3.1 Domain and skills

The results related to domain and language skills are presented in Appendix D and Appendix E, respectively. The domain included language learning (n = 23), cross-cultural learning (n = 2), and distance learning (n = 1). According to the results, in the language learning domain, SRT was applied to English (n = 17), Dutch (n = 3), French (n = 2), and Spanish (n = 1) learning. In the cross-cultural learning domain, scholars used SRT to promote multilingual communication and cross-cultural understanding of the participants. In the distance learning domain, SRT was used to support educational technology courses taken by students from China and Japan. The results showed that SRT was used to assist different language skills: pronunciation (n = 15), listening (n = 5), writing (n = 3), communication (n = 3), grammar (n = 2), word recognition (n = 1), and vocabulary (n = 1). Some studies focused on a single skill and others on more than one. For example, Tsai (Reference Tsai2019) used SRT to improve student pronunciation only, whereas Cavus and Ibrahim (Reference Cavus and Ibrahim2017) focused on facilitating pronunciation as well as listening.

Some of our results echo those obtained in earlier studies. For example, Ehsani and Knodt (Reference Ehsani and Knodt1998) and McKechnie et al. (Reference McKechnie, Ahmed, Gutierrez-Osuna, Monroe, McCabe and Ballard2018) mentioned that SRT was applied to the domain of language learning. Studies by Shadiev et al. (Reference Shadiev, Hwang, Chen and Huang2014) and Radha and Vimala (Reference Radha and Vimala2012) are not comparable because they reported applications of SRT to education in general. In contrast to previous studies, the findings of the present study revealed new domains, such as cross-cultural learning and distance learning. Perhaps due to advancements in SRT, educators and researchers have considered its applications to such new domains. For example, the students in Shadiev et al. (Reference Shadiev, Wu, Sun and Huang2018) represented 13 nationalities, and SRT combined with machine translation enabled them to communicate with each other in their native languages. A student from one nationality spoke in his or her native language, and a student from another nationality could understand speech content because the technology received speech input and translated it. In this way, the students exchanged culture-related information. Another difference between previous studies and the present research is that none of the earlier studies considered any skills that could be supported by SRT.

Based on our results, we suggest that SRT can be applied not only to assist language learning but also in other learning contexts, such as cross-cultural learning and educational technology learning. In the language learning domain, successful research projects were implemented with respect to English, Dutch, French, and Spanish. In addition, SRT was employed to assist cross-cultural learning and educational technology learning of distant students. Educators and researchers who are planning to use SRT in these domains may find useful information from the reviewed studies to guide their teaching and research.

The results also showed certain domains and languages that received more attention from researchers. That is, language learning was the main domain, and English was the dominant language in the reviewed studies. English was the dominant language in the reviewed articles, as it is considered one of the most commonly spoken languages (Cavus & Ibrahim, Reference Cavus and Ibrahim2017). This finding is in line with those obtained in other review studies (McKechnie et al., Reference McKechnie, Ahmed, Gutierrez-Osuna, Monroe, McCabe and Ballard2018; Shadiev & Yang, Reference Shadiev and Yang2020). For example, McKechnie et al. (Reference McKechnie, Ahmed, Gutierrez-Osuna, Monroe, McCabe and Ballard2018) found that SRTs were used to support the learning of 13 different languages, most commonly English.

Our results demonstrated that SRT could be applied to facilitate skills such as pronunciation, listening, writing, grammar, and word recognition. Therefore, educators and researchers who apply SRT may focus on these skills. The results suggest that the most commonly supported skill was pronunciation and that the remaining skills were targeted least frequently. The reason is that SRT is capable of receiving voice input, analyzing it, and providing corrective feedback so that language learners can improve their pronunciation (Cavus & Ibrahim, Reference Cavus and Ibrahim2017; Tsai, Reference Tsai2019). In terms of improving listening, SRT can generate texts from the speech input of the instructor, and language learners are able to listen to the instructor and read transcriptions simultaneously to confirm unfamiliar vocabulary or what they misheard from the speech (Mirzaei et al., Reference Mirzaei, Meshgi, Akita and Kawahara2017; Shadiev et al., Reference Shadiev, Huang and Hwang2017). Furthermore, Shadiev et al. (Reference Shadiev, Sun and Huang2019), Shadiev et al. (Reference Shadiev, Wu, Sun and Huang2018), and Yueh, Lin, Liu, Shoji and Minoh (Reference Yueh, Lin, Liu, Shoji and Minoh2014) proposed extending the speech-to-text recognition process with computer-aided translation to assist communication among learners without a common language. Learners spoke to the system, generated translated texts from speech input, and translated texts were shown to language partners. For writing, SRT can help learners demonstrate compositional ability independent of transcription; for example, technology mitigates challenges with spelling and allows students to compose written content using their rich oral vocabulary (Arcon et al., Reference Arcon, Klein and Dombroski2017; Baker, Reference Baker2017). For vocabulary, word recognition and grammar learning, learners had to speak vocabulary words and their order in a sentence to SRT so that it could analyze learners’ speech and provide immediate corrective feedback, such as the correct pronunciation, their recognition, and order (Bodnar et al., Reference Bodnar, Cucchiarini, de Vries, Strik and van Hout2017; Cavus & Ibrahim, Reference Cavus and Ibrahim2017; Matthews & O’Toole, Reference Matthews and O’Toole2015).

3.2 Technology and its application

Seven types of SRTs (Appendix F) and three different applications (Appendix G) were identified. Scholars used Dragon Naturally Speaking (n = 4), Google speech recognition (n = 4), Windows Speech Recognition (n = 2), automatic speech recognition (ASR)-based CALL system (n = 1), partial and synchronized captioning (PSC) (n = 1), and Julius (n = 1). The results show that SRTs were not specified in 13 studies. According to the data, SRT was used to provide feedback (n = 19), showing texts generated by technology (n = 9) and giving commands to the system (n = 3).

The results showed that the SRTs Dragon Naturally Speaking, Google speech recognition, Windows Speech Recognition, ASR-based CALL system, PSC, and Julius have been used to assist language learning. Most recognition technologies feature continuous speech recognition (i.e. when SRT transcribes naturally dictated speech) and thus were used in reviewed studies in writing practice, listening skills development, and pronunciation training. Therefore, they might be considered by educators and researchers in their future studies. Two recognition technologies, Dragon Naturally Speaking and Google speech recognition, were the most frequently used. One reason is because both Dragon Naturally Speaking and Google speech recognition are designed to recognize speech independent of the speaker (i.e. they are designed to respond to a word or phrase regardless of who speaks) and can be installed on mobile phones, tablet PCs, or desktop and laptop computers with either Microsoft Windows or Mac operating systems. Although Dragon Naturally Speaking is not free, it is mature and undergoes constant improvement. For this reason, Dragon Naturally Speaking is used mostly by professionals (e.g. doctors, lawyers, court stenographers). On the other hand, Google speech recognition is a cloud-based service that is available free of charge and features a high rate of accuracy. Therefore, Google speech recognition is mostly used by casual users. Microsoft Windows Speech Recognition is the third most popular technology. Windows Speech Recognition comes pre-built into the Windows operating system, and thus is free to anyone with a Windows PC but unavailable to those who use different computer operating systems.

In comparison to Dragon Naturally Speaking, Google speech recognition, and Windows Speech Recognition, SRTs such as the ASR-based CALL system, PSC, and Julius were developed by researchers for academic purposes only and therefore are not publicly available. These tools were used as an alternative to existing speech recognition methods. For the ASR-based CALL system, researchers judged learners’ pronunciation by comparing HMM with built-in corpus approaches (Wang & Young, Reference Wang and Young2015). In PSC, the system transcribed speech input in the form of a set of words or phrases, which appeared on the screen in a one-by-one sequence synchronized with the utterance (Mirzaei et al., Reference Mirzaei, Meshgi, Akita and Kawahara2017). Julius was implemented to control a camera and to access demographic information and learning statistics of students using voice commands in a distant learning environment (Yueh et al., Reference Yueh, Lin, Liu, Shoji and Minoh2014). It should also be noted that in contrast to Dragon Naturally Speaking and Google speech recognition, Windows Speech Recognition, ASR-based CALL system, PSC, and Julius are speaker-dependent recognition systems; that is, they are trained by the individual who will be using the system. Furthermore, they are only available through desktop and laptop computers.

Some of our results are similar to those that were reported in other review studies. For example, Shadiev et al. (Reference Shadiev, Hwang, Chen and Huang2014) also found that Dragon Naturally Speaking was among the most frequently used SRT. Shadiev et al. (Reference Shadiev, Hwang, Chen and Huang2014) also mentioned two other popular systems, Windows Speech Recognition and IBM ViaVoice. Although IBM ViaVoice was popular a decade ago, it has since been discontinued and thus is no longer used. On the other hand, Google speech recognition was not mentioned in Shadiev et al. (Reference Shadiev, Hwang, Chen and Huang2014) because Google’s first effort at speech recognition came only recently. It was not widespread among educators and researchers before, and therefore previous review studies did not mention it.

The results showed that SRTs were not specified in 13 studies. That is, half of the reviewed studies simply did not specify their technology, which is problematic. We suggest that scholars should specify their technology in order to disclose what SRTs were used, their applications to learning, and their impacts (i.e. either positive or negative) on learning outcomes. According to the results, providing feedback was the most frequent application. A learner spoke a word or read a sentence, and the technology provided feedback so that a learner could see where and what kind of mistakes she made. For example, students in Ahn and Lee (Reference Ahn and Lee2016) interacted with the system to identify errors in their speech, and students in de Vries, Cucchiarini, Bodnar, Strik and van Hout (Reference de Vries, Cucchiarini, Bodnar, Strik and van Hout2015) received different types of corrective feedback (e.g. implicitness and explicitness, reformulations or prompts to improve pronunciation). In showing texts generated by the technology applications, a lecturer or student spoke to the system, which then generated texts from the voice input and showed them a whiteboard or computer screen (Shadiev et al., Reference Shadiev, Huang and Hwang2017). Such applications are popular in lectures for enhancing students’ comprehension of lecture content, especially when lectures are delivered in a foreign language as a medium of instruction (e.g. English as a medium of instruction) (Mirzaei et al., Reference Mirzaei, Meshgi, Akita and Kawahara2017). To give commands to the system applications, a lecturer or student spoke to the system to execute a command. For example, students in Arcon et al. (Reference Arcon, Klein and Dombroski2017) composed and modified their essays by executing specific word-processing editing commands through a speech recognition interface. Such an approach was useful in improving speaking skills. Previous related studies reported similar applications of SRT. For example, the most popular were providing feedback (Ehsani & Knodt, Reference Ehsani and Knodt1998; McKechnie et al., Reference McKechnie, Ahmed, Gutierrez-Osuna, Monroe, McCabe and Ballard2018) or dictating (Shadiev et al., Reference Shadiev, Hwang, Chen and Huang2014). Other related review studies focused on the technical aspect of technology instead of the educational aspects (Ehsani & Knodt, Reference Ehsani and Knodt1998; Radha & Vimala, Reference Radha and Vimala2012).

3.3 Participants and intervention duration

Identifying the educational level and size of a population is essential for researchers and language teachers in their design of future research. Studies usually cover diverse contexts, target learners at different educational levels, and involve different numbers of learners. Approaches that were found to be successful for learners of one education level may not be as effective for learners of another education level. For example, Arcon et al. (Reference Arcon, Klein and Dombroski2017) carried out their research with primary school students in the context of persuasive writing. Arcon et al. (Reference Arcon, Klein and Dombroski2017) warned that their results could not be generalized to other populations and contexts; for example, written compositions made by adult learners may produce different patterns of results. It will also be inappropriate for young learners (e.g. preschool or primary school students) to participate in language learning activities designed for adult learners (e.g. university students), as the activities might be more complex and difficult. The number of participants can give confidence in the obtained findings as well as recommendations to researchers and educators about how many participants should be involved in specific learning activities. Therefore, without knowing what kind of learners or how many of them were involved in a particular study, it is difficult to develop appropriate and effective learning activities supported by technology.

The educational level and the number of participants are shown in Appendix H and Appendix I, respectively. College (n = 20), elementary school (n = 4), junior high school (n = 2), and preschool (n = 1) students were involved in the reviewed studies. In two studies, the educational level of participants was not specified. The number of participants in the reviewed studies ranged from 10 to 341. There were 12 studies in which the number of participants was less than 30, 10 studies with a number between 30 and 60, and four studies in which more than 60 participants were involved.

Intervention duration is shown in Appendix J. The results demonstrate that there is a range of intervention durations: less than one hour (n = 5), less than one day (n = 2), less than one week (n = 1), less than one month (n = 9), more than one month (n = 6), or not specified at all (n = 3). Some examples of short-term and long-term interventions are provided. For example, in Arcon et al. (Reference Arcon, Klein and Dombroski2017), students trained on SRT and composed persuasive texts on assigned topics. They did so in three sessions, each lasting for approximately 20 minutes. Students in Dalim, Sunar, Dey and Billinghurst (Reference Dalim, Sunar, Dey and Billinghurst2020) interacted with the system to improve their language skills, and the entire session lasted between 30 and 35 minutes. Hsu (Reference Hsu2016) arranged a three-month self-regulated speech-recognition-based course, and Wang and Young (Reference Wang and Young2015) asked participants to practice their language skills twice a week and to complete their learning content in eight weeks.

According to the results, college students made up the majority of the participants. This is because college students are a large group of language learners, and it is convenient for researchers to conduct experiments with such learners. Primary and secondary school students were involved in research to a lesser extent. This is because this group of learners has less experience in both language learning and technology usage. Our results are different from those obtained by McKechnie et al. (Reference McKechnie, Ahmed, Gutierrez-Osuna, Monroe, McCabe and Ballard2018). In their review study, scholars mostly involved participants up to 16 years (i.e. secondary or high school students). Only one study had participants up to 21 years (i.e. college students). This is perhaps because McKechnie et al. (Reference McKechnie, Ahmed, Gutierrez-Osuna, Monroe, McCabe and Ballard2018) focused on a specific group of learners (e.g. children with speech sound disorders).

In the face of such research information, we believe that younger groups will be more involved in research and that with the continuous development of technology, more suitable SRT for young children will be developed. Thus, it is necessary for researchers to take other groups of learners into consideration, provide them with necessary and easy-to-use software, and teach them learning strategies to apply technology to language learning more efficiently (e.g. Shadiev et al., Reference Shadiev, Hwang, Huang and Liu2016).

The number of studies with small and large samples (i.e. small and large numbers of participants) was almost the same. Although small samples can reduce the workload of the experiment, we found that studies using small sample populations acknowledged it as a limitation and noted that their results could not be generalized beyond that particular group involved. For example, 21 students participated in the study of Arcon et al. (Reference Arcon, Klein and Dombroski2017), and 22 participants were involved in Baker (Reference Baker2017). Some scholars argued that they aimed for a large group of participants to obtain a sample that is likely to be representative of population values. To make their conclusions robust and valid, Hsu (Reference Hsu2016) had 341 students participate in the experiment, and it is worth mentioning that in Ahn and Lee (Reference Ahn and Lee2016), there were more than 1,000 participants.

Regarding duration, eight studies were relatively short (less than one week). It is possible that such a short duration may result in frustration with the learning experience involving technology and negatively affect the learning outcomes. For example, students need to train on SRT and learn how to use it to learn more efficiently. Scholars have suggested that at least two weeks are needed for this process (Shadiev et al., Reference Shadiev, Sun and Huang2019; Shadiev et al., Reference Shadiev, Wu, Sun and Huang2018). Therefore, educators and researchers need to consider having their intervention for a longer period of time, with at least a few weeks dedicated to technology training.

3.4 Measure

The primary goal of including this dimension was to explain the overall data collection preference and the tendency of the research. The measures (i.e. data collections instruments) used are shown in Appendix K. They are a questionnaire (n = 24); pre-/posttest (n = 14); interviews (n = 11); content of reflective notes, created texts, and think-aloud protocols (n = 3); learning logs (n = 3); EEG recordings (n = 2); fieldwork method (n = 1); eye tracking (n = 1); task analysis (n = 1); language learning logs (n = 1); and usability review (n = 1). Scholars used questionnaires to investigate participants’ perceptions (n = 17) of SRT usefulness or student cognitive load (n = 3). For example, Bodnar et al. (Reference Bodnar, Cucchiarini, de Vries, Strik and van Hout2017) measured student perceptions of corrective feedback based on SRT. Pre-/posttest was administered in reviewed studies to measure student proficiency before and after the intervention (n = 9). For example, Mirzaei et al. (Reference Mirzaei, Meshgi, Akita and Kawahara2017) employed comprehension tests to measure the effects of the intervention. In the interviews, scholars mostly explored student learning experiences (n = 8). For example, teachers were interviewed in Baker (Reference Baker2017) to gather information on how students were learning with the support of technology.

During the application of SRT, various measurements provide not only scientific evidence of the effectiveness of the intervention but also useful feedback on the technology so that it can be improved. We can also see that researchers pay attention to both learning outcomes and learning experiences in the technology use process. Therefore, educators and researchers may consider these aspects in future planned studies.

Research methods can be divided into qualitative and quantitative research. Scholars used pre-/posttest design to investigate the effectiveness of applying SRT to assist learning by comparing learning outcomes of the control and experimental groups. For example, in quantitative research, Cavus and Ibrahim (Reference Cavus and Ibrahim2017) administered pre- and posttests to explore the improvement of language skills. Scholars compared the results of the tests between two groups to demonstrate the effects of their intervention. Questionnaires were mostly used to measure participants’ perceptions of the intervention. For example, in qualitative research, Ahn and Lee (Reference Ahn and Lee2016) aimed to measure the perceptions and attitudes of their students toward the application of speech-to-tech recognition in terms of its design, convenience, and efficacy. Interviews are mostly administered by researchers to obtain insights into participants’ learning experiences with the technology. For example, in qualitative research, Liakin, Cardoso and Liakina (Reference Liakin, Cardoso and Liakina2017) interviewed the participants to explore their learning experience with SRT during language learning. Some scholars even collected the data using mixed methods (e.g. Arcon et al., Reference Arcon, Klein and Dombroski2017; de Vries et al., Reference de Vries, Cucchiarini, Bodnar, Strik and van Hout2015; Liakin et al., Reference Liakin, Cardoso and Liakina2017). This approach is common in research because it enables the triangulation of data from different sources to make results and conclusions reliable and robust (Shadiev et al., Reference Shadiev, Huang and Hwang2017).

Log data were collected by scholars to objectively reflect on the state of students’ use of technology. When such data are combined with an interview or questionnaire data, the results become more convincing because researchers are able to triangulate data collected from various sources. Interestingly, among these measures, very few studies have employed physiological measurement tools to obtain physiological data. Such data objectively reflect the physiological state of learners and enable us to know whether SRT is effective in improving physiological learning outcomes (e.g. learning or visual attention) and to make appropriate changes in the intervention design.

3.5 Results in reviewed studies

The results in the reviewed studies are shown in Appendix L. The results can be divided into five areas: (1) gains in proficiency – there were changes in certain language skills after the intervention; (2) perceptions – students’ perceptions of the intervention; (3) questions, suggestions, or approaches – some questions were raised after the intervention and suggestions or approaches were proposed; (4) system design – results concerned the system design; and (5) learning logs – records of system usage for language learning. Proficiency gains in terms of writing (Baker, Reference Baker2017; Haug & Klein, Reference Haug and Klein2018), pronunciation (McCrocklin, Reference McCrocklin2016; Tsai, Reference Tsai2019), and grammar (de Vries et al., Reference de Vries, Cucchiarini, Bodnar, Strik and van Hout2015) were reported in reviewed studies. For example, Haug and Klein (Reference Haug and Klein2018) argued that SRT played an important role in learning writing strategies. Haug and Klein (Reference Haug and Klein2018: 1) found that using the technology resulted in “large gains in text quality, word count, and variety of argument moves.”

The effects of SRT applications are different depending on students’ abilities. For example, the technology significantly improved the writing outcomes of students who struggled with writing (MacArthur & Cavalier, Reference MacArthur and Cavalier2004) but it did not help students with better writing abilities (Arcon et al., Reference Arcon, Klein and Dombroski2017). In another study by Shadiev et al. (Reference Shadiev, Huang and Hwang2017), texts generated by SRT from the instructor’s speech during lectures in a foreign language were very useful for low language ability students. They listened to the instructor and read texts to better comprehend lecture content. However, the same texts were not useful and were even distracting for high language ability students because listening to the instructor was enough for them (Shadiev et al., Reference Shadiev, Huang and Hwang2017). This issue should be considered in future studies – for example, how technology can help students with lower language abilities without hurting the learning of students with greater abilities.

In pronunciation learning, SRT helped improve pronunciation significantly, not only in terms of the accuracy of pronunciation but also in awareness of pronunciation errors and linguistic features (Tsai, Reference Tsai2019). SRT is a potential tool for practicing oral grammar, as it detects errors in the speech of learners, and then learners receive immediate corrective feedback (de Vries et al., Reference de Vries, Cucchiarini, Bodnar, Strik and van Hout2015). SRT feedback can be of different forms, such as texts generated by technology, graphs with sound waves, or echoed speech. Of course, more feedback and details on the mistakes made or specific instructions on how to improve them will be beneficial for improving learning outcomes. However, it is still not clear how to combine multiple feedback forms, what combination of feedback forms, and in what quantity feedback needs to be provided. Future studies may consider exploring this aspect.

Regarding perceptions, participants generally had positive perceptions of their experiences using SRT. SRT helped students practice speaking in private spaces (Ahn & Lee, Reference Ahn and Lee2016), enabled them to use the target language in situated contexts and assisted their understanding of language appropriateness (Wang & Young, Reference Wang and Young2014). In addition, students became aware of their errors and were able to modify their utterances (Liakin et al., Reference Liakin, Cardoso and Liakina2017). Most of this evidence is subjective in nature and prone to bias. Therefore, future studies may consider exploring learning experiences (cognitive load, learning attention, learning emotions, etc.) using objective approaches based on psychological data (EEG or eye tracking). For example, it is possible to know objectively and precisely what part of the learning content is more difficult or when learners are overloaded when learning a new language.

In the reviewed papers, some scholars posed questions and proposed suggestions or approaches. For example, Liakin et al. (Reference Liakin, Cardoso and Liakina2017) claimed that language learners do not have enough time and opportunities for pronunciation practice in language classrooms. To address this issue, they introduced SRT implemented on mobile devices so that learners can practice this skill for as long as they need, anytime and anywhere. Tsai (Reference Tsai2019) reported that although SRT used for pronunciation learning was useful to provide feedback related to such linguistic features (e.g. pitch variation), some learners doubted the validity of its scoring system and complained about its feedback. That is, learners followed the model utterance regarding pausing, and their scores were lower than when they read without any pauses. Other learners complained that feedback from the system was not as clear as the feedback from the instructor or that no feedback on how to correct their production was provided. Therefore, researchers have suggested that instructors may integrate peer interaction activities into the learning process; such an approach may enable assistance from peers that cannot be provided by the software (e.g. some support or strategies to obtain better learning outcomes).

Results related to SRT design were positive. The SRT “interface appeared good and learners could find the areas of their interests easily” (Yu et al., Reference Yu, Pan, Li, Zhang, Shi, Chu, Liu and Zhu2016: 997). Teachers in van Doremalen, Boves, Colpaert, Cucchiarini and Strik (Reference van Doremalen, Boves, Colpaert, Cucchiarini and Strik2016) thought that SRT was easy to use and that students were satisfied with the tool. Scholars used learning logs to know how often and how long learners used SRT (Wang & Young, Reference Wang and Young2014). In addition, we found that scholars do not get the most out of their systems. For example, it is possible to add some features that can record the learning behavior of learners (e.g. what functions they used and how frequently). Learning behavior analysis did not receive much attention in the reviewed studies, but it can help reveal behavior patterns, behavior habits, and behavior rules to promote our understanding and optimization of language learning through SRTs. Therefore, scholars may consider exploring learning behavior in the future. We could not find learning analytics and big data approaches in the reviewed studies. It is possible to employ such approaches to further optimize language learning experiences; they can help instructors better support struggling language learners by personalizing the learning process and adapting their teaching when necessary.

Finally, scholars may consider combining SRT with other tools according to different learning purposes. For example, in the process of assisting word recognition and pronunciation learning, dictionaries, voice recording and playback, as well as other functions that may be frequently used by learners, can be integrated. In writing training, SRT can be combined with a proofreading function so that feedback on grammar, spelling, and punctuation can also be provided.

Most previous review studies also reported similar results, and most of them were positive. For example, Ehsani and Knodt (Reference Ehsani and Knodt1998), McKechnie et al. (Reference McKechnie, Ahmed, Gutierrez-Osuna, Monroe, McCabe and Ballard2018), and Shadiev et al. (Reference Shadiev, Hwang, Chen and Huang2014) found an improvement in students’ language learning. However, the results of the reviewed studies did cover other dimensions – for example, cross-cultural learning and learning system design. Other review studies, such as that by Radha and Vimala (Reference Radha and Vimala2012), did not consider such results at all.

3.6 Advantages and disadvantages of using speech recognition technology

The reported advantages and disadvantages are shown in Appendix M. The advantages can be divided into 10 different aspects: improving affective factors (n = 18), enhancing language skills (n = 14), promoting interaction (n = 5), creating a self-paced learning environment and improving autonomy (n = 5), increasing learning involvement (n = 3), self-monitoring errors (n = 2), enhancing intercultural sensitivity (n = 2), supporting learner differences (n = 1), reducing task completion time (n = 1), and developing awareness of intelligibility (n = 1). Scholars have reported that using SRT is beneficial for improving affective factors during language learning by, for example, decreasing frustration and anxiety or increasing enjoyment, confidence, and motivation. Dalim et al. (Reference Dalim, Sunar, Dey and Billinghurst2020) received many positive comments regarding the system – for example, using the system was enjoyable and exciting. Mroz (Reference Mroz2018) reported affective gains, such as an increase in learning motivation and self-confidence, as well as learners’ willingness to communicate, particularly for apprehensive learners. Language skills development – for example, pronunciation (Cavus & Ibrahim, Reference Cavus and Ibrahim2017), listening (Mirzaei et al., Reference Mirzaei, Meshgi, Akita and Kawahara2017), or writing (Arcon et al., Reference Arcon, Klein and Dombroski2017) – was also reported. Cavus and Ibrahim (Reference Cavus and Ibrahim2017) claimed that student skills (i.e. listening, vocabulary, or pronunciation) showed significant improvements as a result of using SRT. In terms of promoting interaction, SRT enabled a novel type of interaction between language learners and technology. Students were able to speak to the system and be provided with immediate feedback. Students in Ahn and Lee (Reference Ahn and Lee2016) liked the system, as they found speaking practice with SRT to be more motivating, enjoyable, and interactive; students felt like conversing with a real person.

Some disadvantages were also reported: the low accuracy rate of the system (n = 9), its insufficiency (i.e. SRT lacked some useful features to support learning efficiently) (n = 8), the system placing a burden on some students (n = 3), and being time-consuming to use (n = 1). Disadvantages were not specified in nine studies. Scholars have reported that the system generates texts from speech input with errors (Arcon et al., Reference Arcon, Klein and Dombroski2017). Usually, the accuracy rate decreases in a noisy environment when several people speak at the same time (Dalim et al., Reference Dalim, Sunar, Dey and Billinghurst2020) or speech input is lengthy (Mroz, Reference Mroz2018). To combat these issues, Shadiev et al. (Reference Shadiev, Huang and Hwang2017) and Yueh et al. (Reference Yueh, Lin, Liu, Shoji and Minoh2014) provided strategies, such as training on the technology beforehand or using it in a quiet environment. More useful strategies to use SRT for learning and achieving higher accuracy rates were proposed by Shadiev et al. (Reference Shadiev, Hwang, Huang and Liu2016). In terms of insufficiency, some tools lacked features that could efficiently support learning. For example, Wang and Young (Reference Wang and Young2014) found that implicit feedback showing learners’ pronunciation scores and audio waveforms was insufficient to recognize their pronunciation errors. Therefore, it was suggested that feedback be provided in an “explicit format of immediate audio replay (recast) with a textual description” (Wang & Young, Reference Wang and Young2014: 230). Scholars also warned that recognition tools might place a burden on students. For example, students felt frustrated by the low accuracy rate of SRT as a result of their accent when they practiced writing skills (Arcon et al., Reference Arcon, Klein and Dombroski2017) or pronunciation (McCrocklin, Reference McCrocklin2016), and they had to spend much time completing their tasks using SRT. To avoid this problem, students need to be trained on SRT beforehand, the instructor needs to provide students with necessary practice strategies, and certain SRTs should be used that are able to recognize speech in English with certain types of accent. Another example is that texts generated by SRT were distracting to students with high language abilities because they were able to understand the content without any support (Shadiev & Huang, Reference Shadiev and Huang2020). Therefore, a more flexible approach to using SRT was proposed; learners may switch on displays with texts when they need them and switch them off when support is no longer required. It is possible that the quick conversion from speech to text may, to a certain extent, indulge some students’ laziness. To avoid this, educators and researchers need to design learning activities in such a way that students have manageable assignments, understand the affordances of SRT for their learning, know what to do and how to do it, have a positive learning experience, are motivated, and can obtain assistance and guidance whenever needed.

Some results of the present study on advantages and disadvantages of using SRT are in line with the results from related review studies; for example, SRT is beneficial for enhancing language skills, or problems associated with accuracy rates of SRT have been reported previously (Ehsani & Knodt, Reference Ehsani and Knodt1998; McKechnie et al., Reference McKechnie, Ahmed, Gutierrez-Osuna, Monroe, McCabe and Ballard2018; Shadiev et al., Reference Shadiev, Hwang, Chen and Huang2014). However, reported advantages and disadvantages of SRT for language learning varied in the present review studies compared to earlier review studies. For example, we found that SRT was beneficial for improving affective factors or enhancing intercultural sensitivity. Reviewed studies also mentioned the insufficiency of SRT or that it placed a burden on language learners. Such variety of reported advantages and disadvantages can be accounted for the fact that earlier review studies focused either on SRT applications to support education in general but not on language learning specifically (Shadiev et al., Reference Shadiev, Hwang, Chen and Huang2014) or on technical aspects of SRT instead of educational (Ehsani & Knodt, Reference Ehsani and Knodt1998; Radha & Vimala, Reference Radha and Vimala2012). For these reasons, our findings may be of greater interest to those who target language learning process supported by SRT.