Syntactic pausing? Re-examining the associations

2.1 Pauses and the syntax-prosody interface

Silent pauses have often been recognized as one of the most important signals of a prosodic boundary (Du Bois et al. 1992: 17, Himmelmann et al. 2018, Himmelmann 2022: 718, Ladd 2008: 288). They have also long been observed to pattern with syntactic boundaries, especially with clause boundaries (e.g. Cooper and Paccia-Cooper 1980; Goldman-Eisler 1968; Nespor and Vogel 2007). However, pauses do not necessarily co-occur with major syntactic boundaries, that is, they can be observed outside of syntactically determined positions, including word-internally (Bundgaard-Nielsen and Baker 2020; Paschen 2023). Such pauses are usually attributed to performance; for example, for hesitation, processing, style (cf. Nespor and Vogel 2007: 189, 219), and breathing, especially in faster speech (cf. Grosjean and Collins 1979). Syntactic structures are not always sufficient to predict the prosodic structure of an utterance, and several prosodic realizations of the same syntactic unit are possible.

There are thus various reasons to assume a certain degree of independence between syntactic and prosodic structures, while allowing for a high degree of interdependence at the same time. Such a position on the relation between syntax and prosody is captured by approaches which assume an intermediary prosodic structure to which morphosyntax is mapped (see e.g. Nespor and Vogel 2007; Selkirk 2011, 1984). Our study is compatible with the theoretical assumptions of this line of research, but we take an empirical approach by using cross-linguistic corpus data of spontaneous speech to examine to what extent pauses are associated with syntactic boundary positions.

2.2 Pause types and functions

There are different ways in which silent pauses can be distinguished, such as according to where they occur. Previous studies have distinguished between pauses at sentence, clause, and phrase boundaries as well as pauses that do not coincide with any major syntactic boundaries. Example 1 illustrates different possible locations of silent pauses in English in an excerpt from a spontaneous spoken text from the DoReCo corpus (Schiborr 2022). It contains four consecutive sentences split up for convenience, so that each line corresponds to a chunk of speech delimited by a silent pause whose length is indicated in parentheses.

The pauses following (1a) and (1d) coincide with main clause boundaries. The pause at the end of (1b) occurs at a dependent clause boundary. The three remaining rows include pauses that do not coincide with clausal or phrasal boundaries. In (1c), a pause occurs between a transitive verb and its object, where it follows a false start. A similar situation can be seen in the transition between (1e) and (1f), where a pause separates a determiner from its head noun.

Silent pauses such as those in (1a), (1b), and (1d) are typically understood within psycholinguistics as helping to facilitate the parsing of speech into syntactic units. This has been shown for both adult L1 language use (e.g. Frazier et al. 2006; Kjelgaard and Speer 1999; Petrone et al. 2017; Schafer et al. 2000) and for L1 and L2 language acquisition (e.g. Christophe et al. 2008; Goad et al. 2021; Hawthorne and Gerken 2014; Speer and Ito 2009). At the same time, these pauses also help the speaker to prepare and plan further utterances (e.g. Cooper and Paccia-Cooper 1980; Ferreira 1991; Fuchs et al. 2013; Goldman-Eisler 1968; Krivokapić 2007; Krivokapić et al. 2020). They can also be used for interactional purposes, such as yielding the floor during a conversation (Levelt 1993; Maclay and Osgood 1959; Wennerstrom and Siegel 2003).

Other pauses, such as the clause-internal ones in (1c), (1e), and (1f), are argued to result from high processing demands in situations in which speech planning is comparatively difficult. Planning difficulties can be caused by various factors, such as a long or complex following syntactic or prosodic unit, or difficulties in lexical access with low-probability items (see e.g. Beattie and Butterworth 1979; Belz et al. 2017; Goldman-Eisler 1961a; Hartsuiker and Notebaert 2010). Clause-internal pauses may also be associated with attempts to hold the floor during an interaction (Himmelmann 2014; Local and Kelly 1986; Wennerstrom and Siegel 2003).

As we work with corpus data in the present study, we cannot test for specific cognitive processes that would explain pausing in a given context. As such, we do not distinguish between pauses that occur for hesitation, planning, or stylistic reasons. We instead distinguish between three types of pauses based on their location in relation to syntactic boundaries: pauses within clauses, pauses at main clause boundaries, and pauses at dependent clause boundaries (see Section 3 for more details and examples). This distinction allows us to measure the association of different clausal contexts with pauses, and to assess how robust the patterns found are across languages (see Section 4.2).

2.3 Pause duration

Studies into the distribution of pause duration often show a distinction between brief pauses and longer pauses, with longer pauses typically found at sentence boundaries and shorter pauses at dependent clause boundaries and within clauses.^[2] In an early study, Goldman-Eisler (1972) reports for spontaneous spoken English data that pauses between sentences mostly have durations above 500 ms, while pauses between clauses (within sentences) tend to be shorter than 500 ms. Grosjean and Deschamps (1972) report comparable pause durations for other French and English datasets. Fletcher et al. (2004) similarly find that pauses within and between words in Dalabon (Gunwinyguan, Australia) fall into two groups in terms of duration: medium pauses (around 500 ms) and long pauses (around 1,000 ms or more).

Pause distribution and duration in spontaneous French data show similar trends, even when more detailed distinctions of pause type are made. Duez (1982: 24) finds that pauses within phrases (∅ = 401 ms) and within clauses (∅ = 632 ms) tend to be shorter than pauses between clauses (∅ = 802 ms) in casual interviews. In a more recent study, Candea (2000) examines French spontaneous speech data and reports that pauses at sentence boundaries have average durations of 900 ms, while pauses at clause boundaries within sentences average 600 ms (Candea 2000: 168). In contrast, hesitation pauses within clauses are reported to have an average duration of 560 ms (Candea 2000: 181).^[3] Campione and Véronis (2002) focus on pause duration in French spontaneous speech data and read speech from five European languages. They find a trimodal distribution of pauses in terms of duration and distinguish between brief (<200 ms), medium (200–1,000 ms) and long (>1,000 ms) pauses. Read speech of all five languages only featured brief and medium pauses, with long pauses occurring in addition to short and medium pauses in spontaneous speech.

While the exact details differ across studies, the picture emerges that we can distinguish between different pause types in terms of their duration, and that pauses at clause boundaries tend to be longer than pauses within clauses, at least for English and French. We assess to what extent this association of pause location and duration is reflected in our data, how it relates to previous findings, and how robust the patterns are across languages (see Section 4.3).

3.1 Dataset

In our study, we investigated how pauses were distributed across seven typologically distinct languages in the Multi-CAST corpora (Haig and Schnell 2021). The languages are listed in Table 1 together with information on the number of different speakers in the corpus, the number of utterances (as defined by Multi-CAST), and the genres of texts used. We chose a subset of data which consisted of monological data, specifically mostly autobiographical narratives (AN) and traditional narratives (TN), with the addition of one stimulus-based narrative (SN) in Tondano. As can be seen in Table 1, each corpus ranges from 160 to 1,300 utterances from two to six different speakers. We used the WAV files along with the accompanying annotations in EAF format provided in Multi-CAST (Haig and Schnell 2021).

3.2 Clause annotation

The majority of spontaneous speech corpora are created for language documentation purposes, and as such, differ as to how elaborate their morphosyntactic annotation is. Information on clausal boundaries below the utterance level is usually not explicitly annotated for, Multi-CAST being an important exception.

We took the annotation of clausal boundaries from GRAID (Grammatical Relations and Animacy in Discourse) annotations in Multi-CAST (see Haig and Schnell 2014). In GRAID, a clause is generally defined as a predicate with its arguments, with more language-specific criteria for how to deal with, for example, multi-verb predicates (Haig and Schnell 2014: 45). Main clauses are signalled with an annotation of “##” at their left edge, while dependent clauses are annotated with “#” at the left edge. The end of a dependent clause is generally not marked explicitly, as it usually coincides with the beginning of a new main clause. However, the end of a centre-embedded dependent clause that does not correspond to the end of a main clause is additionally signalled by “%” in GRAID.

Examples (2)–(4) show how clause boundaries are annotated in GRAID (see Haig and Schnell (2014: §2.6, §4.1) for more details).^[4] Example (2) consists of a main clause in Mandarin, marked as such by the initial “##”.^[5]

In example (3), we see an utterance from the Northern Kurdish corpus, consisting of a main clause with a dependent relative clause. In this case, the end of the relative clause corresponds to the end of the main clause and does not receive an explicit annotation.

The Nafsan example in (4) shows a somewhat more complex utterance. It is made up of three main clauses, the first of which contains an additional dependent clause. Note that the end of the dependent clause does not coincide with the end of the main clause, which is why it is additionally marked with the annotation by “%” in (4a).

We used this clause information from the original GRAID annotation to distinguish between main clause boundaries (cb_main), dependent clause boundaries (cb_dep), and clause-internal positions (no_cb). Main and dependent clause boundaries as shown in (2)–(4) were directly taken from the annotation. The remaining annotations were classified as clause-internal.

3.3 Pause annotation

Spontaneous speech corpora often do not include annotations for phonological or prosodic properties.^[6] As such, the re-use of the Multi-CAST corpora for this paper required us to additionally annotate silent pauses in order to explore any potential associations between pauses and clause boundaries. The pause annotation process we followed is described in detail in the supplementary document “data-extraction-processing.pdf”; we provide a brief summary here and some illustrative examples for the reader’s orientation.^[7]

We firstly annotated silent pauses using a silence recognizer and followed this with a round of manual checks and corrections. This resulted in an annotation of pause and speech segments. We then classified both pause and speech segments into one of three types (cb_main, cb_dep, no_cb) according to whether or not a main or dependent clause boundary fell into their duration. This was done by automatically comparing the time stamps of the pre-existing clause boundary annotations and those of the newly annotated silent pause and speech segments. Since the Multi-CAST data is only aligned with the acoustic signal on the utterance level and not on the clausal level, we could not automatically specify the location of a pause occurring within the utterance further. As described in more detail in Section 4 of the supplementary materials, we performed selected manual checks of the position of clausal boundaries in relation to pauses, which suggested that the automatic extraction works sufficiently well for our purposes.

Tables 2–6 show examples of the resulting pause annotation for examples (2)–(4) from Section 3.2. The first row shows the utterance including the original clause boundary annotation from Multi-CAST. The second “pause” row then shows whether or not a silent pause was detected. If a pause was detected, the “type” row indicates which type of pause we are dealing with, and the “dur” row shows its duration. In the Mandarin example in Table 2, we see that both the initial and final clause boundaries coincide with a pause, which we annotated as cb_main. In addition, we find another silent pause occurring within the clause, which was marked as no_cb.

In the Northern Kurdish utterance in Table 3, the main clause boundaries also coincide with silent pauses. The dependent clause boundary between the two main clauses, however, is not accompanied by a silent pause. In addition, no pauses occur within the two clauses shown in Table 3. The Nafsan example from (4) is shown in Tables 4 to 6. Table 4 contains the first main clause together with its dependent clause. Here, silent pauses only occur at the end of the dependent clause (cb_dep) and within the second part of the main clause (no_cb). Table 5 shows that another silent pause coincides with the boundary between the two main clauses in Tables 4 and 5. Finally, we see in Table 6 that both clausal boundaries co-occur with a pause (cb_main), and that the main clause itself also contains a silent pause (no_cb).

4.1 Overall tendencies

We first inspect the distribution of clause boundaries. Table 7 shows how the three clausal contexts are distributed across speech and pause segments (“No. of segments”) in the seven languages. By segments we refer to contiguous chunks of speech or silent pause in the speech signal. Pause/speech segments with no clause boundaries (no_cb) make up around 55–70 % of annotations in all languages, with an average of 61 % across the dataset. The proportions of cb_main and especially cb_dep contexts are more language-dependent. For example, dependent clause boundaries make up 23 % of all boundaries in Northern Kurdish, while they only amount to 4 % in Mandarin.

Figure 1 builds on the distribution of clausal contexts shown in Table 7 by including information about how they relate to pause and speech segments in the seven languages. Comparing the distribution of the three clausal contexts in pause and speech segments (left vs. right bars), we see clearly that pause segments contain a higher proportion of clause boundaries than speech segments in all languages. The exact proportions differ across languages, with clausal boundaries co-occurring with 75 % of all pauses in Northern Kurdish but with only around 40 % of all pauses in Tondano.

Figure 1:

The raw distribution of clausal contexts across speech and pause segments.

The overall distribution of clausal contexts across our dataset suggests that clause boundaries more often co-occur with a pause than with speech. However, in Tondano and Arta, pauses are more likely to occur within clauses than in either of the clause boundary conditions. The probability of pauses co-occurring with dependent clause boundaries is also substantially lower in Tondano, Mandarin, and Teop than in other languages, likely due to the overall low proportion of dependent clause boundaries in these languages to begin with (see Table 7). This suggests that the presence of a clausal boundary does not necessarily entail the occurrence of a silent pause and that the association between clause boundaries and pauses is indeed more complex than a simple one-to-one relationship.

4.2 The association of clause boundaries with pauses

The first question that this paper addresses is to what extent the different clausal contexts are associated with silent pauses. To do this, we fitted a Bayesian logistic regression model to assess the probability of pauses across the three types of clausal contexts. We added varying intercepts as well as varying slopes over clausal contexts for individual speakers.^[8] The model was fitted using Bayesian methods with Stan (Carpenter et al. 2017) and brms (Bürkner 2017) in R (R Core Team 2021).

Figure 2 shows the conditional effects of clausal contexts on the probabilities of pauses. The points in Figure 2 correspond to the mean of the probability density of the predictions; the whiskers correspond to the 95 % uncertainty intervals. This means that we can be 95 % confident that the actual probability will lie within this interval based on the data and the model.

Figure 2:

Predicted probability of pauses across clausal contexts.

While we see a certain degree of variation across the seven languages, the overall pattern is robust. The probability of pauses occurring gradually increases from environments within clauses (no_cb) to dependent clause boundaries (cb_dep) and main clause boundaries (cb_main). For main clause boundaries, the probability of co-occurrence with pauses is between 0.5 and just over 0.75 in all languages, the lowest prediction being 0.53 for Tondano and the highest being 0.76 for Arta. For dependent clause boundaries, the mean posterior co-occurrence probabilities for all languages are slightly lower, ranging from 0.29 in Teop to 0.57 in Nafsan. For the no clause boundary context, we find the overall lowest probability of pause co-occurrence, from 0.14 in Teop to 0.28 in Tondano. We furthermore generally find a higher level of cross-linguistic variation in the cb_main and cb_dep contexts than in the no_cb condition.

The robust patterning of pause co-occurrence probabilities across our dataset illustrates a clear trend. As the level of the syntactic boundary juncture increases, so too does the overall predicted probability of a pause co-occurring. Looking at the uncertainty intervals, we find that they overlap between “neighbouring” categories, that is, within clauses and dependent clause boundaries, as well as dependent and main clause boundaries. Only Nafsan differs in that both types of clause boundaries have a similarly higher probability of pauses compared to boundaries within clauses. Still, the overall results suggest that there is no cross-linguistically robust, binary distinction either between main clause boundaries and other contexts or between non-clausal boundaries and clausal boundaries. Instead, we found a gradual increase in pause probability from no_cb to cb_dep and cb_main.

4.3 Pause durations

The second question that this paper aims to address concerns pause durations in relation to their location. Figure 3 shows the raw distribution of pause durations across the seven languages of the dataset. The general trends are fairly robust. The majority of silent pauses in the dataset have a duration up to 750 ms, few pauses are longer than 1,000 ms, and pauses longer than 2,500 ms are rare. However, Arta, Nafsan, and Northern Kurdish have a somewhat larger portion of silent pauses longer than 750 ms.

Figure 3:

Observed pause duration by language.

Furthermore, the average pause duration is more or less comparable across languages. As can be seen in Figure 3, the median pause duration (the solid vertical line) is between 500 ms and 1,000 ms for all languages, apart from Mandarin. However, we see more variation of the mean pause duration (dotted black line) across languages. Arta, Nafsan, and Tondano show a larger difference between median and mean pause duration, as these three languages have more longer pauses than the other languages of the dataset. Still, we can say that the distribution of pause length, albeit not identical, is comparable across the seven languages.

Figure 4 shows the distribution of pause duration together with median and mean values across the three clausal contexts with data from all seven languages pooled together. These distributions along with the median and mean suggest a difference in pause duration associated with the absence or presence of a clause boundary. Most pauses within clauses have a length of below 750 ms, with only 15 % of all pauses being longer than 1,000 ms. In contrast, we find quite a substantial proportion of pauses with durations above 1,000 ms up to 3,000 ms for both types of clause boundaries, namely 40 % for cb_dep and 38 % for cb_main, respectively. This binary difference between pauses within clauses and pauses at clausal boundaries is further reflected in their median pause durations: 537 ms for no_cb versus 796 ms and 788 ms for cb_dep and cb_main, respectively.

Figure 4:

Observed pause duration by clausal context.

To test the robustness of the observations from the raw distributions, we fitted a Bayesian regression model to predict the duration of pauses from clausal contexts and languages while controlling for the effects of single speakers.^[9] Figure 5 shows the conditional effects of clausal contexts and languages on pause durations. The model predictions support the previous observation of a two-way distinction of pauses based on their duration. Pause durations are predicted to be between 524 ms and 781 ms very consistently across languages when no clause boundary is present. At main and dependent clause boundaries, the average predicted pause duration varies to a greater extent within and across languages. Pauses are predicted to be between 1,000 ms and 1,300 ms for both types of clausal boundaries in all languages except for Arta. Arta exhibits shorter pause durations in all three contexts compared to the other languages. However, the relative difference between the boundaries within clauses (525 ms) and clausal boundaries (802 ms and 814 ms) is maintained. We should be careful before concluding that this indeed reflects a real cross-linguistic difference, as the Arta subcorpus is comparatively small (see Table 1).

Figure 5:

Estimated pause duration by clausal context.

Larger datasets with data from more speakers are needed to be more certain about what seems to be a language-specific preference towards shorter pauses in our data.