Linguistic disfluencies in Russian-speaking typically and atypically developing children: individual variability in different contexts

ABSTRACT Disfluency in children and adults seems to occur like errors of speech but, at the same time, is an essential feature of spontaneous (unprepared) speech. The present study aimed to evaluate linguistic disfluencies in typically and atypically developing Russian-speaking children from the perspective of the dynamic adaptive model of self-monitoring in speech production. The study collected four language samples from 10 six-year-old children with developmental language disorder and 14 typically developing peers: two storytelling tasks, structured conversation, and a play argument. After transcribing audio-recordings and marking linguistic disfluencies, the authors conducted structured distributional analysis. The distribution of several indexes of disfluency was estimated to assess the prevalence and profiles of different (sub)types of disfluencies. The disfluency rate statistics were similar between the typically developing children and children with developmental language disorder. The distributional indexes score showed that tasks significantly impacted the rate of different (sub)types of disfluencies. Task-related patterns in a set of the distributional indexes significantly distinguished the groups. Thus, changes in the disfluency profile related to different external factors, as a sign of a flexibility of an adaptive self-monitoring system, may be limited in children with developmental language disorder.


History of studies in linguistic disfluency
Children and adults make speech errors regularly.In adults, such occasional errors are termed slips of the tongue (Fromkin, 1973), but in children, they are typically treated as signs of limited language competence (Shanker, 2002).However, even preschool age children can predict and, thus, try to prevent, some speech errors, as well as notice and repair them.These attempts, paradoxically, also lead to specific types of disfluencies, such as self-repairs, false starts, and pauses (McDaniel et al., 2010).From this perspective, a comparative analysis of disfluencies between typically developing (TD) children and children with developmental language disorder (DLD) might shed light on disfluency nature recognition in the developmental context.
The current study is devoted to linguistic disfluency.The following types of disfluencies may be noted as the most regularly attributed: hesitations, revisions, repetitions, and false starts (Altıparmak & Kuruoğlu, 2018;Gyarmathy & Neuberger, 2013;Jansson-Verkasalo et al., 2021;Manfra et al., 2016;McDaniel et al., 2010;Morgenstern et al., 2013;Navarro-Ruiz & Rallo-Fabra, 2018).In almost all these studies, hesitations have been further divided into filled and unfilled pauses, whereas revisions (also termed 'self-repairs') have been classified in multiple ways.Among others, such subtypes as 'phonological revisions', 'lexical revisions', 'morphological revisions', 'syntactic revisions', 'self-repaired words', 'semantic self-repairs', 'corrections of word choice', 'corrections of reference', 'word order corrections', 'grammatical corrections', and 'accentual correction' have been mentioned.Thus, considering the lack of commonly accepted taxonomy, 'linguistic disfluency' may be recognised rather as an umbrella term.It covers the functionally different tricks that speakers apply to resolve a competition between the quick online flow of a speech string, on the one hand, and the necessity to execute several tasks related to utterance content planning, linguistic structuring, and self-monitoring in parallel, on the other.

Language development and speech fluency
From the first steps into verbal communication development, children acquire language units and system and master speech production skills -they learn to combine two or more words having different syllable structures integrated in proper prosodic patterns.At the final stage of language acquisition, a child is expected to build conceptual structures, to transform them fluently into relevant linguistic constructions, and, almost simultaneously, to produce prosodically proper articulatory sequences.A progress in language and speech development is most obviously reflected in two measures: speech accuracy and utterance production fluency (DeJoy & Gregory, 1985;Holm et al., 2007Holm et al., , 2022;;Jones, 2020).In this way, children reduce speech error rate continuously and become more skilful in automatised fluent utterance execution.However, in both children and adults, only prepared speech is fluent; spontaneous (unprepared) speech is characterised as being disfluent (Dufour et al., 2009).

Self-monitoring and disfluency
Research has proposed several psycholinguistic models of utterance programming.Within the framework of these models, despite some theoretical distinctions, speech errors (deviations from a speech plan), disfluencies (interruptions in the execution of a speech plan), and self-repairs (prevention or corrections of the speech errors) are considered as being connected to one another (Levelt, 1989;Postma, 2000).Structurally, most scholars agree on the existence of three major components of speech production, which work relatively autonomously to handle: conceptualisation, formulation, and articulation (Blackmer & Mitton, 1991).
Conceptualisation is recognised as a controlled process, whereas formulation and articulation are highly automatised (Bock, 1982;Levelt, 1989).Self-control and self-repair processes are closely related to speech production (Levelt, 1983(Levelt, , 1989;;Postma et al., 1990;Postma, 2000).Self-control is addressed to all levels of speech production (Blackmer & Mitton, 1991).Levelt (1989) proposed three monitoring processes that are employed: 1) during conceptualisation, 2) during the inner speech resulting from the speech planning process, and 3) at the stage when the articulating program was made and executed, and one can listen to what one says (Levelt, 1989).However, owing to timing pressure and resource limitations, speakers '[. ..] do not continuously attend to all these things simultaneously.Attention is on the one hand selective, and on the other hand fluctuating.Which aspects of speech are attended to is highly dependent on the context and on the task.A speaker can be set to attend to certain kinds of errors or dysfluencies, and to ignore others' (Levelt, 1989, pp. 499-498).MacWhinney and Osser (1977), in a developmental study with five-year-old TD children, assumed three major functional categories of speech planning, namely, pre-planning, co-planning, and avoidance of superfluous verbalisation.These styles in verbal planning reflect basic differences in cognitive processing.The authors expected that while a speaker is attempting to figure out what to say and how to say it, a conversation moves on and a child must choose one of two ways.The first way is to attempt to formulate fully what is going to be said before they say it.The second way is to start talking and hope to be able to figure out their utterance in time but with higher risk of making errors.
According to developmental psycholinguistic and neurobiological data, developing speech and language devices are served by the maturity of feedback mechanisms that underly the self-monitoring activity (Nozari et al., 2011).The developed self-monitoring activity prevents the speaker from speech errors and/or allows to correct them but, at the same time, provokes different types of disfluencies (McDaniel et al., 2010).Previous studies have evaluated child speech disfluency and observed great inter-and intra-individual variability (Silverman, 1971(Silverman, , 1973)).Several reports have shown that some circumstances of speech elicitation (e.g.age and familiarity degree of the partner, stressful conditions of communicative situation, familiarity of the experimental procedure) impact the number and distribution of disfluency (Silverman, 1971(Silverman, , 1973)).

Discourse acquisition and social determinants of disfluency
Becoming a 'proficient speaker', in the sense of Berman and Slobin (1994), means mastering many complex and flexible skills and strategies for producing cohesive and coherent discourse texts according to the multiple social and cultural traditions a child faces within their community.Children learn to speak via interaction with adults, grounding on their positive and negative evidence.Children learn to revise and repair their speech to meet the requirements of the language and of different activities (Salonen & Laakso, 2009).Four-year -olds and older children often repair their utterance to adjust the speech acts to the listener or communicative context.In parallel to the line of speech development, children learn to use language in different pragmatic, social, and communication contexts.From the first steps into language acquisition, children master oral discourse skills moving from colloquial conversation to personal narratives, fictional stories, and other genres.Some studies have reported that genres of discourse differ in the cognitive demands they present to a speaker (Levin & Silverman, 1965;Manfra et al., 2016;McDaniel et al., 2010;Navarro-Ruiz & Rallo-Fabra, 2001).Meanwhile, McDaniel et al. (2010) presented much evidence on the mean length of utterance (MLU) and syntactic complexity impact on the number of disfluencies.Thus, discourse genre can be expected to be related with disfluency rate in children.Studies have revealed an impact of genre of discourse on the distribution of the parts of speech (Balčiūnienė & Kornev, 2020), error types (Kornev & Balčiūnienė, 2021), and syllables with different complexities (Kornev & Balčiūnienė, 2020).Additionally, some external circumstances of speech production in children have been identified to influence error and disfluency rates (Silverman, 1971(Silverman, , 1973)).In TD children, linguistic disfluencies have been characterised as being caused by distinct strategies of speech production attributable to the subject's language competence (Balčiūnienė & Kornev, 2016).

Speech/Language disorders and disfluency
Among clinical populations of children with impaired speech, a great number of studies has been carried out in stutterers (Boscolo et al., 2002;Logan et al., 2011;Ratner & Sih, 1987).Most of them have highlighted stuttering-like disfluencies.Studies in non-stuttering children with DLD are relatively scarce, and, moreover, they have presented contradictory evidence.In some studies, children with DLD have been reported as being more disfluent than their TD peers (Cáceres-Assenço et al., 2014;Finneran et al., 2009).In other studies, TD children and children with DLD do not have differences in disfluency rate (Lees et al., 1999).Such a dissociation may be caused, presumably, by the great variability in the existing measures of disfluency, its taxonomy, or the analysis methods (qualitative vs. quantitative evaluation; percentage vs.raw data).Mainly, stuttering-like disfluency has been identified as a characteristic for distinguishing between TD children and children with DLD (Befi-Lopes et al., 2014;Boscolo et al., 2002;Hall, 1999;Lees et al., 1999).Meanwhile, linguistic disfluency has been reported as being similar in the given populations (Lees et al., 1999;Sormani, 2010).

Aims of the current study
Research has identified various complexes of different adaptive tricks related to the speech self-monitoring sub-system (Gauvin et al., 2016;Levelt, 1983) and the executive control of the speech production performance (Bourguignon, 2014).We hypothesised that in pre-and during utterance production, different external circumstances (e.g.communication context, speech elicitation method, socially determined discourse genre), as well as internal factors (e.g.language development status), would impact the distribution of linguistic disfluency types and sub-types.We predicted that in each child, the probability of linguistic disfluency phenomena would be occasional.Linguistic disfluencies are supposed to be influenced by such determinants as flexibility of speech production (high in TD children and low in children with DLD) and discourse genre.We presumed that the more adaptive and flexible their speech production and self-monitoring systems, the more variable patterns of disfluency should occur in different circumstances (i.e.different discourse genres in our model) (Cañas et al., 2003).To test this hypothesis, we compared disfluency distribution, represented by four discourse genres, in the speech of TD children and children with DLD.

Participants
Our participants were 10 monolingual Russian-speaking pre-schoolers with DLD and 14 TD peers (the mean age in both groups was 6 years and 5 months).All participants lived in Saint Petersburg, Russia, and attended state kindergartens daily.All children with DLD were clinically referred and received a two-year course for speech therapy (five sessions per week) at the kindergarten; nevertheless, various phonetic, lexical, and grammatical errors still occurred in their speech.Previous studies of the same sample revealed that among others, lexical and morphological errors were dominant in the DLD children (syntactic and derivational errors were much less numerous) but only the morphological errors (namely, the number of morphological errors per word) significantly (p = 0.047) distinguished the DLD children from TD peers (Kornev & Balčiūnienė, 2021).Before the experiment, the children were assessed by a speech-language pathologist using Russian language assessment tools (Kornev, 2006) to confirm the TD vs. DLD status and to exclude children with language comprehension disorders from the study.(Children with language comprehension disorders were excluded to escape the non-relevant variable impact and to highlight the discourse production features related to the expressive language limitations.)Thus, the DLD group may be characterised as children with developmental expressive language disorder with intact language comprehension.The children were also assessed by a psychologist using the Raven's Coloured Progressive Matrices Test to check their nonverbal IQ.Children with hearing or visual disorders, neurological disorders, or nonverbal IQ < 84 were not included in the study (nonverbal IQ was p = 113.25,SD = 3.15 in the children with DLD; and p = 121.00,SD = 4.90 in the TD children).Before the study, informed consent was obtained from all subjects' parents.The study was approved by the Ethics Committee of Saint-Petersburg State Pediatric Medical University (#4/2, 13 July 2011).

Procedures
Each participant took part in two individual interactive play sessions of a (semi-) structured speech elicitation.During Session 1, fiction narratives (storytelling and retelling according to picture sequences) and conversations based on comprehension questions were collected.During Session 2, playful argumentation conversations were elicited.

Session 1
The children were asked to tell and retell a story according to different picture sets.We used two sequences consisting of six coloured pictures each.At the beginning of the assessment, the experimenter placed the pictures of one of the sequences in a single horizontal row in front of a child.First, the child was encouraged to look at the pictures to get the gist of the story.Second, the child was asked to tell a story according to the pictures (for storytelling mode) or to listen to the story read by the experimenter and then to retell it (for retelling mode).During the telling/retelling, the pictures remained on the table, and the child had a possibility to look at them throughout the assessment.
After the telling/retelling, the child was asked 10 comprehension questions.This type of conversation may be equivalent to a structured interview.The tasks were separated by a few minutes of chat between the experimenter and the child.The methodology is described in more detail in Balčiūnienė and Kornev (2016).

Session 2
To elicit argumentation conversations, the study used so-called 'nonsense pictures'.An experimenter showed a child several ridiculous pictures containing unrealistic content (e.g. a fish sitting in a cage, a pig climbing a tree) and asked the child to evaluate its plausibility.During the assessment, the experimenter used as many provoking questions and statements as possible to involve the child in a discussion and argumentation.The conversations were carried on in the manner of a spontaneous colloquial chat.The methodology is described in more detail in Kornev and Balčiūnienė (2020).

Data
All the sessions were audio-recorded using a portable recorder.The audio-records were transcribed orthographically using CHAT tools (MacWhinney, 2000).The transcripts were segmented into utterances.Two experts double-checked the transcripts independently and extended them by encoding language errors and linguistic disfluencies.The developed corpus (27,322 words in total) was divided into four sub-corpora of different discourse genres (Table 1).
Our interest was typical (also termed 'linguistic') disfluency, not the stuttering-like one.All linguistic disfluencies in our data were classified into five types: hesitations, repetitions, revisions, false starts, incomplete utterances.Among them, three types of disfluencies (hesitations, repetitions, and revisions) were selected for further classification into subtypes.
Repetitions were considered as cases in which the child repeated a part of a word (Example 4), a single word (Example 5), or a string of words (Example 6) within their utterance.Typical examples of the given subtypes were as in the following.
'Rats (incorrect phonology) . . .rats cannot be there'.False starts were identified as utterances dropped after just one or a few words (Example 10).
'She. . .I presume a lunch will be served to her'.
Incomplete utterances were considered as utterances where the obligatory ending is missing, although the general idea of the entire utterance may be predicted by a listener (Example 11): (11) Potomu, cho vse tut nepraviln--. . .'Because everything here is incorrect (incomplete word) . . .' Speech disfluency has often been discussed as errors, with an implication of a relation between error and disfluency rates.In this regard, we measured the rate of linguistic (lexical, morphological, syntactic, and derivational) errors in the corpus (for further details on the error analysis, see Kornev & Balčiūnienė, 2021).

Variables analysed
The frequency of disfluencies has been estimated by the raw numbers per text (Navarro-Ruiz & Rallo-Fabra, 2001), per word (Boscolo et al., 2002;Enger et al., 1988;Ratner & Sih, 1987), or even per syllable (Logan et al., 2011).Other researchers have estimated the mean number of disfluencies per utterance (Bortfeld et al., 2001;Plevoets & Defrancq, 2018;Schnadt & Corley, 2006;Zackheim & Conture, 2003).From the perspective of our hypothesis, we deemed it relevant to count the mean number of linguistic disfluencies per utterance, given that both self-monitoring and non-fluent speech manifest in the utterance (but not word or syllable) context.
We estimated three types of variables: disfluency rate (three different indexes), error rate per utterance, and MLU.The error rate and MLU were supposed to be the main determinants that might have an impact on the disfluency variables.As for disfluencies, the following indexes of quantitative measures of disfluency distribution were estimated: 1) disfluency (sub)type rate per utterance (DSRpu), 2) disfluency type ratio within all disfluencies (DTRwad), and 3) disfluency subtype ratio within the type (DSRwt).
The DSRpu was estimated by dividing the number of linguistic disfluencies of the given (sub) type by the total number of utterances in the (sub)corpus.It represents a general probability of the given (sub)type of disfluencies in discourse.The DTRwad was estimated by dividing the DSRpu of the given type of linguistic disfluencies by the DSRpu of all disfluencies.It represents a probability of the given (sub)type of disfluencies among all disfluencies.The DSRwt was estimated by dividing the subtype DSRpu on the type DSRpu.It represents a probability of the given subtype of disfluencies among the given type of disfluencies.
Statistical differences between the measure distribution were calculated by means of the Kruskal -Wallis H-test.During the statistical analysis, it was revealed that some variables in our sample did not fit a normal distribution and, thus, the nonparametric statistical test was considered as more appropriate.On the other hand, the given statistical test was the most suitable for the comparison of distributions between the groups.In each statistical comparison, only one hypothesis was tested, i.e. whether the distributions of the analysed index were significantly different between the samples; thus, a correction for multiple comparison was not applied.Generalised linear model (GLM) (Myers & Montgomery, 1997) was applied to estimate the impact of the independent variables on the dependent variables.Causal relations between variables were measured by multiple regression analysis.Within each of the groups (TD and DLD), two independent variables (mean error rate and MLU rate) were included in the model; the DSRpu of all disfluencies merged was included as the dependent variable.
The GLM-analysis of the entire corpus (including all participants, regardless of language development status) revealed that the DSRpu index for all disfluencies per utterance was uneven among the participants and sub-corpora and depended on the disfluency (sub)type and discourse genre (conversation vs. narrative) (Table 3).The language development status (Group), however, did not determine a prevalence of all disfluencies per utterance rate.

Linguistic disfluency in TD children vs. children with DLD
The between-group comparison of DSRpu (disfluency (sub)type rate per utterance) in the entire corpus (all genres merged) revealed only significant difference in filled hesitations (Table 3).According to the main question raised, between-group differences of the disfluency rate within different discourse genres were analysed.The results are presented in the Figure 5 and Table 4.The distribution of DSRpu within different genres was very similar between TD children and children with DLD (Table 4), with only a few differences.Children with DLD produced more lexical revisions in telling, fewer filled hesitations in retelling, and fewer unfilled hesitations in the argumentation conversation in comparison to their TD peers.
The question about the DSRwt (disfluency subtype ratio within the type) in children with DLD in different genres compared to their TD peers was raised as being one of the most important in the framework of the current study.Comparative analysis of the DSRwt between the genres revealed only few significant differences in children with DLD, in contrast to their TD peers: in this group, only the filled and unfilled hesitations discriminated the retelling from argumentation conversation (H = 4.345; p = 0.037).As for the percentage of different sub-types of revisions among all revisions, most of them were represented by lexical and grammatical revisions in both groups.

Regression analysis of relations between disfluency, linguistic error, and MLU rates
To analyse potential relations between the error rate and MLU on one hand and the disfluency rate on the other, we conducted multiple regression analysis.Within each of the groups (TD and DLD), two independent variables (mean error rate and MLU rate) were included in the model; the DSRpu (disfluency (sub)type rate per utterance) of all disfluencies merged was included as the dependent variable (Figures 7-8).Statistical analysis revealed that only the MLU rate significantly regressed on the DSRpu of all disfluencies merged both in the TD group (β = 0.450; t = 4.118; r 2 = 0.181; t = 0.001) and DLD group (β = 0.371; t = 3.009; r 2 = 0.124; t = 0.004).

Discussion
Similar to previous research in child speech (Silverman, 1971(Silverman, , 1973)), our results confirmed that both TD children and children with DLD produced plenty of disfluent passages during speech production.Our data demonstrated great variability of disfluency in both TD children and children with DLD.The given variability was evidenced both in quantitative measures (number per utterance) and qualitative features of disfluency distribution.Many differences were related to discourse genre demands.
The GLM results highlighted an internal factor ((sub)type of disfluency) and external determinant (discourse genre) that impacted significantly on the changes in the flexibility of the disfluency rate.Among them, Disfluency type had the strongest power.The given results are congruent with other previous studies.
Both in TD children and children with DLD, hesitations occupied the main portion among all disfluency types.Other studies have also shown that hesitations play an essential role in utterance planning, especially if the utterance has a complex syntactic and/or cognitive structure or the speaker struggles with finding a proper word (Schnadt & Corley, 2006).Filled and unfilled hesitations have different functions.Unfilled hesitations give a speaker extra time for speech programming, whereas filled hesitations also function as discourse markers (Gósy, 2001;Roggia, 2012).The latter role has been identified as less developed in children with DLD than in TD children (Kornev & Balčiūnienė, 2020).
The next most common type was repetitions.According to the hypothesis, linguistic disfluency, as a dynamic system related to the self-monitoring system (Levelt, 1983), was expected to be flexibly adaptive to changes in various internal and external conditions of speech elicitation/production.In our study, the main changes the children faced were in task distinction, which demanded the proper genre of discourse.We assumed that more information on disfluency flexibility may be given by the between-genre discriminability of the DTRwad (disfluency type ratio within all disfluencies) and DSRwt (disfluency subtype ratio within the type) in both groups.They should be treated as signs of some reorganisation in the self-repairing activity.In the TD children, we found 23 between-genre distinctions in the DTRwad and DSRwt; in children with DLD, the between-genre distinctions were very rare: we found only seven distinctions in the DTRwad and one in the DSRwt.The self-monitoring and self-repair systems in the children with DLD may be less flexible compared with the TD children.During speech production, the children with DLD reacted to different tasks and circumstances by the same or very similar manner -almost without any adaptations.
Several studies have reported that programming and executing different discourse genres cause different cognitive demands to a speaker (Berman & Slobin, 1994;Cannizzaro et al., 2019).These circumstances presumably impact both speech accuracy and self-monitoring strategies children apply to avoid speech errors.A previous study argued that the genre of discourse impacts the distribution of linguistic errors and speculated that the linguistic error rate and disfluency production may be related (Kornev & Balčiūnienė, 2021).However, regression analysis carried out in the current study did not support this expectation.Moreover, we found that the MLU rate significantly positively impacted the disfluency rate.The longer the utterance is planned, the more difficulties arise with fitting cognitive demands to available resources.Similarly, long and/or complex utterances have been identified as provoking more linguistic disfluencies (Didirkova et al., 2019;Lee & Hwang, 2001).
We found relatively low between-group differences in the disfluency rate per utterance, in contrast to several other studies that found many more disfluencies in children with DLD.This difference may be partially explained by the fact that all our participants with DLD were recruited from a specialised day-care centre for children with speech/language disorders; the children with DLD had been taking a remedial course (2-3 years) with a consequence of partial compensation of DLD.The children became more fluent in their discourse but still had residual limitations in their flexibility and adaptability in terms of their self-monitoring and self-repairing systems.
Within the discourse genre (story telling/retelling vs. conversation), we found some differences between the disfluency types (hesitations, repetitions, revisions, false starts, and incomplete utterances).Different proportions of false starts and incomplete utterances were highlighted between question -answer and argumentation conversations.Although both conversation types refer to the same discourse genre and both were executed in the form of a dialogue, the question -answer conversation was based on prepared verbal content (story) and elicited as a semi-formal dialogue, whereas argumentation conversation was based on a complex picture (a drawing representing unnatural content for argumentation) and organised in the manner of a playful chat.Similar findings were obtained by Manfra et al. (2016) in their study with preschool children: differences between private and social speech have an impact on the number of disfluencies.As such, the difference in the disfluency (sub)type rate per utterance (DSRpu index) and that in the disfluency subtype ratio within the type (DSRwt index) may be related to some qualitative features of individual self-monitoring systems, on the one hand, and to the complexity of utterance and its resource demands, on the other.
Different types or subtypes of disfluency play different functional roles in speech production (Levelt, 1989;MacWhinney & Osser, 1977).For example, revisions and repetitions are assumed to be a part of self-repair processes (Levelt, 1983), whereas hesitations are considered related, to some extent, to utterance planning (Postma, 2000).In this regard, different proportions of the (sub)types among all disfluencies may refer to different difficulties and the coping strategy choice of the speaker.
A novel finding of our study was on the disfluency rate between the groups of TD children and children with DLD.We proved our prediction that the higher the proficiency and flexibility of speech, the more probable the genre influences the distribution of (sub) types of disfluencies.

Limitations
Our study has some limitations.The sample size was rather small, which restricted the validity of our conclusions.Another limitation is the cross-sectional design of the study.We plan to continue investigations from the perspective of longitudinal observation.

Figure 1 .
Figure 1.DSRpu in different discourse genres in TD children.

Figure 2 .
Figure 2. DTRwad in different discourse genres in TD children.

Figure 3 .
Figure 3. Repetitions DSRwt in different genres in TD children and children with DLD.

Figure 4 .
Figure 4. Revisions DSRwt in different genres in TD children and children with DLD.

Figure 5 .
Figure 5. DSRpu in different discourse genres in children with DLD.

Figure 6 .
Figure 6.DTRwad in different discourse genres in children with DLD.

Figure 7 .
Figure 7. Regression of the MLU on the DSRpu in TD children.

Figure 8 .
Figure 8. Regression of MLU on the DSRpu in children with DLD.

Table 2 .
Descriptive statistics of the data from the two groups of children.

Table 3 .
Impact of Group, Genre, and Disfluency type on disfluency rate per utterance.
Note: p was estimated by Wald Chi-Square Statistics.

Table 4 .
DSRpu rate in different discourse genres.