Speech fluency is a temporal feature of speech that can be defined both in a broad and a narrow sense (Lennon, 1990). In the broad sense, fluency refers to “spoken command of a foreign language” (Lennon, 1990, p. 389). In a narrow sense, it is used in research settings and L2 proficiency testing to refer to a specific aspect of oral proficiency that refers to the speed, smoothness and effortlessness of L2 speech (Chambers, 1997; Lennon, 1990). The present study concerns the narrow sense of fluency.

To understand how L2 fluency is achieved, it is necessary to look at the speech production process. Levelt (1989) or Levelt et al. (1999) distinguishes four main processes in his L1 speech production model: conceptualizing, formulating, articulating, and self-monitoring. In L1 (or highly proficient L2) speech, formulation (including lexical retrieval and grammatical encoding) and articulation are mainly automatic and allow for parallel processing with other modules (i.e., occurring virtually simultaneously and requiring little cognitive effort). Conceptualization (utterance planning) and monitoring processes require attention. These L1 speech production features explain “speed and fluidity, that is, a speaker’s fluency” (Tavakoli & Wright, 2020, p. 9).

For less proficient L2 speakers, not only conceptualization and monitoring, but also lexical retrieval, grammatical encoding and articulation might require conscious attention (de Bot, 1992; Kormos, 2006). The process will then occur less automatically, in a consecutive sequence (serial processing) and hence more slowly and effortfully. When L2 learners have access to a larger repertoire of lexical and grammatical knowledge and progress to a higher proficiency level, their lexical access and grammatical encoding becomes more automatic, and they therefore produce output of higher fluency. The fluency of L2 performance thus reflects the efficient functioning of speech production processes, namely conceptualization, formulation, articulation and monitoring (Kormos, 2006). This aspect of fluency is referred to as cognitive fluency (Segalowitz, 2010).

The degree of cognitive fluency is reflected in speech in the form of so called utterance fluency (Segalowitz, 2010), which can be further divided into speed fluency, or the number of words or (pruned²) syllables that are produced; breakdown fluency, or the number and length of pauses and repair fluency, which reflects hesitations and repairs (Lambert et al., 2017; Saito et al., 2018; Pérez Castillejo, 2019). Speed fluency reflects the degree of automatization of linguistic knowledge and the ease of accessing it (Pérez Castillejo, 2019, 2021; Bielak, 2022), which refer to the formulation/encoding (Levelt, 1989) aspect of cognitive fluency. Regarding breakdown fluency, pauses within clauses, which can reflect processing difficulties (Kahng, 2014), may be indicative of formulation /encoding difficulties/processes. Pauses between clauses may be indicative of content planning and conceptualization (Götz, 2013). Repair fluency is assumed to reflect the effectiveness of the monitoring system (Kormos, 1999).

In summary, while L1 speech processing is more automatized and stable, L2 speech processing, especially at lower proficiency levels, usually requires more control or conscious attention. Besides the degree of automatization, fluency can be affected by FLA, which will be discussed in the next section.

Over time, attempts have been made to classify the FLA construct as trait anxiety (i.e., a permanent likelihood of experiencing anxiety regardless of the situation); state anxiety (which may arise in response to a situation at a particular time); or situation-specific anxiety (i.e., a personal predisposition to experience anxiety in particular situations, Baran-Lucarz, 2022). Since a pioneering article by Horwitz et al. (1986), FLA has been treated as a situation-specific type of anxiety (MacIntyre & Gardner, 1989). Recently, Horwitz (2017) characterized a person experiencing FLA as “having the trait of feeling state anxiety when participating in (or sometimes even thinking about) language learning and/or use” (p. 33). Dewaele (2007, pp. 405–406) explains that FLA is “probably situated halfway between trait, situation-specific anxiety and state, more sensitive to environmental factors than personality traits and yet more stable than states since it remains relatively stable across languages.”

One explanation for the idea that FLA may influence the speech production process (i.e., more specifically fluency) comes from the Processing Efficiency Theory by Eysenck and Calvo (1992). They suggest that anxiety impairs the efficiency of cognitive processing during L2 speech production by posing additional demands on working memory resources and attentional processes.

The current study investigates the relationship between FLA and utterance fluency. Studies examining this link are limited. Sanaei et al. (2015) found strong negative correlations between FLA and speed fluency (i.e. more anxious participants had shorter runs between pauses and shorter phonation time) during a non-exam situation. Pérez Castillejo’s (2019) examined how FLA and proficiency relate to L2 fluency during a final oral exam in learners with low-intermediate proficiency level. The results indicated medium and strong negative relationships between FLA and those fluency measures that reflect speech formulation/encoding (i.e. more anxious participants had a smaller phonation time, shorter runs between pauses and paused more frequently within clauses). The negative relationship between FLA and the length of pauses between clauses, which is mainly related to message conceptualization, was weaker. Regression analyses showed that FLA was a stronger predictor of fluency than proficiency. Whereas Pérez Castillejo (2021) replicated Pérez Castillejo’s (2019) results regarding the relationship between FLA and the fluency of speech formulation/encoding (i.e., mean length of run and number of mid-clause pauses), their results regarding the relationship between FLA and conceptualization fluency (the length of pauses) were not replicated, indicating that there is more evidence that formulation/encoding is affected by anxiety than conceptualization is. Pérez Castillejo (2021) also found that repair fluency was very weakly related to FLA and proficiency, probably due to the low proficiency level of participants. Proficiency was however a much stronger predictor of L2 fluency than FLA when fluency was measured in a task following another similar task, that is, in a condition of prior L2 processing (understood as L2 use earlier in discourse).

In addition to anxiety, cognitive fluency may also be influenced by task complexity. According to Robinson’s Cognition Hypothesis (2001), for instance, cognitively demanding tasks focus L2 speakers’ attention on language form, to increase the grammatical accuracy and linguistic complexity of their L2 production at the cost of fluency (Michel et al., 2007; Robinson, 2001). The Limited Attentional Capacity Model (Skehan & Foster, 2001) on the other hand suggests that complex tasks may negatively affect both grammatical accuracy and fluency. Skehan and Foster (2001) argue that learners focus more on meaning than on form, and more so in cognitively demanding tasks. Both Robinson (2001) and Skehan and Foster (2001) thus predict that task complexity would negatively impact fluency. De Jong et al. (2012) examined how task complexity affected L2 speaking performance. Results showed that complex tasks led to more pausing and repair behavior than simple tasks, but no significant difference in speed fluency was found. Regarding lexical complexity, L2 speakers produced a wider range of words in complex tasks compared to simple tasks.

The effect of task complexity may vary depending on the proficiency and anxiety level of the speaker. Since the automaticity of linguistic encoding is related to proficiency (Kormos, 2006; Segalowitz, 2010), more proficient L2 speakers may have more attentional resources available to conceptualize the message than less-proficient speakers, who must divide their attention between conceptualization, formulation and monitoring (Lambert et al., 2017). Additionally, the effect of task complexity on speech processing might vary depending on the anxiety level of the speaker completing the task. According to Eysenck and Calvo (1992), the more cognitively complex the task, the stronger the effect of anxiety on task performance. Robinson’s (2001) Cognition Hypothesis also states that individual learner differences, such as anxiety, have a greater influence on spoken task performance if tasks are more cognitively complex. To date, only Bielak (2022) examined the relationship between FLA and L2 fluency in advanced L2 learners. Participants completed both a group task and a monologue (i.e., complex task). Numerous negative correlations between FLA and fluency measures, reflecting formulation (i.e., more anxious participants had shorter runs and lower articulation rate), and conceptualization (more end-pauses), were found in the group task, and only one negative correlation (more end-pauses) was found during the complex task. Regression analyses showed that proficiency was a stronger predictor of fluency than FLA. This could be attributed to the advanced proficiency level of the participants and consequently more automatized L2 use. A limitation of this study was that FLA was measured twice, after the group and complex tasks, but fluency was measured only once, during the complex task; therefore, it is not certain whether the correlations between the first anxiety measurement and fluency are reliable.

Based on these previous studies (Sanaei et al., 2015; Pérez Castillejo’s, 2019, 2021) one might expect that anxiety affects attentional processes, which in turn affects fluency in L2. Furthermore, based on Robinson (2001), Eysenck and Calvo (1992) and Bielak (2022) it is expected that anxiety will negatively affect speech fluency during a cognitively demanding task.

The present study investigates the effect of anxiety on speed, breakdown and repair fluency in two types of tasks, one simple and one complex, among low-intermediate L2 learners. On the basis of previous research, we hypothesize that a cognitively more demanding task negatively affects speed fluency of anxious learners. With respect to breakdown fluency, we expect that complex tasks elicit more pausing behavior. Considering repair fluency, we hypothesize that a cognitively demanding task will lead to fewer self-corrections and reformulations by anxious learners. The idea underlying this hypothesis is that at lower proficiency levels, when L2 knowledge is not automatized, learners’ attentional capacity can be strained too much by other processes, particularly formulation (Kormos, 2000; Pérez Castillejo, 2021), so that there is no room left to make corrections.

Our research question is therefore as follows: “To what extent is there a relationship between speed, breakdown and repair fluency measures and FLA and proficiency, and how does this interact with task complexity?”.

The participants were 61 German learners of Dutch (20 male and 41 female) from three secondary schools, all of whom were in their fifth year of Dutch. The L2 learners reported German as their L1. Six participants had two L1s: three German/Niederdeutsch, one German/Russian, one German/Kurdish and one German/Ukrainian. Their ages ranged from 14 to 16 with an average age of 15.2 years. All students participated in this study voluntarily with signed informed consent.

The original FLCAS questionnaire (Horwitz et al., 1986) consists of 33 items with a 5-point Likert scale concerning learners’ feelings and behavior in the foreign language classroom, and in conversation with native speakers. This instrument has been validated and widely used in language learning research.

Mak (2011) found that the FLCAS tests five subconstructs of anxiety: (a) speech anxiety and fear of negative evaluation, (b) uncomfortableness when speaking with native speakers, (c) negative attitudes toward the L2 class, (d) negative self-evaluation, and (e) fear of failing the class or the consequences of personal failure. Considering our research question only three of these subconstructs with in total 21 items were of relevance and used in our study: the first two (a) and (b), as well as (d). The “negative attitudes toward the Dutch class” subconstruct was not selected because the questionnaire was completed in the first weeks of the school year and hence, participants’ responses were less likely to be influenced by their attitudes toward the class. The “fear of failing the class or the consequences of personal failure” subconstruct also was not relevant because the speech production tasks did not constitute an exam and not related to the class in that sense.

A native German speaker with a Master’s degree in English translated the original English version into German. A professional translator subsequently back-translated the questionnaire and confirmed that the English and German versions were completely comparable. The questionnaire was then piloted with participants not involved in the study to test whether the items were clearly formulated. Feedback showed that one item had a translation error which was subsequently corrected. Participants completed the questionnaire in approximately five minutes. The reliability of the questionnaire as measured by Cronbach’s alpha was 0.89, which indicates a high level of internal consistency.

The participants completed the Dutch version of the LexTALE (Lemhöfer & Broersma, 2012), a five-minute, 60-item vocabulary test, in which they must indicate whether or not a string of letters represents an existing word in the language. The test measured participants’ general proficiency in Dutch. A benefit of the LexTALE is that the ranges of scores are associated with the proficiency levels as described in the Common European Framework of Reference (CEFR) (Council of Europe, 2001; see Lemhöfer and Broersma (2012) for the correspondence between LexTALE and CEFR).³ According to Lemhöfer and Broersma (2012), LexTALE scores of 80–100 % represent CEFR level C1–C2, 60–80 % level B2, and below 59 % level B1 and lower. Cronbach’s alpha assessed the LexTALE’s reliability and was 0.71, which indicates an acceptable level of internal consistency.

Speech data was collected using two tasks that differed in complexity. To operationalize complexity, we followed De Jong et al. (2012). In their opinion, complex tasks contain more elements than simple tasks; complex tasks concern a more general topic as opposed to simple tasks, which concern topics of personal life; and complex tasks involve more abstract notions as opposed to simple tasks, which involve mostly concrete notions. In our simple task, the participants discussed their weekend activities. This is a concrete topic of personal life, and the task consists of one element: “What did you do last weekend?”. In our complex task, participants were asked to persuade the class of their opinions and arguments about smartphones in the classroom. This is a more general topic; it is more abstract and consists of two elements (providing their opinion and arguments; see Appendix A). The regular class teachers confirmed that the tasks were appropriate for the participants’ Dutch level. Moreover, 14 German students of Dutch piloted the tasks. Their feedback showed that the instructions were clear and that the more cognitively demanding task was perceived as more complex than the simple task.

The participants completed the FLCAS questionnaire, as well as the demographic and language background questionnaire and the proficiency test some weeks before the speech production tasks. We measured anxiety independently of the speaking tasks, to be better able to compare our data to similar studies (Sanaei et al., 2015; Pérez Castillejo, 2019).

The data from the oral tasks were collected on regular school days where the participants met individually with the researcher whom they had already met during the first session (completion of questionnaires and proficiency test). The participants were given the simple and complex tasks on paper and, after checking for comprehension, were instructed to answer without time to prepare. They were asked to speak for two minutes per task and told that their performances were audio recorded. In both tasks, participants spoke only to the researcher, not to a classmate. The role of the researcher during the performance was limited to nodding to show understanding. The tasks had not been practiced before. For the sake of comparability, the tasks were always taken in the same sequence; first the simple task and then the complex task. All speech samples were recorded on an Olympus LS 10 digital voice recorder.

In the simple task, participants spoke on average for 58.00 seconds (SD = 25.41) from the first to the last syllable, and in the complex task for 54.33 seconds (SD = 22.39). All speech recordings of the two tasks of the 59 participants (roughly two hours of speech in total) were transcribed in detail, including information regarding pauses and repairs. Two of the 61 participants were absent when the speech data was collected. Three datasets were discarded due to technical problems. One recording lasted five seconds: we excluded it from the analyses, as fluency measures are unstable for short speech samples. This resulted in 56 datasets from the simple task and 58 from the complex task. These 114 recordings were transcribed into Analysis of Speech (AS) units following the procedures of Foster et al. (2000). The AS unit is defined as “a single speaker’s utterance consisting of an independent clause, or sub-clausal unit, together with any subordinate clause(s) associated with either” (Foster et al., 2000, p. 365).

As a first step, the “textgrid-to-silences” function in PRAAT (Boersma & Weenink, 2015) was used to automatically mark silent pauses. Since the background noise was occasionally disturbing automatic detection of silent pauses, the identified pauses were manually checked and adjusted as necessary. Only silent pauses (i.e., greater than or equal to 0.25 seconds) and filled pauses were marked as pauses. Short “micropauses” (Riggenbach, 1991) that are irrelevant for measures of L2 fluency were excluded (De Jong & Bosker, 2013). All filled pauses were manually marked.

For each pause it was determined whether it was mid-clause or end-clause through careful listening to the recordings and examination of the transcript, which had been marked with clause boundaries. Mid-clause pauses were filled and silent pauses occurring in the middle of clauses. Pauses between clauses and when a conjunction occurred between clauses were considered to be end-clause pauses.

For each speech recording, syllables were counted manually. Syllables were counted for the purpose of calculating speech rate, articulation rate and mean length of run (i.e., stretches of speech uninterrupted by pauses > 0.25 seconds). The number of self-corrections and reformulations during each performance was also counted manually.

Regarding fluency measures, it has been argued (Bosker et al., 2013; Huensch & Tracy-Ventura, 2016; Hunter, 2017; Skehan, 2015, Tavakoli et al., 2020) that composite and pure fluency measures need to be distinguished. Speech rate, mean length of run and phonation time ratio are described as composite measures because they combine two or more of the aspects of fluency. A measure such as mean length of run combines speed and breakdown fluency. In contrast, pure measures relate to one of the three aspects of fluency: speed, breakdown, or repair. For example, articulation rate relates to speed fluency. Pure measures have been argued to tell more about the underlying processes involved in speech production (Huensch and Tracy-Ventura, 2017; i.e., linking specific fluency measures to stages of speech production; Lambert et al., 2017).

The present study used several different (composite and pure) fluency measures (see Table 1). Sanaei et al. (2015), Pérez Castillejo (2019, 2021) and Bielak (2022) included a limited number of fluency measures. To capture a more complete picture of how anxiety relates to speed, breakdown and repair phenomena in a simple and complex task, we used a broad range of fluency measures (see Table 1 for details on their operationalization). For comparison purposes fluency measures from FLA and fluency studies were used, namely: speech rate, articulation rate; phonation time ratio; mean length of run (Sanaei et al., 2015); number of mid-clause and end-clause pauses; mean length of mid-clause; mean length of end-clause pauses (Pérez Castillejo, 2019; Bielak, 2022) and number of reformulations and self-corrections (Pérez Castillejo, 2021). In addition, we used other commonly used speed, breakdown and repair fluency measures (i.e., number of silent and filled pauses per minute, number of pauses per minute, and mean length of pauses) based on earlier fluency studies (See Kahng, 2022).

Data for six participants (i.e., 12 recordings; 10.53 % of the data) was re-examined for syllable count by an independently trained teacher/researcher. Additionally, the rater analyzed and coded 10 % of the data again for pause type, pause length, pause location, self-corrections and reformulations. Interrater reliability was 98 % agreement on syllable count, 98 % on pause type, 90 % on pause length, 92 % on pause location and 92 % on self-corrections and reformulations, demonstrating high interrater reliability.

As a first step, we checked the data (fluency measures, anxiety, and proficiency) for normality. The results of Shapiro-Wilk test indicated that the fluency measures: mean length of run, number of filled pauses, mean length of pauses, mean length of mid and end-clause pauses, and reformulations were not normally distributed (p < 0.05).

To explore the research question, a correlation analysis was run with the L2 learners’ data on speed, pause phenomena, repair, anxiety and proficiency scores. Due to non-normality of some fluency measures, we decided to use Spearman rank-order correlation test. Based on this exploration, we decided to run regression analyses. Prior to the regression analyses, the assumptions of LMM were checked, including linearity, absence of collinearity, homoscedasticity, and normality and no violations were noted. Then a mixed-effects analysis was conducted, using the lme4 package (Bates et al., 2015) in R Version 4.0.2 (R Core Team, 2020). First, we tested interactions, for every fluency measure that correlated with anxiety and/or proficiency (i.e. speech rate, number of mid-clause pauses, number of pauses, number of silent pauses and phonation time ratio) and center the predictors involved in interactions (i.e., Anxiety, Proficiency and Task type). However, for all five fluency measures the models turned out to be non-significant and did not improve the model fit. We also added random slopes with interactions, but we received a warning that the data was not large enough to estimate random effects. Finally, for each fluency variable, we built a model with Anxiety, Proficiency and Task type as fixed effects and Subject as a random effect and then added School as a fixed effect. The fixed-effect factor, Task type, was added to control for differences between performing the simple and complex tasks. School was added as a fixed effect to control for possible differences between the three schools that participated in the study. The variable Subject was added as a random effect to control for differences between participants. The models were subsequently compared; in essence, we tested whether to include the fixed effect School or not, and if the first model was preferable compared to second one. Model comparisons were assessed using the ANOVA (Model 1, Model 2) function in R.

The correlation analysis results that served as a first exploration of the data showed that more anxious speakers used significantly more mid-clause pauses in the complex task, but with a small effect size. The results generally support the theoretical claims of the Cognition Hypothesis (Robinson, 2001) and Processing Efficiency Theory (Eysenck & Calvo, 1992) regarding the negative role of anxiety in L2 speech processing, as well as previous research (Pérez Castillejo, 2019, 2021). In contrast, Bielak (2022) found no significant correlations between FLA and mid-clause pauses in a cognitively demanding task among advanced L2 learners. Mid-clause pauses reflect cognitive efficiency during formulation. Especially lexical retrieval and syntactic encoding (Kahng, 2014) may require attention. The participants in the present study had generally lower proficiency, which suggests that their ability to formulate may require greater attention control (Kormos, 2006). Thus, any interference that limits attention control during L2 speech production, such as FLA, may impact pausing behavior within clauses (Pérez Castillejo, 2021).

There were positive correlations between proficiency and fluency (i.e. speech rate and phonation time ratio) and negative correlations (i.e. number of silent and mid-clause pauses and number of pauses) in the simple task. These speed and pausing phenomena show how effectively L2 knowledge can be accessed and are linked to message formulation and encoding (Kahng, 2014). In contrast, fluency measures that may be associated with message conceptualization such as number of end-pauses were not related to proficiency in a statistically significant way. For L2 speakers, especially at lower proficiency levels, formulation and encoding may require more attention control than conceptualization (De Bot, 1992; Kormos, 2006; Segalowitz, 2010). This could explain why proficiency relates differently to the fluency measures associated with each of these processes. Note that the task we used, LexTALE has been validated for English, but not yet for Dutch, so especially with respect to assigning scores to the European Reference Framework, we cannot form any strong conclusions. The scores on the task can be used to correlate individual variation in proficiency though, which we have done here.

Furthermore, repair fluency was not significantly related to FLA or proficiency. This is in line with the findings of Pérez Castillejo (2021). Self-repairs may reflect that attentional resources are being used for monitoring (Kormos, 2000), but at lower proficiency levels, when L2 knowledge is not automatized, attention may be needed for other processes (particularly formulation), and hence FLA or proficiency may not significantly affect monitoring.

We additionally investigated the contributions of the predictors (i.e., FLA, proficiency and task type) on the fluency measures. First, proficiency was a significant predictor of two fluency measures, whereas FLA was not a significant predictor of any fluency measure. In Pérez Castillejo’s study (2019) and its replication (2021) on the relationship between FLA, proficiency and fluency during a final oral exam, the roles seem reversed: FLA was a significant predictor of the fluency measures analyzed and proficiency was not. Explanations for the contradictory results could be that in the present study, the tasks were not performed during an exam situation, and the participants’ anxiety levels were lower compared to the levels measured in previous research (Bielak, 2022; Pérez Castillejo, 2019, 2021; Saito et al., 2018).

Recall that the complex task resulted in a faster speech rate and articulation rate, longer runs, more phonation time, more and longer mid-clause pauses, and fewer and shorter end-clause pauses than the simple task. We found a significant positive effect of the predictor task type on speech rate. The more complex the task, the higher the speech rate. It is somewhat surprising that a cognitively demanding task results in a higher speech rate, as the Cognition Hypothesis (Robinson, 2001) and Limited Attentional Capacity Model (Skehan & Foster, 2001) predict that these tasks may negatively affect fluency. One possible explanation is that the complex task may not have been sufficiently complex, thus not resulting in more formulation problems. Another explanation is that participants had already used the L2 in the simple task and benefited from it while performing the complex task. According to McDonough and Trofimovich (2008), earlier activation of language material facilitates later access to and use of related linguistic knowledge, which can lead to improved performance. This is also what Pérez Castillejo (2021) and Bielak (2022) found in their research.

While this study is one of the first to address a broad range of fluency measures and address the influence of FLA, proficiency and task complexity, it also has some limitations. First, the participants produced relatively short speaking samples. The participants were instructed to speak for two minutes for each task, but on average they did not speak for more than one minute. Another limitation is that the tasks were always taken in the same sequence. This should have been varied to avoid the effect of novelty (better performance) in the first task and maybe the effect of repetitiveness (worse performance) in the second task. Future research might compare whether it makes a difference to vary the sequence of the simple and complex tasks. Finally, in the absence of L1 data from the participants, it is difficult to claim that the results obtained here are only due to the speakers’ L2 fluency behavior and not affected by individual differences in their L1 speaking style (De Jong et al., 2015).

As far as pedagogical implications are concerned, the results of this study suggest that FLA may not have a strong influence during speaking tasks that are not part of formal assessments. This could encourage teachers to base grades for speaking proficiency on classroom observations as well, rather than only on formal tests (Bielak, 2022).