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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.)

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.

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.

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.

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.