagob
    2. D.
    1. go
    2. pst
    1. oms
    2. home
    1. ǀkha
    2. to
    1. oa
    2. return
    1. sa.
    2. -comp
    1. ‘I am saying that Dandago returned home.’

Overall, then, the results of the model support the hybrid model Hypothesis C: In most embedded clause types, verbs resist sandhi; elsewhere, verbs undergo sandhi exactly when preceded by tense marking.13

5 Discussion

Khoekhoegowab sandhi, at first glance, appears to be left-dominant in the sense discussed by Zhang (2007): Within some domain, the leftmost item retains its underlying tone while all other items undergo sandhi. However, Khoekhoegowab is typologically unusual within this class: left-dominant sandhi systems most typically involve spreading of the leftmost tone over the non-leftmost elements; Khoekhoe sandhi instead involves paradigmatic melodic substitution, which is typically characteristic of right-dominant systems.

Khoekhoegowab verbs present a problem for the characterization of this sandhi process as left-dominant. The experiment reported here shows that verbal sandhi obeys the generalization repeated in Table 12. In matrix clauses, verbal sandhi is plausibly left-dominant: If the verb and its tense marking are taken to form a sandhi domain14, then the verb will only be leftmost in that domain when it precedes the tense marker. However, this apparent relationship is disrupted in embedded clauses: In most embedded clause types, verbs resist sandhi regardless of the position of tense. This draws into question the relevance of Khoekhoegowab sandhi to the typology discussed in Zhang (2007) and elsewhere.

Table 12

Results of experiment (repeated from Table 7).

Matrix Embedded
Preverbal tense Sandhi Citation
Postverbal tense Citation Citation

5.1 Variation

The data reported here expands on previous descriptions of Khoekhoegowab prosody in another respect: All previous descriptions have said that Khoekhoe sandhi is categorical15, while the results of this experiment leave open the possibility that it is variable: No two transcribers agreed on 100% of the tokens.

Some of this variability is certainly due to transcriber error. All transcribers were non-Khoekhoegowab-speaking, and thus it is highly likely that the transcriptions are not perfectly accurate to the true phonological form of the token. That is, there certainly some tokens which have phonologically undergone sandhi but which were transcribed as having citation form, and vice versa.

But transcriber error cannot fully explain the variability in the data. For example, Figure 3 presents the F0 pitch tracks for two tokens of the same verb from the same speaker in the same condition (one from each block) — in both cases, the sentence in (18). Even if we allow for variability in F0 range between utterances, the two tokens here have different contours; it seems likely that one is High-Rising (citation form) while the other is Low (sandhi form). This seems to be a case of intra-speaker variability in verbal sandhi.

Figure 3
Figure 3

Pitch tracks for two tokens of (18) from the same speaker, showing variation in tone of /am/ ‘grill’.

    1. (18)
    1. Aob
    2. man
    1. ge
    2. decl
    1. mai-e
    2. pap
    1. go
    2. pst
    1. huni
    2. stir
    1. tsi
    2. and
    1. ǁgam-e
    2. meat
    1. am.
    2. grill
    1. ‘The man stirred the pap and grilled meat.’

There are a variety of known sources of variation that this experiment was not designed to control for. For example, speech rate is known to affect prosodic phrasing, with higher speech rates being associated with fewer prosodic boundaries (Fougeron & Jun 1998). It’s possible that, at slower speech rates, speakers may insert a prosodic break before the verb, allowing it to retain citation form (by virtue of being at a left edge) even when the syntactic structure would normally lead to a different prosodic structure. It’s also possible that this variation is either disfluency (i.e. the speaker simply misspoke) or an effect of the experimental task (for example, list intonation).

However, the experimental results do show that the generalizations reported here are strong trends and seem to reflect the normal case. As such, further research is required to determine the sources and extent of variation in Khoekhoegowab tone sandhi.

Matrix:

    1. (19)
    1. Khoeb
    2. man
    1. ge
    2. decl
    1. oms
    2. home
    1. ǀkha
    2. to
    1. go
    2. pst
    1. oa.
    2. return
    1. ‘The man returned home.’
    1. (20)
    1. Khoeb
    2. man
    1. ge
    2. decl
    1. oms
    2. home
    1. ǀkha
    2. to
    1. oa
    2. return
    1. tama.
    2. neg.nf
    1. ‘The man didn’t return home.’
    1. (21)
    1. Khoeb
    2. man
    1. ge
    2. decl
    1. oms
    2. home
    1. ǀkha
    2. to
    1. go
    2. pst
    1. -ro
    2. -imp
    1. oa.
    2. return
    1. ‘The man was returning home.’
    1. (22)
    1. Khoeb
    2. man
    1. ge
    2. decl
    1. oms
    2. home
    1. ǀkha
    2. to
    1. oa
    2. return
    1. hâ.
    2. perf
    1. ‘The man has returned home.’
    1. (23)
    1. ǀGôab
    2. boy
    1. ge
    2. decl
    1. mai-e
    2. pap
    1. go
    2. pst
    1. huni.
    2. stir
    1. ‘The boy stirred the pap.’
    1. (24)
    1. ǀGôab
    2. boy
    1. ge
    2. decl
    1. mai-e
    2. pap
    1. go
    2. pst
    1. -ro
    2. -imp
    1. huni.
    2. stir
    1. ‘The boy was stirring the pap.’
    1. (25)
    1. ǀGôab
    2. boy
    1. ge
    2. decl
    1. mai-e
    2. pap
    1. huni
    2. stir
    1. tama.
    2. neg.nf
    1. ‘The boy didn’t stir the pap.’
    1. (26)
    1. ǀGôab
    2. boy
    1. ge
    2. decl
    1. mai-e
    2. pap
    1. huni
    2. stir
    1. hâ.
    2. perf
    1. ‘The boy has stirred the pap.’

Nominalized:

    1. (27)
    1. Mî
    2. say
    1. ta
    2. I
    1. ge
    2. decl
    1. ra
    2. imp
    1. [
    2.  
    1. Dandagob
    2. D.
    1. go
    2. pst
    1. oms
    2. home
    1. ǀkha
    2. to
    1. oa
    2. return
    1. -sa.
    2. -comp
    1. ]
    2.  
    1. ‘I am saying that Dandago returned home.’
    1. (28)
    1. Mî
    2. say
    1. ta
    2. I
    1. ge
    2. decl
    1. ra
    2. imp
    1. [
    2.  
    1. Dandagob
    2. D.
    1. oms
    2. home
    1. ǀkha
    2. to
    1. oa
    2. return
    1. tama
    2. neg.nf
    1. -sa.
    2. -comp
    1. ]
    2.  
    1. ‘I am saying that Dandago didn’t return home.’
    1. (29)
    1. Mî
    2. say
    1. ta
    2. I
    1. ge
    2. decl
    1. ra
    2. imp
    1. [
    2.  
    1. ǀgôab
    2. boy
    1. go
    2. pst
    1. mai-e
    2. pap
    1. huni
    2. stir
    1. -sa.
    2. -comp
    1. ]
    2.  
    1. ‘I am saying that the boy stirred the pap.’
    1. (30)
    1. Mî
    2. say
    1. ta
    2. I
    1. ge
    2. decl
    1. ra
    2. imp
    1. [
    2.  
    1. ǀgôab
    2. boy
    1. mai-e
    2. pap
    1. huni
    2. stir
    1. tama
    2. tama
    1. sa.
    2. -comp
    1. ]
    2.  
    1. ‘I am saying that the boy didn’t stir the pap.’

Coordination:

    1. (31)
    1. Aob
    2. man
    1. ge
    2. decl
    1. [
    2.  
    1. mai-e
    2. pap
    1. huni
    2. stir
    1. ]
    2.  
    1. tsi
    2. and
    1. [
    2.  
    1. ǁgan-e
    2. meat
    1. go
    2. pst
    1. am.
    2. grill
    1. ]
    2.  
    1. ‘The man stirred the pap and grilled the meat.’
    1. (32)
    1. Aob
    2. man
    1. ge
    2. decl
    1. [
    2.  
    1. mai-e
    2. pap
    1. go
    2. pst
    1. huni
    2. stir
    1. ]
    2.  
    1. tsi
    2. and
    1. [
    2.  
    1. ǁgan-e
    2. meat
    1. am.
    2. grill
    1. ]
    2.  
    1. ‘The man stirred the pap and grilled the meat.’
    1. (33)
    1. Aob
    2. man
    1. ge
    2. decl
    1. [
    2.  
    1. mai-e
    2. pap
    1. huni
    2. stir
    1. ]
    2.  
    1. tsi
    2. and
    1. [
    2.  
    1. ǁgan-e
    2. meat
    1. am
    2. grill
    1. tama.
    2. neg.nf
    1. ]
    2.  
    1. ‘The man didn’t stir the pap or grill the meat.’
    1. (34)
    1. Khoedages
    2. K.
    1. ge
    2. decl
    1. [
    2.  
    1. omsa
    2. house
    1. om
    2. build
    1. ]
    2.  
    1. tsi
    2. and
    1. [
    2.  
    1. ǁgam-e
    2. water
    1. go
    2. pst
    1. ā.
    2. drink
    1. ]
    2.  
    1. ‘Khoedage built the house and drank water.’
    1. (35)
    1. Khoedages
    2. K.
    1. ge
    2. decl
    1. [
    2.  
    1. omsa
    2. house
    1. go
    2. pst
    1. om
    2. build
    1. ]
    2.  
    1. tsi
    2. and
    1. [
    2.  
    1. ǁgam-e
    2. water
    1. ā.
    2. drink
    1. ]
    2.  
    1. ‘Khoedage built the house and drank water.’
    1. (36)
    1. Khoedages
    2. K.
    1. ge
    2. decl
    1. [
    2.  
    1. omsa
    2. house
    1. om
    2. build
    1. ]
    2.  
    1. tsi
    2. and
    1. [
    2.  
    1. ǁgam-e
    2. water
    1. ā
    2. drink
    1. tama.
    2. neg.nf
    1. ]
    2.  
    1. ‘Khoedage didn’t build the house and drink water.’

Relative:

    1. (37)
    1. [
    2.  
    1. Oms
    2. home
    1. ǀkha
    2. to
    1. go
    2. pst
    1. oa
    2. return
    1. ]
    2.  
    1. khoeb
    2. man
    1. ge.
    2. decl
    1. ‘It was the man who returned home.’
    1. (38)
    1. [
    2.  
    1. Oms
    2. home
    1. ǀkha
    2. to
    1. oa
    2. return
    1. tama
    2. neg.nf
    1. ]
    2.  
    1. khoeb
    2. man
    1. ge.
    2. decl
    1. ‘It was the man who didn’t return home.’
    1. (39)
    1. [
    2.  
    1. Mai-e
    2. pap
    1. go
    2. pst
    1. -ro
    2. -imp
    1. huni
    2. huni
    1. ]
    2.  
    1. ǀgôab
    2. boy
    1. ge.
    2. decl
    1. ‘It was the boy who stirred the pap.’
    1. (40)
    1. [
    2.  
    1. Mai-e
    2. pap
    1. huni
    2. huni
    1. hâ
    2. perf
    1. ]
    2.  
    1. ǀgôab
    2. boy
    1. ge.
    2. decl
    1. ‘It is the boy who has stirred the pap.’

Quotative:

    1. (41)
    1. Mî
    2. say
    1. ta
    2. I
    1. ge
    2. decl
    1. ra
    2. imp
    1. [
    2.  
    1. arib
    2. dog
    1. ge
    2. decl
    1. ǀhôasa
    2. cat
    1. go
    2. pst
    1. mû
    2. see
    1. ti.
    2. c.quot
    1. ]
    2.  
    1. ‘I am saying that the dog saw the cat.’
    1. (42)
    1. Mî
    2. say
    1. ta
    2. I
    1. ge
    2. decl
    1. ra
    2. imp
    1. [
    2.  
    1. arib
    2. dog
    1. ge
    2. decl
    1. ǀhôasa
    2. cat
    1. mû
    2. see
    1. tama
    2. neg.nf
    1. ti.
    2. c.quot
    1. ]
    2.  
    1. ‘I am saying that the dog didn’t see the cat.’
    1. (43)
    1. Mî
    2. say
    1. ta
    2. I
    1. ge
    2. decl
    1. ra
    2. imp
    1. [
    2.  
    1. ne
    2. this
    1. khoes
    2. woman
    1. ge
    2. decl
    1. ǁgan-e
    2. meat
    1. go
    2. pst
    1. am
    2. grill
    1. ti.
    2. c.quot
    1. ]
    2.  
    1. ‘I am saying that this woman grilled the meat.’
    1. (44)
    1. Mî
    2. say
    1. ta
    2. I
    1. ge
    2. decl
    1. ra
    2. imp
    1. [
    2.  
    1. ne
    2. this
    1. khoes
    2. woman
    1. ge
    2. decl
    1. ǁgan-e
    2. meat
    1. am
    2. grill
    1. hâ
    2. perf
    1. ti.
    2. c.quot
    1. ]
    2.  
    1. ‘I am saying that this woman has grilled the meat.’

Question:

    1. (45)
    1. ǁNa
    2. that
    1. tarasa
    2. woman
    1. go
    2. pst
    1. tae-e
    2. what
    1. am?
    2. grill
    1. ‘What did that woman grill?’
    1. (46)
    1. ǁNa
    2. that
    1. tarasa
    2. woman
    1. tae-e
    2. what
    1. am
    2. grill
    1. tama?
    2. neg.nf
    1. ‘What didn’t that woman grill?’
    1. (47)
    1. ǁNa
    2. that
    1. ǀgôaba
    2. boy
    1. go
    2. pst
    1. -ro
    2. -imp
    1. tae-e
    2. what
    1. ā?
    2. drink
    1. ‘What did that boy drink?’
    1. (48)
    1. ǁNa
    2. that
    1. ǀgôaba
    2. boy
    1. tae-e
    2. what
    1. ā
    2. drink
    1. hâ?
    2. perf
    1. ‘What has that boy drunk?’

Filler:

    1. (49)
    1. ǁAri
    2. yesterday
    1. =b
    2. =3ms
    1. ge
    2. decl
    1. ne
    2. this
    1. khoeba
    2. man
    1. ǁnaba
    2. there
    1. ǂna
    2. dance
    1. tama.
    2. neg.nf
    1. ‘This man didn’t dance there yesterday.’
    1. (50)
    1. ǁAri
    2. yesterday
    1. =b
    2. =3ms
    1. ge
    2. decl
    1. ǁnaba
    2. there
    1. ne
    2. this
    1. khoeba
    2. man
    1. ǂna
    2. dance
    1. tama.
    2. neg.nf
    1. ‘This man didn’t dance there yesterday.’
    1. (51)
    1. Nesi
    2. now
    1. =b
    2. =3ms
    1. ge
    2. decl
    1. ariba
    2. dog
    1. ǀhôasa
    2. cat
    1. nâ
    2. bite
    1. tide.
    2. neg.fut
    1. ‘Now the dog will not bite the cat.’
    1. (52)
    1. Nesi
    2. now
    1. =b
    2. =3ms
    1. ge
    2. decl
    1. ǀhôasa
    2. cat
    1. ariba
    2. dog
    1. nâ
    2. bite
    1. tide.
    2. neg.fut
    1. ‘Now the dog will not bite the cat.’
    1. (53)
    1. Netsē
    2. today
    1. =b
    2. =3ms
    1. ge
    2. decl
    1. khoeba
    2. man
    1. oms
    2. home
    1. ǀkha
    2. to
    1. go
    2. pst
    1. oa
    2. return
    1. ǀkhi.
    2. come
    1. ‘Today the man came back home.’
    1. (54)
    1. Netsē
    2. today
    1. =b
    2. 3ms
    1. ge
    2. decl
    1. oms
    2. home
    1. ǀkha
    2. to
    1. khoeba
    2. man
    1. go
    2. pst
    1. oa
    2. return
    1. ǀkhi.
    2. come
    1. ‘Today the man came back home.’
    1. (55)
    1. ǁNaba
    2. there
    1. =s
    2. =3fs
    1. ge
    2. decl
    1. tarasa
    2. woman
    1. ǃgâise
    2. well
    1. go
    2. pst
    1. -ro
    2. -imp
    1. ǁnae.
    2. sing
    1. ‘The woman was singing well there.’
    1. (56)
    1. ǁNabas
    2. there
    1. ge
    2. =3fs
    1. ǃgâise
    2. decl
    1. tarasa
    2. well
    1. go
    2. woman
    1. -ro
    2. pst
    1. ǁnae.
    2. -imp
    1.  
    2. sing
    1. ‘The woman was singing well there.’
    1. (57)
    1. Netsē
    2. today
    1. =b
    2. =3ms
    1. ge
    2. decl
    1. axaba
    2. boy
    1. ǃhaese
    2. quickly
    1. ǂû
    2. eat
    1. hâ.
    2. perf
    1. ‘Today the boy has eaten quickly.’
    1. (58)
    1. Netsē
    2. today
    1. =b
    2. =3ms
    1. ge
    2. decl
    1. ǃhaese
    2. quickly
    1. axaba
    2. boy
    1. ǂû
    2. eat
    1. hâ.
    2. perf
    1. ‘Today the boy has eaten quickly.’
    1. (59)
    1. Tsī
    2. and.then
    1. =b
    2. =3ms
    1. ge
    2. decl
    1. ǀgôaba
    2. boy
    1. ǀhūsa
    2. spider
    1. go
    2. pst
    1. mû.
    2. see
    1. ‘And then the boy saw the spider.’
    1. (60)
    1. Tsī
    2. and.then
    1. =b
    2. =3ms
    1. ge
    2. decl
    1. ǀhūsa
    2. spider
    1. ǀgôaba
    2. boy
    1. go
    2. pst
    1. mû.
    2. see
    1. ‘And then the boy saw the spider.’
    1. (61)
    1. Dandagob
    2. D.
    1. ge
    2. decl
    1. ǂkhanisa
    2. book
    1. ǁîb
    2. his
    1. ǀgôasa
    2. daughter
    1. khomai
    2. read
    1. -ba
    2. -appl
    1. hâ.
    2. perf
    1. ‘Dandago has read the book to his daughter.’
    1. (62)
    1. Dandagob
    2. D.
    1. ge
    2. decl
    1. ǁîb
    2. his
    1. ǀgôasa
    2. daughter
    1. ǂkhanisa
    2. book
    1. khomai
    2. read
    1. -ba
    2. -appl
    1. hâ.
    2. perf
    1. ‘Dandago has read the book to his daughter.’
    1. (63)
    1. Khoedages
    2. K.
    1. ge
    2. decl
    1. ǁgauǃna-aoba
    2. teacher
    1. ǁnaba
    2. there
    1. ra
    2. imp
    1. ǃhoa-u.
    2. talk.to
    1. ‘Khoedage is talking to the teacher over there.’
    1. (64)
    1. Khoedages
    2. K.
    1. ge
    2. decl
    1. ǁnaba
    2. there
    1. ǁgauǃna-aoba
    2. teacher
    1. ra
    2. imp
    1. ǃhoa-u.
    2. talk.to
    1. ‘Khoedage is talking to the teacher over there.’
    1. (65)
    1. Tita
    2. I
    1. ge
    2. decl
    1. ǂkhanisa
    2. book
    1. ǁkhawa
    2. again
    1. ra
    2. imp
    1. xoa.
    2. write
    1. ‘I am writing a book again.’
    1. (66)
    1. Tita
    2. I
    1. ge
    2. decl
    1. ǁkhawa
    2. again
    1. ǂkhanisa
    2. book
    1. ra
    2. imp
    1. xoa.
    2. write
    1. ‘I am writing a book again.’
    1. (67)
    1. ǀHôas
    2. cat
    1. ge
    2. decl
    1. ariba
    2. dog
    1. netsē
    2. today
    1. mû
    2. see
    1. tama.
    2. neg.nf
    1. ‘The cat didn’t see the dog today.’
    1. (68)
    1. ǀHôas
    2. cat
    1. ge
    2. decl
    1. netsē
    2. today
    1. ariba
    2. dog
    1. mû
    2. see
    1. tama.
    2. neg,nf
    1. ‘The cat didn’t see the dog today.’
    1. (69)
    1. ǁNa
    2. that
    1. ǀgôa-i
    2. child
    1. ge
    2. decl
    1. khoe-e
    2. someone
    1. ǂanebega-se
    2. on.purpose
    1. nâ
    2. bite
    1. tama.
    2. neg.nf
    1. ‘That child bit someone on purpose.’
    1. (70)
    1. ǁNa
    2. that
    1. ǀgôa-i
    2. child
    1. ge
    2. decl
    1. ǂanebega-se
    2. on.purpose
    1. khoe-e
    2. someone
    1. nâ
    2. bite
    1. tama.
    2. neg.nf
    1. ‘That child bit someone on purpose.’
    1. (71)
    1. ǁGauǃna-aos
    2. teacher
    1. ge
    2. decl
    1. ne
    2. this
    1. axaba
    2. boy
    1. netsē
    2. today
    1. ǃhoa-u
    2. talk.to
    1. tide.
    2. neg.fut
    1. ‘The teacher didn’t talk to this boy today.’
    1. (72)
    1. ǁGauǃna-aos
    2. teacher
    1. ge
    2. decl
    1. netsē
    2. today
    1. ne
    2. this
    1. axaba
    2. boy
    1. ǃhoa-u
    2. talk.to
    1. tide.
    2. neg.fut
    1. ‘The teacher didn’t talk to this boy today.’

Abbreviations

3fs = 3rd person feminine singular, 3ms = 3rd person masculine singular, appl = applicative, comp = complementizer, c.quot = quotative complementizer, decl = declarative clause type marker, fut = future, imp = imperfect, neg = negative, nf = nonfuture, perf = perfect, pst = past tense.

Notes

  1. I will follow the tonal notation convention used for Khoekhoegowab by Brugman (2009), in which the diacritics /a̋ á à ȁ/ correspond to superhigh ([5]), high ([4]), low ([2]), and superlow ([1]), respectively. A vowel with no tone marked indicates that no tone target is associated with it; this results in F0 interpolation between the last tone target and the next. Other than the addition of tone marking where relevant, all examples are presented in Khoekhoegowab standard orthography. [^]
  2. Functional items like auxiliary verbs or nominal affixes also have contrastive tone, but that tone system works differently from the tone on lexical vocabulary; see Brugman (2009) for details. [^]
  3. All observations about the distribution of sandhi in DPs are due to Brugman (2009) and confirmed by my own fieldwork. [^]
  4. Brugman does note an exception to this generalization: Example (73) on page 260 shows that it is possible “to reverse the order of the verb and the [normally preverbal] tense marker, and in such cases, the verb takes its citation melody”. If true, this is an extremely interesting and suggestive fact; however, in my own fieldwork, my language consultants universally rejected the sentences in Brugman’s example (73) (and others of the same form) as ungrammatical. As such, I’ll ignore this wrinkle in Brugman’s data going forward. [^]
  5. The facts leading to this conclusion were originally noted by Hahn (2013). [^]
  6. A reviewer points out that there is a further wrinkle in the data not discussed here: Brugman’s data shows that root clause verbs fail to undergo sandhi when followed by the past-tense copula i; interestingly, this is true even if the verb is also preceded by a preverbal tense marker, making this phenomenon difficult to account for even in Haacke’s model. Unfortunately, my consultants were reluctant to produce sentences bearing both a preverbal tense marker and the past-tense copula; they preferred to place even normally-preverbal markers after the verb but before the copula. As a result, I cannot say for certain whether my consultants would replicate the exceptional sandhi behavior of i; this complication would be an interesting matter for future investigation. [^]
  7. /huni/ ‘stir’ is often produced as [uni]. [^]
  8. Brugman (2009) analyses these OVS sentences as a kind of TP fronting. My analysis of them as relative clause clefts hinges on three facts. First, the subject obligatorily undergoes sandhi in this context, as though the noun is not leftmost within its own phrase; this implies that the preceding material (i.e. the embedded clause) is contained within the DP. Second, OVS word order is ungrammatical when the subject is immediately preceded by a demonstrative. This is unexpected if OVS is derived by TP fronting, but expected if the OV constituent is a subject relative clause within the DP — determiners precede DP-internal relatives. Finally, this word order has a unique pragmatic meaning: it is used to convey that the subject is new information while the rest of the clause is given, parallel to cleft structures in other languages. [^]
  9. In examples (7) – (11), the first subexample has a tense marker in preverbal position while the second has a tense marker in postverbal position. In example (12), [a] has the tense marker preceding both verbs; [b] has it preceding only the second; and [c] has it fully postverbally. [^]
  10. χ2(1,255) = 20.33, p < 0.0001. [^]
  11. Speakers frequently partially devoiced the vowel of the verb when it was clause-final; the 5 rejected items all had a fully devoiced vowel. [^]
  12. Fleiss’ Kappa is a measure of inter-transcriber agreement; see Fleiss (1971). It generalizes the widely-used Cohen’s Kappa to datasets with more than 2 transcribers. [^]
  13. A reviewer expressed concern that this generalization overstates the results, given that sandhi was transcribed in a non-zero percentage of tokens occurring where Hypothesis C predicts citation form. I agree with the reviewer that, in principle, the results here are open to one of two interpretations: Either 1) sandhi is categorical and obeys Hypothesis C, but some citation-form verbs were transcribed as sandhi due to unrelated factors (e.g. declination); or 2) sandhi is variable (see Section 5.1) and Hypothesis C should be restated as a statistical generalization. In the discussion that follows, I’ll mostly continue to assume the former interpretation; however, Section 5.1 discusses some possible evidence for at least somewhat adopting the latter. It’s worth noting however that under neither of these interpretations are the results of the study strictly compatible with Haacke’s or Brugman’s generalizations. [^]
  14. For example, as the result of a constraint requiring Extended Projections to be prosodically phrased together (as proposed by López 2009), or as the result of a constraint requiring argument-selection relations to be maintained in prosody (as proposed by Clemens 2019). [^]
  15. Brugman (2009) does acknowledge variability in one limited respect: nouns preceded by a relative clause sometimes anomalously resist sandhi. Verbal sandhi, though, is said to be categorical. [^]

Acknowledgements

I would like to thank Professor Levi Namaseb for his help connecting with Khoekhoegowab language consultants; many thanks as well to Irene ǁGaroes, Nadia April, Gerdrut Hevita, Nicoline Geingos, and many other Khoekhoegowab speakers who welcomed me into their communities. Kerri Devlin spent countless hours transcribing recordings; thanks as well to Ivy Hauser, Andrew Lamont, and Maggie Baird. Finally, I’m grateful to Krstine Yu for considerable support and advice.

Funding Information

This research was supported by the National Science Foundation Graduate Research Fellowship under Grant No. 1451512. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the author and do not necessarily reflect the views of the National Science Foundation.

Competing Interests

The author declares that they have no competing interests.

References

Boersma, Paul & Weenink, David. 2001. Praat, a system for doing phonetics by computer.

Brugman, Johanna Christina. 2009. Segments, tones and distribution in Khoekhoe prosody. Cornell University dissertation.

Bruin, Joe. 2011. R library contrast coding systems for categorical variables. https://stats.idre.ucla.edu/r/library/r-library-contrast-coding-systemsfor-categorical-variables/.

Clemens, Lauren Eby. 2019. Prosodic noun incorporation: the relationship between prosody and argument structure in Niuean. Syntax. DOI:  http://doi.org/10.1111/synt.12170

Connell, Bruce. 2001. Downdrift, downstep, and declination. In Proceedings of Typology of African Prosodic Systems Workshop. Bielefeld University, Germany.

Fleiss, Joseph L. 1971. Measuring nominal scale agreement among many raters. Psychological Bulletin 76. 378–81. DOI:  http://doi.org/10.1037/h0031619

Fougeron, Cécile & Jun, Sun-Ah. 1998. Rate effects on French intonation: prosodic organization and phonetic realization. Journal of Phonetics 26(1). 45–69. DOI:  http://doi.org/10.1006/jpho.1997.0062

Haacke, Wilfrid H. G. 1999. The tonology of Khoekhoe (Nama/Damara). Quellen zur Khoisan-Forschung 16. Köln: Rüdiger Köppe Verlag.

Hagman, Roy. 1977. Nama Hottentot Grammar. Indiana University.

Hahn, Michael. 2013. Word order variation in Khoekhoe. In Proceedings of the 20th International Conference on Head-Driven Phrase Structure Grammar, 48–68.

Kusmer, Leland P. 2019. Optimal Linearization: prosodic displacement in Khoekhoegowab and beyond. University of Massachusetts Amherst dissertation.

Lewis, M. Paul & Simons, Gary F. & Fennig, Charles D. (eds.). 2016. Ethnologue: languages of the world. 19th edn. Online version: http://www.ethnologue.com. Dallas, TX: SIL International.

López, Luis. 2009. Ranking the linear correspondence axiom. Linguistic Inquiry 40.2. 239–276. DOI:  http://doi.org/10.1162/ling.2009.40.2.239

McAuliffe, Michael & Socolof, Michaela & Mihuc, Sarah & Wagner, Michael & Sonderegger, Morgan. 2017. Montreal Forced Aligner: trainable textspeech alignment using Kaldi. In Proceedings of the 18th Conference of the International Speech Communication Association. DOI:  http://doi.org/10.21437/Interspeech.2017-1386

Yip, Moira. 2002. Tone. New York: Cambridge University Press.

Yue-Hashimoto, Anne O. 1987. Tone sandhi across Chinese dialects. In Chinese Language Society of Hong Kong (ed.), Wang Li memorial volumes, English volume, 445–474. Hong Kong: Joint Publishing Co.

Zee, Eric & Maddieson, Ian. 1980. Tones and tone sandhi in Sanghai: phonetic evidence and phonological analysis. Glossa 14. 44–88.

Zhang, Jie. 2007. A directional asymmetry in Chinese tone sandhi systems. Journal of East Asian Linguistics 16(4). 259–302. DOI:  http://doi.org/10.1007/s10831-007-9016-2

Zhu, Xianong. 1999. Shanghai Tonetics. München: Lincom Europa.

Zhu, Xianong. 2006. A Grammar of Shanghai Wu. München: Lincom Europa.