Figure 1 shows the patient enrollment process. Eighty people over the age of 60 years with normal cognitive function were recruited through community advertising and enrolled through three community centers. During the recruitment process, 5 potential participants were excluded because they met one of the following exclusion criteria: (1) hearing, visual, or speech disabilities; (2) history of stroke, brain tumor, head trauma, or severe psychiatric disorders; or (3) impaired cognitive functioning (ie, Mini-Mental State Examination score <24). The remaining 75 participants were assigned into either the training group (n=38) or the control group (n=37) using the convenience sampling method. The training group assignees were provided with SK Telecom’s smart speakers, while the control group assignees were not. Three participants in the training group and 6 participants in the control group were dropped from our analysis because they did not complete the poststudy evaluation. Thirty-five participants in the training group finished the smart speaker–based MMT and completed the pre/post cognitive tests. In the training group, we excluded 6 participants who had participation rates below 80%. Thirty-one participants in the control group completed the pre/post examinations. Finally, the analysis was conducted on 29 participants in the training group and 31 participants in the control group using per protocol analysis.

The emergence and integration of voice interfaces using AI technology has led to the recent development of smart speakers. Smart speakers contain software that provides customized information or services to users while communicating with them by voice. The devices receive requests from users in the form of voice, text, or other communication forms and then performs the requested tasks.

As AI technology is at the core of smart speakers, the spoken language dialogue system (SLDS) proposed by Bertrand et al [13] is most similar to the system used in this study. SLDS functions include (1) speech recognition (including language understanding), (2) speech synthesis, and (3) dialogue management; all are similar to the current smart speaker system structure. Speech recognition technology refers to technology that receives human speech and converts it into text. This type of human–machine interaction technology can control various devices and services using human speech [14]. Speech synthesis mechanically produces human speech using a computer and is an active research focus due to the development of automation technology using deep neural networks. Recent synthesis technology that enables speech to reflect a specific tone has been developed to a level that makes various types of technical utilization possible, such as emotional expression [15]. Dialogue management receives text sentences from recognized human speech, understands the speaker’s intentions, and responds appropriately. Goal-oriented dialogue processing technology, which understands speech and maintains conversation on a related topic to accomplish a specific purpose or task, is used to implement smart speaker dialogue.

Based on voice recognition technology, smart speakers have voice interactions that can extract meaning from human voices, which leads to understanding between humans and machines. We call this a voice-user interface [16]. This voice interaction consists of voice recognition, voice synthesis, and dialogue management technologies. A system that provides voice interaction based on these technologies is called a conversational agent. Beyond their use as an interface, voice interactions with smart speakers are the key interaction determining the user’s main service experience. From the user's perspective, this is the most natural method of human interaction with machines.

This study was conducted using NUGU (SK Telecom [17]), a smart speaker platform that is widely used in Korea. According to a survey by SK telecom, there are currently about 1 million consumers with NUGU speakers. In August 2019, 670 million individuals used a NUGU speaker at least once a month (monthly active users [MAUs]), and in the first half of 2020, there were an average of 6 million MAUs of NUGU [18]. NUGU analyzes users’ natural language (voice or text) requests, interprets them, and then provides users with information or related services. For detailed information, see Multimedia Appendix 1.

In this study, the smart speaker–based MMT is called “Brain Toktok” or “Brain Opal.” It was commercialized on July 30, 2020. Users of NUGU speakers can subscribe to it [19]. Brain Toktok is based on a multistrategy MMT program developed by Youn and colleagues [7]. Previous offline MMT studies were conducted by trained psychological experts and MMT has already been well-validated [5-7]. The offline MMT program applied several memory strategies based on the metamemory concept. The program consists of 10 weekly sessions. Each session is comprised of three parts, and it takes 90 minutes to complete a session. The goal of the first part of each session is to strengthen meta-knowledge. The second and third parts of each session focus on meta-monitoring, meta-judgment, and memory training via the practice of external and internal strategies. Brain Toktok was developed through a series of analytic stages (content, learner, technology, and environment analysis) and design stages (information, interaction, synchronous, and evaluation design), followed by expert consultation. In this process, the goal of Brain Toktok was to focus on the steady participation of users in training rather than to increase the user’s correctness rate. Finally, 11 cognitive training exercises were developed. Three programs are randomly selected for use among the 11 cognitive training exercises possible per session. The provided training program was selected randomly without a fixed standard, and the content provided for each session was not recorded separately. Users can respond to multiple-choice or open-response questions specific to the programs provided and consist of more than 100 questions per program. For detailed information about the individual programs, see Multimedia Appendix 2.

This study was conducted from October 2019 to March 2020. Individuals who agreed to participate were assigned into two groups using convenience sampling. The control group did not receive any training; neuropsychological measures were simultaneously taken in the control group and in the training group before and after the training was completed in the latter group. As shown in Figure 1, 6 control group participants were excluded from the study because they did not complete the postevaluation. At the start of the study, participants in the training group were instructed on the use of the smart speaker. Neuropsychological measures (ie, SGDS, SMCQ, MMSE, RI, Digit Span Test, and fluency test) were assessed twice during the study: before the training (pretraining evaluation) and after the training (posttraining evaluation).

The training group received motivation reinforcement education training at the beginning of the smart speaker–based MMT. Group education took place for 1 hour and 30 minutes, and lectures were conducted on the definition, purpose, effectiveness, and necessity of metacognitive training. Afterward, participants had an opportunity to provide feedback. The 8-week training program was designed to include training sessions 3 times a day, 5 days a week (weekdays), for approximately 15 minutes. The patient flexibly set the time without setting a specific time for program participation. During Brain Toktok, the program may be suspended if there are external factors such as personal circumstances of the user or instability of the internet connection. If the program is interrupted due to external factors, it starts again from the beginning when participation resumes. To check the training group’s compliance, the number and duration of times each patient turned on the program and the number of times they completed the session was stored in the server. Participants who did not complete a training session in a given day received an encouraging program message. In the control group, the monitoring system was not implemented as in the experimental group, but the control group was registered in the local welfare center and the situation of the control group was continuously confirmed by the welfare center staff.

All participants provided written consent before study participation. The study was conducted following the Declaration of Helsinki and approved by the ethics review board of the Seoul National University Boramae Medical Center. Participants received a gift card worth US $85 as compensation for their participation in the study.

Focus group interviews were conducted with participants in the experimental group to assess their experience with and the effectiveness of the smart speaker–based MMT program. After completing the smart speaker–based MMT course, interviews were conducted with 28 participants who had agreed to the interview during the course. It was composed of eight open-ended questions, and for each question, the subject was asked to freely describe their thoughts about the item. An experienced clinical psychologist conducted the interviews and recorded both their observations and the participants’ responses.

Statistical analyses were conducted using SPSS software (version 20.0; IBM Corp). Differences between groups in demographics and neuropsychological measures (ie, age, education, gender, and MMSE score) were assessed using independent t tests or χ² tests. Repeated-measures analysis of variance (rmANOVA) was used to evaluate differences between the pre-and posttraining neuropsychological scores in the training and control groups. For model I, no adjusted factors were used. For model II, adjusted factors were used for SGDS and SMCQ scores. Because the SGDS and SMCQ scores varied between the groups at baseline, adjusting for the effects of SGDS and SMCQ scores allowed us to see the true effect of MMT on the memory test score. Statistical tests were two-tailed, with P<.05 indicating significant results.

Figure 1 summarizes the characteristics of the study participants. No difference was found in the drop-out rate between the two groups (MMT group: 9 dropouts, control group: 6 dropouts; χ²=0.65; P=0.42). All but 6 participants in the MMT group achieved more than 80% compliance with the attendance requirement. The 6 subjects who failed to achieve more than 80% compliance did so because of personal reasons (ie, hospital admission, sickness, preparation for moving, etc). The details of compliance are shown in Multimedia Appendix 2. Table 1 summarizes the demographic and clinical characteristics of the MMT and control groups. There were no differences between the groups regarding gender distribution, age, education, or MMSE scores. There was a difference in the SMCQ and SGDS scores between the groups. Therefore, we adjusted the SMCQ and SGDS scores in regression model II.

As shown in Table 2, we found significant interactions between group and score change for the tests of delayed free recall, fluency, digit span forward, and digit span backward. Because SMC may affect cognitive performance regardless of memory training, even after an improvement of depressive symptoms, the change in the SGDS score was controlled for by treating it as a covariate. After controlling for the effects of baseline SMCQ and SGDS scores, the interactions remained significant (model II in Table 3). The change in verbal immediate recall score did not interact significantly with the group.

The scores for delayed free recall, fluency, digit span forward, and digit span backward were significantly higher in the group that participated in the MMT program for 8 weeks; there were no significant cognitive changes in the control group after 8 weeks (Tables 2 and 3, and Figure 2).

Users of smart speakers reported that they were easier to use and less burdensome than other information and communications technology tools such as computers and mobile phones. For example, one user reported, “The smart speaker was convenient because I did not need to press the keyboard or the button, and usually I didn’t use the cell phone very well. Rather, I used it more frequently than the cell phone.”

Although the smart speaker is a machine, users reported that communicating with it was like a human-to-human conversation: “Television is what you see, and you can't talk but the smart speaker is able to communicate.” Some users preferred to communicate with the smart speaker rather than real people, and they enjoyed it: “If I get it wrong repeatedly, people get easily tired. The AI speaker keeps listening and encouraging;” “Smart speakers feel like comfortable friends, sometimes like a family;” and “I don't like to communicate to people very much, but I enjoyed communicating to the smart speaker when I needed it.”

According to the study results, our smart speaker–based MMT program improved memory ability, executive ability, and working memory. After adjusting for SMCQ and SGDS scores, the effects were still significant. These results indicate that cognitive improvements may be achieved through the effects of pure cognitive training. Furthermore, smart speaker–based MMT, which can easily be carried out without trained experts, fostered high compliance. Hence, it offers a good cognitive training tool without the typical limitations imposed by location or time constraints. To the best of our knowledge, this is the first report on a cognitive training program using a smart speaker. We hope that this study offers a useful guideline for future research on cognitive training programs based on smart speakers.

Previous studies on the effectiveness of smart speaker–based MMT reported that healthy older adults achieved improvements in memory ability, verbal fluency performance, and working memory [7]. The results of this study are similar to those of previous studies. Users of smart speaker–based MMT may still experience the same cognitive improvements as people who undertake in-person courses conducted by experts.

In this study, the scores for delayed free recall and verbal memory were significantly higher in this healthy older adult population after the training. Decreased delayed recall scores are considered the best predictor of progression from a dementia-free period to Alzheimer disease dementia (ADD) [35]. A previous study reported that expert MMT was associated with significantly improved cognition in subjects with mild cognitive impairment (MCI) [6]. These results suggested that smart speaker–based MMT may slow not only cognitive decline in healthy individuals but the progression of MCI to ADD.

Among the improvements in executive functions, the improvement in verbal fluency was significant. Previous studies have reported that verbal fluency was significantly associated with social activity [36]. According to the US Census Bureau, in 2010, 30% of people aged 65 years and older lived alone at the time of the census. As people get older, their likelihood of living alone increases [37]. Due to the social conditions associated with aging, older people are more likely to be vulnerable to verbal fluency problems. Smart speaker–based MMT can effectively support the vulnerable cognitive functioning of older adults who live alone. Also, older people who live alone and repeat relatively simple life patterns have a higher risk of depression and poor cognitive abilities compared with older people who do not-live alone. Smart speaker–based MMT is conducted through conversations with users, which can serve as a new stimulus for the elderly user who lives alone [38]. According to our focus group interviews, smart speaker users reported that communicating with it was like conversing with another person, so our smart speaker–based MMT may help those who live alone.

Digit span forward and backward tests are the most commonly used tests in clinical neuropsychology to assess working memory capacity [39]. Working memory is a key focus of memory improvement strategies. It can be trained through smart speaker–based MMT and then employed to increase cognitive reserve to prevent dementia. Furthermore, working memory is one of the core mechanisms involved in higher-order cognitive abilities, such as problem-solving, fluid intelligence, and reading comprehension [40,41]. It is also one of the cognitive processes that suffers a clear and linear decline with aging [42,43]. In this study, digit span forward and backward tests were significantly improved in the group that participated in 8 weeks of training. Smart speaker–based MMT may help maintain or improve higher-order cognitive abilities by enhancing working memory capacity.

There are some limitations to this study, the first being the small sample size. However, this is the first time a cognitive training program has been developed using the smart speaker, so this paper provides meaningful guidelines for future studies. Second, this study did not adopt an active control design, and the control group took no action. Further studies should use an active control design. Third, the SMCQ and SGDS scores were not homogenous between the groups. To remove the effects of SMCQ and SGDS scores, the SMCQ and SGDS values were adjusted during the analysis. We found that the efficacy of the MMT program using the smart speaker was not affected by the SMCQ and SGDS scores. Therefore, the results of this study are reliable.

Fourth, neuropsychological tests performed before and after the experiment are not a direct measure of metamemory function. Still, we used neuropsychological tests to assess the functioning of general memory areas. Specifically, the test of cued recall is the most effective test of memory function status. We also employed the cued recall test to discover the effectiveness of smart speaker–based-MMT. Although a direct measure of metamemory function was not used, the purpose of this study was to examine the enhancement of memory function through MMT. We consider the study methodology to be suitable for its purpose. Finally, we assigned participants to the two study groups by convenience sampling. We are aware of one other recent study that has addressed this issue [44]. Although cross-sectional in design, the previous study found few significant differences when contrasting the demographic characteristics and psychological performance of older adults in a random sample instead of two convenience samples. Specifically, the previous study observed no differences in word list recall results between the random and convenience samples recruited for a study of memory and aging. In further studies, to increase the power and proof of the effectiveness of smart speaker–based MMT, a large sample size and a more elaborate research design such as a randomized controlled study with an active control design should be implemented.

Smart speaker–based MMT, without location requirements and time constraints, may address older adult memory problems and possibly improve their quality of life by helping them cope with the cognitive issues associated with aging. Our smart speaker–based MMT may also be useful in delaying the onset of dementia.

Our study’s strength is that it is the first report on a cognitive training program using a smart speaker. In the future, the use of smart speakers will be more commercialized, expanding with the paradigm shift to “social distancing.” The problem of memory deterioration due to the increase in the elderly population will become more critical. In the future, the combination of a smart speaker and the memory training program will be of great help to older adults. We hope that this study offers a useful guideline for future studies on cognitive training programs based on smart speakers.