2.1 Foreign language anxiety

Decades of research on FLA bear witness to the importance of affective variables in language learning. The term anxiety in psychology is defined as the “subjective feeling of tension, apprehension, nervousness and worry associated with an arousal of the autonomic nervous system” (Spielberger, Reference Spielberger1983: 15). Spielberger (Reference Spielberger1983) distinguishes between “trait anxiety”, which is defined as the individual’s likelihood of becoming anxious in any situation, and “state anxiety”, which is experienced by an individual as a reaction to a specific situation. In language learning, however, a third type of anxiety is defined: “a distinct complex of self-perceptions, beliefs, feelings, and behaviours related to classroom language learning arising from the uniqueness of the language learning process” (Horwitz et al., Reference Horwitz, Horwitz and Cope1986: 128). Several instruments have been designed to measure FLA: Attitudes and Motivation Test Battery (Gardner, Clément, Smythe & Smythe, Reference Gardner, Clément, Smythe and Smythe1979) and Language Class Discomfort (Ely, Reference Ely1986). Furthermore, MacIntyre and Gardner’s (Reference MacIntyre and Gardner1994) Input, Processing, and Output Anxiety scales aimed to measure anxiety at different phases of language learning. The most widely used instrument is Horwitz et al.’s (Reference Horwitz, Horwitz and Cope1986) Foreign Language Classroom Anxiety Scale (FLCAS); it has been used to investigate the reliability and validity of the measure, to successfully test various psychometric properties in the classroom, and to assess the effect of FLA on acquisition and performance. Most studies discussing FLA report on its negative effects on language learners in terms of the learning process (MacIntyre & Gardner, Reference MacIntyre and Gardner1994; Steinberg & Horwitz, Reference Steinberg and Horwitz1986) and assessment results (Zhang, Reference Zhang2013; Zhao, Guo & Dynia, Reference Zhao, Guo and Dynia2013, just to mention a few).

Research on FLA has investigated its role in the development of different skills: reading (Lee, Reference Lee and Young1999; Zhao et al., Reference Zhao, Guo and Dynia2013), writing (Abdel Latif, Reference Abdel Latif2015; Leki, Reference Leki and Young1999), listening (Elkhafaifi, Reference Elkhafaifi2005), and speaking (Çağatay, Reference Çağatay2015; Koch & Terrell, Reference Koch, Terrell, Horwitz and Young1991; Tsiplakides & Keramida, Reference Tsiplakides and Keramida2009). Any of these skills may trigger FLA. However, reading seems to be the least anxiety-provoking skill, whereas speaking is the most anxiety-provoking activity (Young, Reference Young1992). Factor analysis of FLCAS reveals that the most significant triggers of FLA are fear of making mistakes/fear of inadequate performance, fear of negative evaluation, general speech anxiety/fear of speaking spontaneously/understanding, and communication apprehension (Aida, Reference Aida1994; Mak, Reference Mak2011; Park, Reference Park2014; Tóth, Reference Tóth2008; Young, Reference Young1990).

Language practitioners for decades have been looking for ways to create a low-anxiety classroom where students can thrive. There is ample evidence that CMC tools decrease FLA levels when interactions occur in a written form via emails or written chats (Kissau, McCullough & Pyke, Reference Kissau, McCullough and Pyke2010; Satar & Özdener, Reference Satar and Özdener2008). Fewer studies demonstrate that CMC activities are perceived by students as low-anxiety environments even when speaking (e.g. Cooke-Plagwitz, Reference Cooke-Plagwitz2008). Rosell-Aguilar (Reference Rosell-Aguilar2005: 432) says, based on “anecdotal evidence”, that students of beginner Spanish might have chosen online audio-graphic tutorials over F2F ones because “they may feel more comfortable ‘shielded’ behind a PC and not making mistakes in the physical presence of others”. Melchor-Couto (Reference Melchor-Couto2017) examined the evolution of FLA levels in a virtual world (Second Life) tandem learning environment between native speakers and language learners. The author showed that FLA levels were lower, and gradually decreased, in the virtual world group than in the control group. She argues that anonymity helps to reduce anxiety at the beginning of the course, and although it wears off as participants get to know each other better, it is beneficial for instigating oral interaction as it offers a gentle start. Anonymity gives learners more freedom to make mistakes, which reduces FLSA. Other researchers, however, show that the lack of body language and visual information may in itself cause anxiety, which can be aggravated by anxiety related to the use of technology (Hampel, Felix, Hauck & Coleman, Reference Hampel, Felix, Hauck and Coleman2005). De los Arcos, Coleman and Hampel (Reference de los Arcos, Coleman and Hampel2009) conclude that actually both a facilitating and a hindering effect are possible in CMC contexts.

2.2 Asynchronous speaking

There are few studies focusing on the affordances of asynchronous tools in relation to FLA and more specifically speaking anxiety, despite the fact that they are generally believed to promote oral skills and help learners, especially beginners, to prepare for F2F communication. Tecedor and Campos-Dintrans (Reference Tecedor and Campos-Dintrans2019: 117) cite two publishing houses in support of this: “Pearson Education advertised in its platform: ‘Oral Recording Activities enable students to increase oral proficiency without leaving the online environment.’ Vista Higher Learning similarly stated: ‘Language production can be intimidating for new students, but this initial step helps build confidence and better prepare students for face-to-face communication.’” A number of studies show that voice-based tools can indeed have a positive impact on second language (L2) speaking development, especially fluency (Sun, Reference Sun2012; Yanguas, Reference Yanguas2010). As asynchronous tools (audio recordings, video blogs) are one-way communication without response or interaction in real time, they allow planning time and often multiple recordings. Information is transmitted in an intermittent manner. Thus, it is a genre that typically elicits presentational language, more complex syntax, but might hinder spontaneity. Tecedor and Campos-Dintrans (Reference Tecedor and Campos-Dintrans2019) explicitly compare the gains of synchronous and asynchronous modes of learning with regard to the improvement of oral proficiency in two tasks: presentational and interpersonal. After examining complexity, accuracy, and fluency, the authors conclude that different modes of learning have different gains in the two tasks, so the medium has important implications for learning outcomes. The study, however, does not deal with affective factors.

In a study on the use of Wimba Voice Board, a tool that allows users to create and post audio recordings directly in Moodle, McIntosh, Braul and Chao (Reference McIntosh, Braul and Chao2003) observed that using the medium was a positive experience for the majority of the students on an English for academic purposes university course, which primarily consisted of F2F classes. The online environment encouraged the development of communicative competency in a non-threatening setting. The medium was also a valuable tool for self-analysis: 17% of the students recorded their voices between four to six times before uploading their message. More than half of the participants felt more confident speaking in class after participating in asynchronous discussions. Some students, however, were not comfortable posting their voices online and overcoming this barrier remained a challenge.

Podcasts are also reported to reduce FLSA. Sze (Reference Sze2006: 122) claims that their use is beneficial for “less confident students as producing podcasts involves performing behind the scenes”. Hamzaoğlu and Koçoğlu (Reference Hamzaoğlu and Koçoğlu2016), in a study with 30 highly anxious students, showed that by the end of the term students who produced podcasts significantly improved in their oral performance and had significantly lower FLSA levels compared to students in the control group.

Poza’s (Reference Poza2011) findings are not so straightforward. She describes an in-depth study on students’ FLSA levels when using the Wimba voice conferencing environment. The author found that the average levels of anxiety in the classroom and on Wimba were very similar (no statistically significant difference). Students were especially worried about speaking to native speakers in both contexts; they were anxious about making mistakes. Despite the lack of statistically significant differences in the two contexts, the author states that there is a strong potential for the reduction of anxiety within the asynchronous context, as learners in this context were less worried about the teacher correcting their mistakes and others being better. The asynchronous context coupled with anonymity on an LMOOC should be a language learning context with reduced levels of FLSA.

A limited number of studies deal with the acquisition of speaking skills on an LMOOC. Rubio (Reference Rubio, Martín-Monje and Bárcena2014) compares the gains in comprehensibility of students enrolled in a traditional F2F pronunciation course with those enrolled in an LMOOC. He concludes that there was significant improvement in both course formats, but larger effect sizes on the LMOOC. However, LMOOC students in his study received a larger quantity and variety of individualised feedback than students in the F2F group, which makes it difficult to compare the two groups. Furthermore, such individualised feedback is rarely possible on LMOOCs. The focus of Rubio’s study is not on the affective factors related to pronunciation.

Apart from two short conference proceedings – Bárkányi and Melchor-Couto (Reference Bárkányi, Borthwick, Bradley and Thouësny2017) and Bárkányi (Reference Bárkányi, Rosell-Aguilar, Beaven and Fuertes Gutiérrez2018) – no studies focusing on FLSA on LMOOCs have been published so far. Both proceedings show that respondents have positive attitudes towards language learning on LMOOCs, and FLSA levels are quite stable throughout the courses. Conclusions on the relationship of self-efficacy beliefs and FLSA are based on a limited and unbalanced set of data, and neither study included motivation as a potential factor. The present research intends to fill this gap in the literature by analysing the impact of motivation and anxiety on speaking on LMOOCs, thus illuminating the affordances of this learning context in L2 and FL acquisition.

3.1 Background and course design

The Spanish for Beginners programme consists of six four-week courses that together cover level A1 of the Common European Framework of Reference for Languages. Weeks consist of 23–32 steps; that is, activities that involve articles (written texts), quizzes (gap-fill, simple choice, and true or false), audios and videos (often followed by a quiz or discussion), and discussions. Discussions are the only collaborative tool offered by FutureLearn. They are embedded in the learning content and are mainly of two types: they require students to practice their production skills (writing or speaking) in Spanish or offer the opportunity for students to reflect on cultural aspects of the Spanish-speaking societies. Discussions are elicited by an activity or an article produced by the course designers. This is where students can connect to each other, share knowledge and support each other, where feedback and encouragement are provided, and questions are answered. The only way to carry out speaking activities on these LMOOCs is for learners to record their speech on third-party sites and post a link to their recordings. Speaking activities are either of the “listen-and-repeat” type or learners are prompted by a task to use the language content of the previous steps (e.g. “describe a friend or family member”). Instructors and peers can comment on these recordings, ideally in another recording, but most often in a written form. As the number of participants in some courses might be extremely high, it is challenging for instructors to respond to all queries. However, the course team paid special attention to giving feedback on all oral productions.

3.2 Data collection

A mixed-methods approach was applied and both quantitative and qualitative data were collected concurrently. There have been five iterations of the programme to date. The first course started in August 2016 and had 49,120 registered learners, from which 4,903 completed at least 50% of the activities. The last course of the last iteration started in November 2019; it had 7,332 registered learners, only 218 of whom completed at least 50% of the activities. Data in this paper are gathered from courses 2, 3, 5 and 6 of the third iteration; that is, between February 2018 and August 2018. The pre-course questionnaire was not in place for the first course, and due to a technical error, the collector did not register data at the beginning of the fourth course, thus these courses were removed from the analysis. This leaves 978 responses altogether.

Data were collected via pre-course and post-course self-reflective questionnaires (which were voluntary) and from discussion forums. Demographic data were obtained from FutureLearn’s statistical dashboard. FutureLearn allows course owners to add five questions to their surveys. Thus, information referring to learners’ motivations and goals was obtained from FutureLearn’s non-course-specific questions, while self-efficacy ratings and information on FLSA were obtained from the five extra questions (the full list of questions can be accessed from the IRIS databaseFootnote ² ). As a starting point for designing the FLSA questions, the FLCAS (Horwitz et al., Reference Horwitz, Horwitz and Cope1986) was considered and adapted to the LMOOC context. Eight Likert-type questions refer to various aspects of speaking apprehension, six of which are analysed in the present study (Table 1).

Table 1. Statements on foreign language speaking anxiety administered in Likert-type test

Five Likert-type questions address self-efficacy ratings. An example is shown as follows:

For my current level of study in Spanish, I think that my Spanish fluency is …

Very poor   Fairly poor   Neither poor nor good   Fairly good   Very good

Two yes-or-no questions inquire about attitudes towards speaking activities. An example is reproduced as follows:

Do you intend to record yourself in at least some of the speaking activities in this course?

Did you record yourself in at least some of the speaking activities?

Yes   No

Qualitative data were obtained from open-ended questions on the questionnaires and comments in the discussion forums of speaking activities.

3.3 Challenge to data analysis

Ideally, in order to find out how speaking anxiety levels and self-efficacy beliefs change as a result of participation in the courses, learners’ matched answers should be compared from before and after the courses. There are several factors that impede the analysis of such repeated measures in the present study. First, respondents sometimes skipped a question, but completed most of the questionnaire. Second, post-course surveys have considerably fewer responses as compared to pre-course surveys – in our data, 10.3–25.8%. We could have chosen to match up post-course respondents with their pre-course responses only. However, it turns out that from the 978 responses on the four courses only 28 matched pairs were found. For this reason, all pre-course data were compiled to form a pre-course data set and all post-course data formed a post-course set. In this way, individual idiosyncrasies cannot be taken into account, but as a compensation, large participant numbers could be preserved for analysis. Both sets have been treated as two groups, and, following Boone and Boone’s (Reference Boone and Boone2012) recommendations on how to analyse Likert-type data, Pearson’s chi-square test was applied to contrast pre-course and post-course answers.

3.4 Data coding and analysis

All quantitative data were transferred to Excel and additional codes were added (e.g. “pre” and “post” and the number of the course) in order to carry out initial calculations. Then data were uploaded to the software SPSS and statistical analyses, including Kendall’s Tau-b test (where relevant) and the χ² test, were carried out, and, in the case of small sample sizes (not enough items in a given cell), Fisher’s exact test was employed. When handling the data, the 5-point scale used in the questionnaire was maintained; that is, data from various categories were not merged (e.g. “agree” and “strongly agree”). For the correlational analysis, we opted for Kendall’s Tau-b test as it is widely used in similar experiments, although partly because it is suitable for small sample sizes, which is not the case here, but also because it is a more conservative measure than, for example, Pearson’s correlation test. The following variables were included in the correlation analysis: motivation, FLSA, computer-mediated speaking (CMS), and self-efficacy ratings. Motivational data were assigned into three categories: (1) career related, (2) managing or fixing a personal situation, and (3) leisure and pleasure. In order to determine a learner’s FLSA level, the answers were converted into points and each learner’s mean score was calculated. Responses on the negatively worded items were changed accordingly. The same procedure was applied for CMS and self-efficacy responses, although there were no negatively worded questions in these categories.

Data collected were anonymised by FutureLearn and a unique encrypted key was assigned to each participant. Although anonymisation would have allowed us to match pre-course and post-course responses, this was not possible because of the lack of data, as explained in section 3.3. Due to the anonymisation, it is impossible to match survey data with discourse in the forum. The responses to the open-ended questions and from forum discussions made up the qualitative data. A total of 22,081 comments (over 620,000 words) were produced in the comment area of the courses. These include solutions to the tasks and tutor feedback. Comments in which learners reflect on their listening and speaking skills, language learning, and self-efficacy have been selected for analysis; this includes 273 comments resulting in 8,084 words. Thematic analysis was used for analysing qualitative data. However, in the first examination of the data, no categories were predetermined. Therefore, categorisation was carried out by the author in an inductive and recursive manner, based on the phases established by Braun and Clarke (Reference Braun, Clarke, Cooper, Camic, Long, Panter, Rindskopf and Sher2012). The software NVivo 12 was used for this purpose in order to make the process simpler and the analysis of the results easier. Table 2 shows the categories that emerged in the coding process and their frequency.

Table 2. Categories used for qualitative analysis

3.5 Participants

There were 35,644 joiners on the third iteration of the Spanish for Beginners 2, 3, 5, and 6 courses. The majority of learners (43–46%) were registered in the United Kingdom, followed by 6–7% from the United States, and 4–6% from Spain, although there were joiners from 156–174 countries. Out of these, 2,271 learners completed at least 50% of the activities in the courses. Participants were evenly distributed across the age ranges with two small peaks: 26–35 years old (19–23%) and 65+ (14–18%), although on all courses approximately 15% of learners were 56–65 years old.