Direct speech, silent pauses, speech verbs, and basic word order: a comparative corpus study of 12 languages

3.1 Data: corpora on 12 diverse languages

The 12 corpora used in the current study were extracted from the set of 51 time-aligned corpora included in the DoReCo collection (version 1.0, Seifart et al. 2022). Each of these contains on average around 10,000 words of mostly narrative speech that has been collected through on-site fieldwork on often endangered languages, and transcribed, translated, and annotated by experts on these languages. For 38 languages, this annotation also includes morpheme segmentation and interlinear morpheme glosses. From this set we selected 12 languages for the current study (Table 1). This selection was based on the following criteria: (i) equal representation of VO and OV languages as well as inclusion of some languages with no dominant word order (also called non canonical word order), and (ii) world-wide genealogical and areal diversity. While this language sample is still limited, it is three times as large as that of Malibert and Vanhove (2015) and represents the most comprehensive dataset available for this type of labor-intensive manual annotation of direct speech at present. Information in Table 1 includes the Glottocode language identification code, which points to entries in the Glottolog (Hammarström et al. 2022), from where information on geographic area and on top-level genealogical language family was taken. Information on basic word order (BWO) was taken from Dryer (2013), where available, or from published descriptions of the language. We additionally provide information on the “genus” a language belongs to, i.e. on genealogical units of comparable time depths (roughly 4,000 years) that are commonly used in sampling for typological studies (Dryer 1989). This information was taken from the World Atlas of Language Structures (WALS) (Dryer and Haspelmath 2013).

Four languages from the Americas are represented in our sample, three South American (Bora, Movima, and Mojeño Trinitario) and one North American language (Arapaho). The latter is a language of the Algonquian branch of the Algic family, spoken by around 250 speakers according to Cowell and Moss (2008). The language is highly polysynthetic and does not have a dominant word order regarding the placement of V and O. The Bora language is spoken by around 3,000 speakers in the Peruvian and Colombian Amazon and belongs to the small Boran family. Bora is a tonal language with OV word order, where SOV and OSV are equally common. Since we are interested in the order of objects, not subjects, relative to verbs, we group it with other (S)OV languages here. The language isolate Movima is spoken in Bolivia by about 1,400 speakers. Word order in Movima follows complex syntactic rules, with no dominant order for V and O (Haude 2006). Finally, we use a corpus of the Arawakan language Mojeño Trinitario, also spoken in Bolivia. This language has been described as having SVO word order (Rose 2014: 89). For reasons of space, we refer to this language as “Mojeño” in the remainder of the current study.

Among the three African languages in our data set, Beja is an Afroasiatic language from the Cushitic branch spoken by one to two million speakers in Egypt, Sudan, and Eritrea and it has SOV word order (Vanhove 2017). The corpus used for the current study is a subset of the corpus used in Malibert and Vanhove’s (2015) study. Nǁng, from the Tuu family (formerly classified as “Khoisan”), is a critically endangered language spoken in Namibia. It has been described as an SVO language. Ruuli is a Bantu language spoken in Uganda by around 160,000 speakers and it is also an SVO language (Sørensen and Witzlack-Makarevich 2020).

There are three corpora from Oceania (called “Papunesia” in Glottolog) in our study, two Austronesian languages, Fanbyak and Nisvai, and the isolate Savosavo. Both Fanbyak and Nisvai are spoken in Vanuatu, the first one on Ambrym, and the second one in South-East Malekula, by around 200 native speakers each, and they both belong to the Central Vanuatu branch of Oceanic. They are both SVO languages. The Nisvai corpus used for this study is part of a larger corpus that was created for a study of narrative practices within the Nisvai community (Aznar 2019). Savosavo is an isolate language spoken by about 3,000 speakers on Savo Island in the Solomon Islands, and – unlike Nisvai and Fanbyak – it is an SOV language (Wegener 2012).

Two languages from Eurasia are represented in our sample: Dolgan is a Turkic language spoken on the Taymyr Peninsula in Siberia by around 1,000 speakers. Like other Turkic languages, its word order is SOV (Stapert 2013: 246). Sanzhi Dargwha is a Nakh-Daghestanian language spoken in the Northern Caucasus by around 250 speakers, and its basic word order is also SOV (Forker 2019). We refer to this language as “Sanzhi” here.

We selected between one and nine texts from each of these 12 DoReCo corpora that we further annotated for the purpose of the current study, as summarized in Table 2. This selection strikes a balance between keeping the amount of manual annotation feasible, on the one hand, and representing speakers of different sexes and age groups where possible. Data processing was facilitated by consistency across DoReCo data regarding the naming and structure of annotation tiers, such as transcription, word-level segmentation, interlinear glosses, as well as the time-alignment of these annotations.

Despite our efforts to produce a balanced data set, Table 2 shows various biases and imbalances regarding the number of texts, their lengths, and the characteristics of the speakers. For instance, Bora and Savosavo represent almost 40 % of the total corpus in terms of interpausal units or recording duration. Regarding the sex of the speakers, there is also a discrepancy: 25 texts were produced by men and only 11 by women, and this disparity is even more important when we consider the time ratio between male and female speakers, which is almost four times more for male speakers. Another bias is that older speakers are overrepresented in our data: the average age of speakers in our data set is 53.1 years. These biases can be attributed to the fact that most DoReCo data stems from language documentation projects that focus on the documentation of traditional narrative practices, which are more often produced by elders of the communities, and often by men.

DoReCo corpora provide audio files and corresponding annotation files in ELAN format (ELAN Developers 2021), among others. We used these ELAN annotation files as input for building a multilingual corpus with comparable units out of the time-aligned corpora. DoReCo data also already included annotation of silent pauses, as a result of automatic identification of silent pauses during automatic phone-level time alignment with subsequent manual corrections at the level of word start and end times, which paid particular attention to the identification of silent pauses (Paschen et al. 2020). These pauses define interpausal units (IPUs) in our data (see Table 2). To further process these data, a Python module was developed following the rationale laid out in Aznar (2020), in order to compile a single, cross-linguistic corpus according to our study’s requirements, working with interpausal units. Furthermore, a Jupyter notebook (Kluyver et al. 2016) was written and combined with Poetry^[2] to document and ease the reuse of our data processing methods. These tools also aided in producing the results presented in this article, as well as the visualizations that helped to interpret the results.^[3]

3.2 Annotation conventions, data coding, and analyses

We developed a dedicated annotation scheme (Table 3) for testing our hypotheses and applied it across corpora. The annotation tags delimitate and describe direct speech reports and three positions of speech verbs: introductory, middle, and ending verbs. Tags for speech verbs in fact extend over entire verb phrases, potentially including its arguments, TAM marking, etc. Our annotation scheme also includes tags for interjections. Even though their analysis is outside the scope of the current study, we report on their occurrence, along with that of pauses, in Section 4.1. Following Norrick’s proposal (2014) to consider phrases such as Oh my god as interjectional, we annotated not only interjectional morphemes as interjections but, whenever relevant, the whole phrase associated with it. We also annotated appellatives as interjections. These are often vocative kinship terms, as in “Grandfather,” he said, “come here”. Recall that silent pauses were already annotated in the DoReCo data we used as input.

Our annotation of direct speech is framed within the concept of direct speech sequence, which is a larger unit that contains all relevant elements surrounding speech reports, including potentially more than one speech verb. The concept of sequence is inspired by the work of Adam (2011) on text sequences as discourse units associated with linguistic cues. A direct speech sequence is characterized by containing one speech report, unless a middle verb occurs, in which case it contains two speech reports. It can, and often does, extend over multiple interpausal units. In our definition of a direct speech sequence, a sequence cannot contain the same type of speech verb twice: a second introductory verb initiates a new direct speech sequence. On the other hand, a direct speech sequence containing an introductory verb can be followed by another direct speech sequence containing only a middle verb. While according to the formal criteria, both could be within the same sequence, we relied on the content within the narrative for indications that the two speech reports are different events, e.g. when they report speech by two different speakers in dialogical speech, and then we annotate these as two separate sequences.

Each of the corpora in the current study had already been manually transcribed, translated, and morphologically annotated by the teams of authors that created them (see Table 1). We manually added annotations for the current study on direct speech, as given in Table 3, to a copy of the transcription tier. This annotation work was done by the first author for 11 of the corpora used here, and by the second author for the Bora corpus. Our identification of direct speech directly followed the language experts’ use of quotation marks in free translations (and in some cases also in the orthographic transcriptions). This concerns the start- and endpoints of speech reports as well as potential borderline cases such as Life is short, which could be hard to classify as either direct or indirect speech (although such cases only very rarely occurred in our data). We also found that in the narrative texts we used (perhaps unlike in conversation), the context almost always makes it clear if speech is attributed to a character of the narrative or not. Where in doubt, we consulted the language experts that created the corpora. We also made sure to only include instances of speech reports that fully conform to the definition of direct speech used in the current study (see Section 2). Manually annotating data ensured that the authors gained detailed insights into how direct speech is expressed in each of the languages, which facilitated a reliable interpretation of the results of the quantitative analyses.

Regarding the annotation of speech verbs, we only included speech verbs that are actually syntactically linked to a speech report. We thus excluded cases like mwēwar ‘he spoke (to them)’ in example (2) from Dolgan, which describes the speaking event in general. We did include the verb mwigile ‘he said’ (in bold), which actually introduces a speech report.

Example (2) also illustrates the first type of speech verbs, namely introductory verbs. Example (3), from Beja, illustrates an ending verb. In example (4), from Arapaho, the verb nihʔiit ‘PST-said-3.S’ is a middle verb in between two speech reports. Example (5), from Bora, illustrates an introductory, a middle, and a final verb within one direct speech sequence.

For the definition of pauses in the context of direct speech production, we only included silent pauses, and disregarded filled pauses. The reason for this is that filled pauses appear to be alternatives for silent pauses only in their function as hesitation markers, rather than as markers of prosodic boundaries, which is what we are interested in here. Recall that DoReCo’s time-alignment procedure involved manual corrections specifically of start and end times of silent pauses. Therefore we did not exclude any silent pauses, e.g. pauses shorter than any given minimal length.

Regarding the analyses of the data, we combine qualitative analyses with descriptive statistics such as proportions of different verb types in languages with different word orders. To estimate the effects of the different factors on the probability of silent pauses before and after speech reports, we fitted two Bayesian Generalized Linear Mixed Models with logistic (Bernouilli) response, using PyMC 5 (Abril-Pla et al. 2023), one predicting the presence or absence of a pause before the speech report, and another predicting the presence or absence of a pause after. As fixed effects, we included the presence or absence of each of the three types of speech verbs and the basic word order, as well as speaker sex and speaker age, standardized as a z-score, to control for variation according to these two factors (e.g., Jacewicz et al. 2009); for the pause-before model only, we also incorporated the distance between the onset of the speech report and the preceding last pause. Because of the nature of our annotation, we were not able to measure the distance to a following pause, to control for duration (or length) of the following phrase. Languages and speakers were treated as random intercepts to capture variability across both groups. The models include interactions between word order and the respective speech verb (introductory or ending) to determine whether languages with different word orders exhibit distinct pause practices. For the interpretation of results, the coefficients are estimated on a logit scale; each coefficient represents a change in the log odds of a pause occurring per unit change in the predictor. Therefore all the predictor effects are to be interpreted relative to the intercept level. Finally, we also followed the practices of Engelmann et al. (2019), Kumarage et al. (2022), and Donnelly et al. (2024) to distinguish between weak evidence (for results within the 85 % Highest Density Interval [HDI]) and strong evidence (results within the 95 % HDI). Results under the threshold of 85 % were treated as providing no evidence.

4.1 Variability in the structure of speech reports

Before presenting results that directly address our hypotheses (Sections 4.2–4.3) and additional results regarding middle verbs (Section 4.4), we provide in this section some general observations regarding the cross-linguistic variability in speech verb usage and the structure of direct speech sequences. These observations are summarized in Table 4.

In Table 4, observe firstly that the number of speech reports exceeds the number of direct speech sequences in our data in many languages (although not in most SVO languages). This is due to the use of “middle verbs” in these languages, which break up direct speech sequences into various speech reports. Secondly, observe that direct speech may occur without any overt speech verb in all languages of our sample, consistent with Jordanoska et al.’s (2022) findings. The absence of a speech verb is in fact more common than its presence in four out of the 12 languages studied here, including both languages with no dominant order and Mojeño, which behaves like a language with no dominant word order in many respects (see Sections 4.2–4.4). Thirdly, we can also see that the use of interjections in speech reports is recurrent across the languages studied here, again to varying degrees, in line with Güldemann’s (2008: 40) remark that “quote-initial expressive items like interjections, exclamations, vocatives, etc. are quite effective devices that the reporter can use for DRD-marking [Direct Reported Discourse]”.

Another important observation is that the phrases headed by introductory verbs tend to be more complex compared to those headed by ending verbs. In example (6), from Savosavo, direct speech is distributed over three interpausal units and is associated with two speech verbs: the introductory verb savu-li-ghu(e) ‘say/tell-3sg.m.o-nmlz’ and the ending verb tei ‘want.to.do/say/be.like.this’ (another instance of that verb, tei-ghue is used with the meaning ‘want.to.do’ within the speech report in this example). The phrase headed by the introductory verb savu-li-ghu does not only introduce the speech report here, but also specifies the participants involved in the situation, namely the third person speaker, through tulola-lo: (then-3sg.m.nom), and the recipient, mama ‘mother’. On the other hand, the phrase headed by the ending verb tei is much less complex, and shorter.

Example (7) from Arapaho also illustrates this difference. The phrase headed by the first instance the verb ‘iit ‘to say’ explicitly mentions the character who produced the speech report howoh’oe ‘wait!’. This is particularly important in Arapaho narratives, where there are many instances where the characters are not overtly expressed.

Example (8) from Bora contains another example of an introductory verb that heads a complex phrase (see example (5) for yet another example).

We thus observe in all three cases that the introductory verb is used to (re-)specify certain aspects of the narrative situation, such as who the main characters are. Some languages, like Savosavo, in fact use different verb stems as introductory versus ending verbs (example 6).

4.2 Speech verb position and basic word order

In this section, we address our first hypothesis, namely that speech reports pattern with verb complements in terms of word order, i.e. they tend to follow speech verbs in VO language and mostly precede, but sometimes follow them in OV languages. Figure 1 summarizes the proportions of verbs occurring in different positions relative to the speech report found in our data, based on the counts of direct speech sequences and speech reports given in Table 4. The upper part of Figure 1 visualizes the proportions of introductory and final speech verbs (blue vs. green bars), as well as middle verbs (orange bars) in each language. The lower part reports on the number of speech verbs observed and the percentage it represents per language and among the sequences with at least one speech verb.

Figure 1:

Occurrence of speech verbs in different positions with respect to speech reports. See Table 4 for the total numbers of direct speech sequences.

Figure 1 shows that there is a clear relationship between the basic word order of a language and the position of speech verbs: In all five OV languages in our sample, speech verbs more often follow than precede speech reports, consistent with the basic order of verbs and their complements, but may follow them, consistent with the “heavy last” principle. In four out of five of the VO languages in our sample, speech verbs exclusively precede speech reports, i.e. these four languages only use introductory speech verbs, situated before the speech report (blue bars). Overall, results from nine out of ten languages that do have a basic word order are thus consistent with our hypothesis. Mojeño, however, is a stark exception in that there are almost no introductory speech verbs, and the majority are ending verbs. We confirmed the basic order of V and O in that language by analyzing their order in the first 71 instances of transitive non-speech verbs with nominal objects in the data we used here, and found that 68 of them had VO order.

We thus observe that the position of speech verbs in VO languages is practically fixed in the initial position, with the exception of Mojeño, while in other languages, it is much more variable: While all OV languages use more ending verbs (green bars) than introductory verbs (blue bars), they all also use introductory verbs and middle verbs (orange bars), separately or together, as we will discuss in Section 4.4. Speakers of Bora, for instance, use as many ending verbs as middle verbs, in addition to rare cases of introductory verbs. This flexibility is consistent with the fact that the order of objects relative to the verb is relatively flexible in Bora. In our Beja data, on the other hand, the great majority of speech verbs are ending verbs, consistent with Malibert and Vanhove’s (2015) findings, and also consistent with the fairly strict OV order they report for Beja.

Finally, we observe that the two languages with no dominant order of verbs and objects, Arapaho and Movima, pattern with OV languages, rather than VO languages, both in terms of the preference for ending verbs, and in terms of using middle verbs. Regarding middle verbs, we observe that these only occur in languages which also have ending verbs,^[4] and these two verb types are a feature exclusively of OV languages and languages with no dominant order, again with the exception of Mojeño. Finally, as can be seen from the percentages reported at the bottom of Figure 1, and confirming our observations from Table 4, above, languages with no dominant order, and – again – Mojeño, use overall considerably fewer speech verbs than other languages.

4.3 Pauses at the edges of speech reports: complementarity and basic word order

We hypothesized that silent pauses occur at those edges of speech reports where the speech verb does not occur, as a cue for identifying transitions to or from direct speech within a narrative. In this section, we address this by examining (i) the complementarity of pauses and speech verbs, i.e. how pause occurrences relate to occurrences of speech verbs in a given context (Hypothesis 2.1), and (ii) how the occurrence of a silent pause before and after speech reports relates to the basic word order of a language (Hypothesis 2.2). We first present our main results addressing these two hypotheses regarding the presence versus absence of pauses, and then complementary results regarding pause durations.

Figures 2 and 3 present the results from our Bayesian generalized linear mixed models for estimating the effects of speech verb presence, as well as those of other factors, on the probability of a silent pause before and after speech reports.

Figure 2:

Posterior Estimates for the Pause Before Models. The factor “Intro Verb” captures the effect that the presence of a speech verb has on the occurrence of a pause in the same position. Other factors are explained in the main text.

Figure 3:

Posterior Estimates for the Pause After Models. The factor “Ending Verb” captures the effect that the presence of a speech verb has on the occurrence of a pause in the same position. Other factors are explained in the main text.

For pauses before speech reports, Figure 2 shows an intercept mean of 1.951, which corresponds to a probability of 88 %, for the presence of a pause before speech reports under the model’s baseline scenario, which is defined as OV word order, absence of any speech verbs, average speaker age, and baseline sex (defined here as male). When an introductory verb is present, compared to no speech verb in the same baseline scenario, the probability of a pause before speech reports is strongly reduced: the result for “Intro Verb” has a mean of -2.040, which corresponds to a probability of 48 % for a pause to occur before speech reports. Further results reported in Figure 2 put this result into a wider perspective. Firstly, the presence of a speech verb in another position, middle or ending, does not have a credible effect on the production of a pause before the speech report. Regarding the effects associated with speaker characteristics, we observed a credible, positive age effect (mean ≈ 0.506), indicating that older speakers are more prone to pause at the onset boundary. Recall from Table 2 that speakers in the corpus are relatively old, 53.1 years on average, and the youngest speaker 25 years old. The effect of female sex on pausing is less clear: The mean of the effect is slightly negative (mean ≈ -0.323), but its Highest Density Interval (HDI) overlaps with zero, reflecting uncertainty or weaker stability of this effect. In Figure 2, we also report on the effect of the distance between the speech report and the preceding pause (“Distance to Last Pause” in Figure 2). This distance has been calculated by counting the number of transcribed characters between the beginning of the speech report and the preceding pause – regardless of whether there is a pause directly preceding the speech report, which is what we are predicting. Results indicate that the longer this distance, the more likely the presence of a pause before a speech report. This shows that, as expected, there is an increasing need for pausing the longer the distance since the last pause, for reasons of prosodic phrasing, and maybe also the physiological need for breathing. As we investigate this effect in a single model together with the effect of speech verb presence, we are able to show here that the presence of a speech verb independently and additionally has an effect on pause probabilities.

Regarding pauses after speech reports, Figure 3 shows a positive effect of 2.47 on the logit scale for the intercept, which means that in the baseline scenario (OV language, no speech verb, average-age male speaker), there is a probability of 92 % for a pause to occur after a speech report. Compared to the baseline scenario, the presence of an ending verb has a strong negative effect (mean ≈ -4.28), meaning that the probability of a pause is reduced to just 14 % when an ending verb is present. The effects related to the speaker’s characteristics, age and sex, are not conclusive in this context. Taken together, these results indicate that the presence of a speech verb and the presence of a pause are inversely related, both at the beginning and even more strongly at the end of speech reports.

Figures 2 and 3 also include the results that address the effect of basic word order on pause probabilities (Hypothesis 2.2). Regarding pauses before speech reports (Figure 2), we find strong evidence that having no dominant word order has a positive effect (mean ≈ 1.87), indicating that in Arapaho and Mojeno there is a higher probability of pauses before speech reports than in OV and VO languages. On the other hand, there is no evidence for an effect of VO word order on the probability of pauses before speech reports, nor are there credible effects of interactions between the presence of an introductory verb and any word order. For pauses after speech reports (Figure 3), we found no evidence for more pausing in languages with no dominant word order, and also not for VO languages. However, we did find evidence for interactions between word order and ending verbs: For languages with no dominant word order, there is strong evidence for a positive effect (mean ≈ 1.56) and for VO languages there is weak evidence for a likewise positive effect (mean ≈ 1.29). This indicates that, depending on the basic word order of a language, the impact of the presence of an ending verb on the probability of a silent pause after the speech report varies: VO languages and languages with no dominant word order tend to have more pauses in this context, compared to OV languages. However, the result regarding VO languages should be interpreted with caution since Mojeño is the only VO language that uses any ending verbs in our corpus. Overall, our results thus indicate, firstly, that languages with no dominant word order clearly tend to pause more both before and after speech reports. Secondly, there is only weak evidence, and only for VO languages, for conventionalization of pausing in contexts where speech verbs do not occur according to the basic word order, i.e. before in OV languages and after in OV languages, independently of complementarity.

Next, we present two more results that further contextualize our main results. Firstly, Figure 4 provides language-specific results of posterior predictions from the Bayesian models, comparing pause probabilities before, in blue, and after, in red, depending on the presence versus absence of an introductory verb or an ending verb. These results indicate that complementarity between pauses and speech verbs prevails across individual languages and contexts (before vs. after). The results appear to also indicate potential word order effects. In OV languages there appears to be a relatively high probability of a pause preceding speech reports, i.e. in the position where typically there is no speech verb, even if a speech verb happens to be present (dark blue bars for OV languages). In VO languages, on the other hand, there appears to be typically a relatively high probability of a pause following speech reports, i.e. in the position where typically there is no speech verb, even if a speech verb happens to be present (dark red bars for VO languages). Figure 4 also illustrates variation across languages and word order types, emphasizing that while the overall patterns support complementarity and the influence of basic word order, individual languages differ in the magnitude of these effects. Note, however, the uncertainty of some of the individual results, reflected in excessively wide error bars that represent the 95 % HDI credible intervals associated with the model, e.g. for pause probabilities in the presence of introductory verbs in Ruuli and Nǁng. This is partially due to the fact that the models infer the probabilities of a pauses for every language in every position, even if they do not have ending verbs in the corpus, as is the case for Ruuli and Nǁng, for which there are no ending verbs attested in our corpora.

Figure 4:

Pause probabilities in the presence versus absence of speech verbs before and after speech reports in 12 individual languages.

Finally, we report descriptive results on pause durations in Figure 5. These show that before speech reports, the durations of pauses tend to be longer when there is no introductory verb (mean ≈ 0.97s) compared to when there is an introductory verb (mean ≈ 0.81s). In positions after speech reports, when an ending verb is absent, the mean pause duration is also relatively long (mean ≈ 0.78s), but when an ending verb is present, pauses are much shorter (mean ≈ 0.38s). A Bayesian model assessing the standardized pause duration also indicates that the presence of an introductory verb has a credible negative effect on introductory pause duration (mean = -0.220, HDI close to zero), albeit marginal. For ending verbs, the Bayesian analysis is clearer: a negative slope of about -0.569 indicates that ending verbs are associated with shorter pauses at the offset. Taken together, these results provide additional evidence for the complementarity hypothesis, and they also again show that pauses before and after speech reports differ. On the one hand, pauses before speech reports tend to be longer, regardless of the presence or absence of a speech verb, compared to pauses after speech reports. On the other hand, in the position between speech reports and ending verbs it is not only relatively rare for a pause to occur, but if pauses do occur there, they are also remarkably short.

Figure 5:

Duration of silent pauses in the absence versus presence of introductory and ending verbs.

4.4 Pauses before and after middle verbs

As noted in Section 4.1, OV languages and languages with no dominant order employ – in some cases many – middle verbs, in addition to ending and introductory verbs. Figures 2 and 3 suggest that the presence of middle verbs does not have an effect on pauses before or after the speech report as whole, as expected. In the current section, we briefly report on pauses before and after middle verbs. Since there are only relatively few middle verbs in most languages, except in Bora and Dolgan, we do not extend the modeling approach to the study of pauses before and after them, but we simply provide basic descriptive statistics (Table 5). These very preliminary results indicate that in most languages, middle verbs are often followed by pauses, i.e. a pause occurs between the middle verb and the following speech report. Only rarely (except in Arapaho) are middle verbs preceded by pauses, i.e. pauses only rarely intervene between the end of a speech report and a middle verb.

These preliminary results also indicate that middle verbs behave like ending verbs in terms of pausing, in that they follow preceding speech reports with little pausing, as can be seen from comparison of the column “pause before verb” in Table 5 with the estimated probability of, on average, 14 % for pauses before ending verbs (Figure 3). They also behave like introductory verbs in that pauses are more likely to intervene between the middle verb and a following speech report, as can be seen from comparison of the column “pause before verb” in Table 5 with the estimated probability of, on average, 48 % for pauses after introductory verbs (Figure 2). The two languages with no dominant word order appear to be characterized by a lot of pausing in the context of middle verbs, too, although these results are particularly uncertain because both languages use particularly few middle verbs. Additionally, we observed that in terms of their structure, as discussed in Section 4.1, middle verbs also tend to resemble ending verbs in that they tend to head short, one-word phrases consisting only of the speech verb (see examples (4)–(5), above), without further elaboration regarding participants or circumstances, etc., which is typical of introductory verbs.