Anxious Words: How Learners’ Speech Reveals Fear in Online and In-Person English Classes

For millions of adults learning English around the world, the moment of speaking can feel like stepping onto a stage without a script. The heart races, the mind blanks, and sentences that seemed effortless in private suddenly fall apart in public. Psychologists call this foreign language speaking anxiety, and it has been studied for decades as one of the most stubborn obstacles to fluency. Yet a new study from researchers at Yıldız Technical University in Istanbul suggests that the fingerprints of this anxiety may be hiding in plain sight, embedded in the small connective words that pepper everyday speech: and, but, well, because, you know. By tracking how learners deploy these discourse markers in online versus face-to-face classrooms, the research offers a fresh window into how emotion shapes language, and it delivers a surprising verdict on whether the pandemic-era shift to virtual learning actually made speaking easier.

The study, published in Discover Education, was led by Gulumser Efeoglu and Büşra Yurdagül, who set out to answer a question that has grown urgent since online language classes became a permanent fixture of education. Previous research on anxiety in virtual settings has been strikingly divided. Some studies found that learners felt less isolated, less anxious, and more connected to teachers and peers when learning online, while others warned that technology itself, from mandatory cameras to unfamiliar platforms, could be anxiety-inducing. Rather than asking learners simply how nervous they felt, the Turkish team decided to look at what those learners actually said, treating the microscopic architecture of speech as a behavioral record of emotional state.

To do this, the researchers recruited 80 adult native Turkish speakers learning English, all of whom tested at the B1 intermediate level on the Oxford Placement Test, a step taken to ensure that differences in vocabulary knowledge would not contaminate the results. Half of the participants were studying in face-to-face settings, including a university School of Foreign Languages and a private language institute, while the other half attended online classes at the same private institute, which used platforms such as Zoom and Microsoft Teams. Participants ranged in age from 18 to 55, and each completed a battery of instruments: an adapted 18-item version of the Foreign Language Classroom Anxiety Scale, a standardized proficiency test, and a semi-structured interview conducted in Turkish to encourage candid self-expression.

The centerpiece of the methodology was a task drawn from the Language Situations method, designed to elicit naturalistic speech under controlled conditions. Participants watched a video depicting a fictional event, such as a traffic accident, and then narrated what they had seen as if reporting it to a police officer. This formal framing simulated an authentic communicative scenario while keeping the task identical across both groups. Every response was audio- or screen-recorded and transcribed verbatim, yielding a corpus of 6,790 seconds of speech, more than 113 minutes in total, and 7,029 fully uttered words. The face-to-face group contributed 3,101 seconds and 3,059 words, while the online group produced 3,689 seconds and 3,970 words, a difference that would later prove analytically important.

The researchers then turned to corpus analysis software called AntConc to hunt for discourse markers, the small words and phrases that speakers use to organize, connect, and cushion their utterances. Following a widely used four-part taxonomy, the team classified markers as structural, which organize discourse units, as in and, then, or first; referential, which link ideas together, as in because; cognitive, which signal mental processes and reformulations, as in well; and interpersonal, which convey viewpoint and shared knowledge, as in you know. Ambiguous words demanded careful adjudication. When a participant said it was just like so tense, the word so functioned as an ordinary intensifier, but in if the first car didn’t, like, stop, the word like operated as a genuine discourse marker signaling hesitation. In total, the analysis identified 660 discourse marker tokens across the two groups.

The first major finding was, in some ways, a non-finding. A Mann-Whitney U test comparing anxiety scores between the two groups revealed no statistically significant difference, with a negligible effect size. Learners in physical classrooms and learners behind screens reported roughly the same levels of speaking tension. This result runs counter to a body of earlier work suggesting that online environments inherently lower the affective barriers to speaking, and it complicates the popular assumption that virtual classrooms are a safe harbor for the socially anxious. Whatever protection the online medium may offer, it did not translate into measurably lower self-reported anxiety in this sample.

But the discourse marker data told a far more nuanced story. In the face-to-face group, two categories of markers showed a moderate negative correlation with anxiety: referential markers such as because, with a correlation coefficient of -0.343, and cognitive markers such as well, with a coefficient of -0.454. In plain terms, the more anxious a learner was in a physical classroom, the fewer of these idea-linking and self-monitoring words they produced. Interpersonal and structural markers, by contrast, were unaffected by anxiety levels in that setting. Strikingly, none of the four marker categories showed any significant correlation with anxiety in the online group, suggesting that the emotional signature of speech differs fundamentally depending on where the speaking takes place. Meanwhile, the online group used more markers overall, 384 tokens versus 276, though from a typologically narrower set, with and, then, I think, but, and okay topping the frequency list.

Fluency measures added another layer of contrast. When fluency was operationally defined as the relationship between total speaking time and total word count, the face-to-face group showed a stronger positive correlation, r = 0.819, than the online group, r = 0.584. This pattern indicates that learners in physical classrooms produced more continuous and verbose speech, sustaining their output over time more consistently. The researchers speculate that technical frictions inherent to online communication, such as connection problems and the absence of rich verbal and nonverbal cues, may slow the pace of virtual speech, even when anxiety itself is no higher. The interviews reinforced the emotional dimension of these patterns: across both groups, learners attributed their anxiety to personal traits such as limited vocabulary and low self-confidence, to social pressures including fear of being laughed at for mispronunciation and judgmental teacher feedback, and to cultural uncertainties about how their words would be perceived.

The practical implications are considerable. Because discourse marker use appears to track anxiety in face-to-face settings but not online, teachers could potentially use these small words as diagnostic tools, listening for a thinning of referential and cognitive markers as an early warning sign that a student is struggling emotionally. Conversely, explicit instruction in discourse markers might itself serve as a form of anxiety management, giving learners concrete linguistic resources for holding the floor and bridging hesitations. The authors suggest tailoring tasks to each context: using face-to-face classes to stimulate fluency, and online classes for explicit discourse marker teaching. They also recommend affective strategies such as mindfulness, which other recent work has linked to reduced speaking anxiety, alongside classroom cultures that frame pronunciation errors as a normal part of linguistic development rather than a source of shame.

The study’s authors are careful to note its limitations. The sample of 80 participants, while adequate for the statistical tests employed, was recruited through convenience sampling, and the formal police-report task may have constrained spontaneous marker use differently across the two settings. Future research, they suggest, should examine synchronous and asynchronous online formats, varied proficiency levels, and wider age ranges to test whether these patterns hold. Even so, the core insight stands as a compelling piece of the growing science of language and emotion: anxiety does not simply silence learners, it reshapes the fine-grained texture of what they say. The words we dismiss as verbal filler may, in fact, be among the most honest signals of how a speaker feels, and learning environments that understand this could help anxious speakers find not just their voice, but the connective tissue that keeps it flowing.

Subject of Research: Foreign language speaking anxiety and discourse marker use in online versus face-to-face English classes

Article Title: Foreign language speaking anxiety and discourse markers in face-to-face and online classes

Article References: Efeoglu, G., & Yurdagül, B. (2026). Foreign language speaking anxiety and discourse markers in face-to-face and online classes. Discover Education, 5(1), Article 1066. https://doi.org/10.1007/s44217-026-02220-3

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02220-3

Keywords: foreign language speaking anxiety, discourse markers, online learning, face-to-face classes, EFL, fluency, second language acquisition, language education, psycholinguistics, FLCAS, corpus analysis, Turkish learners