Keywords

1 Introduction

Smart speakers (SSp) are voice-controlled hardware devices connected to the Internet. Well-known examples are Amazon Echo or Google Home speakers. Positioned in a user’s home and equipped with artificial intelligence features and natural language processing, they are designed to support users in their daily routines. Users can play music, set reminders, or use search engines [1]. To function as anticipated, SSp require large amounts of personal data and various permissions to process (personal) data [2, 3]. This poses threats to security and privacy (sec/priv) [4, 5]. In recent years, researchers have discovered a variety of security issues that can lead to such things as eavesdropping, remote tampering, and unauthorized access [6]. As SSp use natural language processing and are activated via voice commands, they are in constant listening mode for recognizing “wake-up-words” [4]. The corresponding privacy concerns are diverse in nature, given that SSp consist not only of hardware but also of software, skills, remote cloud services and diverse default privacy settings [2, 5]. Sec/priv Concerns are exacerbated by the fact that SSp have found their way into the most private areas of users’ daily lives [1, 2]. To date, little research on the sec/priv of SSp, including behavioral aspects, has been carried out [5, 7].

Prior research suggests that demographic factors, personality traits, the nature and duration of the interactions with smart speakers, the speaker models, and the individual reasons for the acquisition and utilization of smart speaker devices influence the importance users attribute to security and privacy [7]. Based on data from a user study with N = 1.000 participants from Switzerland, this paper explores in detail how these factors influence users sec/priv concerns and to what extent this leads to the adoption of protective measures (PM). In order to gain an understanding of how convenience is off-set against security and privacy [2], four combinations of factors are distinguished and examined.

The remainder of this paper is structured as follows: Sect. 2 presents theoretical findings on users’ perceptions of sec/priv and the PM that are taken, which are used to generate six hypotheses to answer the research question. In Sect. 3, we present our methodology. In Sect. 4, we present the results and discuss them in Sect. 5. We conclude with an outlook on future research opportunities in Sect. 6.

2 Prior Research

2.1 Privacy and Security Threats of Smart Speaker Technology

SSp require permissions, including: access to user’s phone contacts to enable interactions with them, or to access account data to make purchases [2, 3]. The type and amount of data used differs between applications and use cases. Some data are collected and compiled into profiles without the user being aware of this [4]. This can give rise to privacy concerns, as the date and time of a query, for example, allow conclusions to be drawn about the person’s daily routine. In addition, account data, such as name, place of residence and payment methods are linked to interactions with the SSp. Speakers and their applications record the voice pitches not only of their owners, but also of anyone talking in the vicinity of the device at the time of recording [4]. Skills interactions may also require additional user data, such as date of birth for reminders or blood type for health apps [4]. To interact with other smart home devices, SSp exchange various types of (personal) data with other devices and servers [8]. This again allows for analyses and inferences to be made about user behavior [4], for example, about their physical health or the relationship between people in a shared household interacting with the speaker [4].

Security threats posed by SSp include unauthorized access, remote tampering or eavesdropping [6]. Mistakes in accuracy of voice verification may have comparatively benign consequences, such as a child making purchases by speaking to an Amazon Echo [9]. Other types of attacks can cause much greater damage, e.g., when attackers obtain physical access to the device. For example, it is possible to obtain personal data by inputting voice recordings [7]. Zhang et al. [10] find that hackers can emit ultrasound commands to gain complete control of the speaker without authorized users even noticing. By combining this attack with resampled audio snippets of legitimate user commands, the authors were able to fully bypass the biometric voice authentication [10]. Other attacks invoke malicious add-ons for SSp (so-called “skills”), that may operate permanently and stealthily [11]. Most SSp provide no other way to authenticate the owner than through the wake-up-word [4], and where alternatives exist, they may be poorly secured (e.g., PIN mechanisms allowing for an infinite number of re-tries [12]).

Quite frequently, a compromised SSp is a single point of failure [4] for gaining access to other smart home services, which can all be interrupted in a single strike [5, 8]. This can lead to (physical) safety-critical situations, if the speaker is connected to a smart door lock [13]. For these reasons, the sec/priv features of SSp have become an important topic in recent academic research [4, 5, 8]. Researchers are investigating technical PM [14, 15], the usability of frameworks for end users [16] and recommendations for SSp and skill providers [7].

2.2 Privacy and Security Perception and Concerns of Smart Speaker Users

Sec/priv perceptions and the concerns of end users in traditional IT systems are widely researched in computer science. For example, [17] identifies antecedents of privacy concerns as early as the early 1990s, [18] hypothesized that security concerns result from corporate actions, industry risks and individual awareness.

In the research domain of SSp, the concerns are more diverse [5] and arise from users’ perceptions of not having control over the devices’ various data flows [6]. [2] use data from online reviews and a customer survey and find seven different types of privacy/security concerns. These are (1) device hacking (69%), (2) personal data collection (16%), (3) constant eavesdropping (10%), (4) recording of private conversations (12%), (5) the fundamental disregard for privacy (6%), (6) data retention (curiosity about how and where information is being stored) (6%), and (7) the “creepy” behavior of devices (4%) (verbatim quote from a respondent). According to [5], security concerns can relate to different stakeholder types: co-users, service provider, authorized (law enforcement) and non-authorized (hackers) external actors. From a technical perspective, built-in cameras and microphones cause the highest user concerns [1]. In addition, users’ concerns result, for example, in the rejection of online shopping via a SSp [7] or controlling smart door locks [13]. Specific features of SSp also raise concerns, in particular the always-listening mode [2, 19].

Several factors influence user concerns. These include users’ demographics and personality, the nature and duration of their interactions with the speaker, and their motivation to acquire and use a speaker. These influences are discussed in more detail in the following paragraphs.

Context of Use:

There are different groups of users who interact with SSp: active users who purchase and set up a speaker themselves, users who jointly purchase, set up and use a speaker, bystanders who do not actively use a speaker but are affected by a speaker set up by family members or flat mates and non-users [20, 21].

Literature allocates different levels of privacy concerns to different users, based on individual characteristics [17]. In the field of SSp “a bystander’s perceived privacy can diverge from their actual privacy” [21], if they even recognize they are a bystander at all. Furthermore, it is reasonable to assume that users who actively install and use a SSp are less likely to have sec/priv concerns, because they have a sense of control over the use of the SSp [6]. Thus, we hypothesize that users who use their speaker by themselves will use fewer PM than users who share a speaker or a bystander (H1).

Further research found that users of different gender, age, and education level perceive SSp differently [22]. Thus, older individuals would need to weigh up privacy concerns against the increased autonomy they may get from using them [1]. In general, users with higher levels of education are less concerned [23]. In [22], the authors introduce three different relationships between user groups and the speaker. They find that female users are significantly more likely to view SSp as a friend (54.0%) or pet (20.7%), in contrast to male users who are more likely to view SSp as a servant (18.3% vs 11.5% females). Such differences in relationship perceptions may explain the extent to which users perceive the need to protect themselves and/or others against threats to sec/priv. Therefore, we assume that age, gender and level of education have a significant impact on PM (H2).

Usage Changes Over Time:

Users change the way they interact with a speaker over time [5], with automation tasks in particular increasing until the 150th day of use [24]. At the same time, research indicates that users’ sec/priv concerns change over time [25]. For example, in [6], users of SSp are found to develop concerns due to the increasing capabilities they associate with the technology. Other research finds that users slowly overcome their fear of using them over time [9]. Therefore, we define our H3 as follows: Users privacy PM change depending on the length of ownership.

Emotion Recognition Through SSp:

There is a future feature that researchers believe has great potential: Emotion recognition through both the voice and spoken content analysis [26]. Amazon has already announced plans to include these features in its Echo speakers [27]. However, privacy concerns are regarded as a major obstacle regarding the deployment of the technology [28]. To understand whether emotion recognition has an impact on privacy concerns and PM in SSp, we hypothesize that users who are against the use of emotion recognition will tend to be more protective (H4).

Acquisition Reasons Influence Privacy Perception:

There are different reasons, why users acquire SSp: cheap prices [29], donation [2, 9], or social influence from peers and media [9, 30]. For seniors in particular, (grand) children might set up the SSp. However, the main drivers are the perceived usefulness [9] and interest in the technology [25]. We hypothesize that there may be differences in users’ perception of sec/priv concerns between these groups, which could translate into differences in PM. For example, a user study [9] finds that receiving a speaker “as a gift triggered a privacy consideration”. However, to our knowledge, there is no work that investigates the relationship between acquisition reasons and PM. To investigate this further, we hypothesize that different acquisition reasons have an impact on whether users have privacy concerns and whether they adopt PM (H5).

Reputation and Perception of Smart Speaker Manufacturers:

Globally, Amazon, Google, Baidu, Alibaba and Apple are the market leaders in terms of SSp sold [31]. In Switzerland SSp have been late to the market and in some cases have not been launched at all. This is mainly attributed to the difficulty in understanding Swiss dialects [32]. At the time of writing, only Google and Sonos devices are officially available in Switzerland. Amazon Echo and Apple’s HomePod are only available via resellers.

Technically speaking, the speakers from Amazon, Apple and Google have similar data protection functions, although they are equipped with different hardware. However, Google stores the least voice recordings [31] and Apple does not use the data to create a marketing profile, which leads different authors to consider the HomePod as the best choice in terms of privacy [31]. A survey of UK users finds that they have slightly more privacy concerns towards Amazon Echo (2.62 on Likert-5) than Google Home (2.48) [5]. These concerns in turn result in a variety of PM [9, 33]. Further, [1] found that trust depends on “established, mostly positive relationships with these companies”. Thus, we assume that there are notable differences between SSp and hypothesize that users show different PM for speakers from different manufacturers (H6).

2.3 Protective Behavior and Measures in Smart Speaker Usage

There is a variety of research streams in SSp sec/priv protection. These include both those of technical and non-technical nature [5, 33]. However, [7] finds that SSp users mainly engage in non-technical PM. Although PM are rarely taken by end-users [5], their behaviors are still manifold. These range from reviewing and changing privacy settings [5], to using multiple profiles [9], to giving misleading information or avoiding private conversations [3, 5]. There are also more pragmatic approaches such as muting, covering, turning off, unplugging speakers or speaking quietly near them [2, 5]. At the same time, however, users report that these PM are not very convenient because they regularly interfere with the purpose of using the speaker, which may include spontaneously giving commands without first plugging it in [9]. Therefore, deleting recordings [1, 5] is regarded to be the most prevalent PM [25]. Returning a speaker was the final resort reported by users [2].

Finally, we divided the different users into four clusters of user groups, as shown in Fig. 2. (1) Users with no concerns and no PM, (2) users with concerns and PM, (3) users without concerns but PM nevertheless. Finally, research on the Privacy Paradox in a SSp domain [5] suggests that there can be (4) users with concerns but no PM.

2.4 Research Question and Research Model

Based on the insights from the reviewed literature, our research question (RQ) is: Do users protect themselves from the security and privacy threats posed by smart speakers, and what factors influence the PM they reported using? We address this RQ with the research model in Fig. 1. It includes six hypotheses and considers users’ statements to take sec/priv PM as a dependent variable.

Fig. 1.
figure 1

Research model

Full size image

3 Methodology and Survey Design

The insights from the literature reviewFootnote 1 in Sect. 2 served as input for the survey questions. These insights suggest that an initial understanding of the reasons for use and acquisition is necessary, that emotion recognition is a major concern and indicated a variety of user protection behaviors and mechanisms. Accordingly, the questionnaire involved in total thirteen questions about two main topics: (1) SSp acquisition and usage and (2) related sec/priv protection aspects (see Fig. 2). Each had on average six possible answers and an open-text field. These included single and multiple choices, open-text fields and mainly Likert scales (agreement, probability, appropriateness and importance). The questionnaire in German, Italian and French can be found in an online appendix (https://doi.org/10.24406/fordatis/211). Additional demographic information included gender, age, canton of residence and education, as well as Swiss-specific characteristics such as the linguistic region. The data was analyzed using the SPSS 25 statistical software suite. The overall structure of the questionnaire is shown in Fig. 2.

Fig. 2.
figure 2

General structure of the questionnaire

Full size image

The questionnaire was provided in German, Italian and French. Together with a set of six experts (business informatics, political science, data analytics and surveying background, as well as a proofreader), we iteratively reviewed and adjusted the questions, scales, and grammar. In a further step, we executed a pre-test with n = 55 participants in early October 2021. After a final round of adjustments, the questionnaire was provided to the entire sample. The survey itself was carried out by IPSOS, a professional survey company that was responsible for sampling. They used acknowledged standard mechanisms and programming in order to implement the survey. IPSOS used one of their well-tested samples that ensures at recruitment stage, that participants are human and only apply once. To ensure a high quality of answers, a straightlining test was conducted to identify robotic responses and respondents who gave the same answer patterns or polar opposite statements. Furthermore, the answering speed was measured and respondents were given a notification if they were too fast (or in good time) halfway through the questionnaire. However, no entry had to be removed on the basis of the above criteria. IPSOS carried out the analysis, e.g., using Chi2 tests for calculating significance. Data cleaning was not needed for a majority of the fields, as categories were predefined as Likert scales. Open text fields were analyzed manually (as the number was not too high) and categorized by the experts. For example, open text questions about “reasons for purchase” resulted in “curiosity”, “prestige” and “increased autonomy due to walking disability”. The online survey with N = 1000 representative participants in Switzerland was conducted in October 2021 (Table 1). These included German, Italian and French speaking respondents from both rural and urban areas.

Table 1. Demographics (in %, N = 1000)
Full size table

4 Results

In this section, we give insight into the general usage of SSp (Table 2) and present the correlated sec/priv PM stated by the respondents. We aggregate these PM depending on different demographic characteristics (Tables 3, 4, 5, 6, 7 and 8).

4.1 Usage of Smart Speakers

SSp are present in 36.7% of all households in Switzerland. Of these, a majority possesses a Google Home (45.5%). Amazon Echo is present in 25.6% of the households, followed by Apple HomePod (20.3%), others (8.7%, mostly devices from Swisscom), and 10.4% possess more than one SSp (Table 2, significant differences in the sample structure at p < 0.01 in gray). More male (43.3% male, 30.0% female) and young respondents between 16 and 34 years old (52.8%, p < 0.01) reported having and using a SSp in their household. In general, gender, age and education level have significant influence on SSp ownership and use (p < 0.01, respectively p = 0.04 for education). Area of residence (urban or rural), on the other hand, has no significant influence.

The results show that 20.4% of users set up and use the device themselves. A shared-usage is present in 10.8% of the households, with a third person, such as flatmates or the children of elderly persons, installing the speakers in 5.5% of cases.

We find significantly younger users (16–34 years) in the category own use (30.4%) and shared use (16.1%) than users aged 35 and above. Participants with a university degree (25.6%) are more likely to use and install speakers by themselves than users without a university degree (18.2%). French-speaking users use SSp significantly less (13.4%) than Italian-speaking (30.0%) or German-speaking (23.1%) ones. A possible explanation could be found in [34], where 67% of French respondents stated that they do not see a need, and 59% have the kind of privacy concerns that prevent them from buying an SSp.

Table 2. Distribution of demographics and SSp usage with the context of use (in %)
Full size table

4.2 Factors Influencing Protective Measures

Figure 2 shows three different contexts of use (own, shared and bystanders) as well as four combinations of user concerns and PM. For the correlation analysis, the variables “no concerns/no protection” and “concerns/no protection” were combined (CombNP). The results in Table 3 show a very low Pearson correlation (PC) of −.045 for own usage and CombNP. However, there is a significant (p < 0.01) negative correlation between own usage and “concerns/no protection” (C/NP, −.147). A significant correlation exists between bystanders and concerns (.148 and .149 at p < 0.01). There is no difference between bystanders who take PM and those who do not. A binary logistic regression shows a Cox & Snell R2 of .023. Age serves as a control variable, as it has no significant influence (sig = 0.430) it can be disregarded.

Table 3. Pearson correlation for context of use and characteristics of concerns/protection
Full size table

The PC analysis of demographic factors and concerns/protection (Table 4) shows that gender is significantly (p < 0.01) positively correlated to C/NP (.168) and negatively correlated to NC/NP (−.101). Correlating genders separately shows that females (.168, p < 0.01) have significantly more concerns, but take no PM. For men (−.171, p < 0.01), the correlation with this observation is significantly negative. Other factors show only a very low correlation.

Table 4. Pearson correlation for demographics and characteristics of concerns/protection
Full size table

Correlating the usage period with concerns/protection, Table 5 shows that users who possessed a SSp for less than one month are significantly unconcerned and took PM to a lesser extent (.163, at p < 0.01). This PC completely reverses when a speaker is used between one and twelve months (−.162 at p < 0.01). PC values decrease sharply as the duration of use increases to over a year. This observation also holds for the combination of the two behaviors without protection (CombNP). However, this is largely explained by the PC value for NC/NP. A usage time of more than one year shows little correlation (absolute PC values < .09). However, there is a negative correlation (−.108 at p < 0.05) observed between a usage time of 1 to 3 years with regard to having concerns and taking PM (C/P). A binary logistic regression shows a Cox & Snell R2 of 0.051, with gender and age serving as control variables. Both are not significant (0.4 and 0.68). In this regression, duration of use from one to twelve months is significant at 0.055.

Table 5. Pearson correlation table for usage period and concerns/protection
Full size table

Users were asked about their reasons for acquisition using a 5-Likert agreement scale. Spearman correlations (SC) for own usage are presented in Table 6Footnote 2. Price significantly influences users concerns and whether or not they take PM (.204 at p < 0.01). Furthermore, price also leads users to take PM, even when there are no concerns (C/NP −.169). Persuasion appears to have a positive influence on taking PM (CombP .160 at p < 0.05). Persuaded users have fewer concerns but take PM at a significantly higher rate (.176 at p < 0.05). This observation is supported by the significant positive SC of .162 for C/NP. Logistic regression shows that the control variables gender and age are not significant (sig 0.93 and 0.80). Persuasion is significant (sig 0.02). High self-interest in using a SSp significantly leads to having concerns and taking PM (.142 at p < 0.05). Users, who purchase a speaker because friends use one show less concerns but still take PM (C/NP −.176 at p < 0.05). The expectation that the speaker will make life easier does not influence concerns or PM. The results differ greatly when it comes to shared-use. There is no significant effect on taking PM (or not). Only when the device is purchased by someone else (e.g., because the user is inexperienced) is there a significant influence on not having concerns but still taking PM (.198 at p < 0.05).

Table 6. Spearman correlation for reasons for acquisition and concerns/PM for own usage
Full size table

Table 7 shows SC for different speaker models. For the context of use “all” there is a positive correlation between “other SSp” and CombNP (.109). The only significant negative correlation is for “other SSp” and NC/P (−.117 at p < 0.05). For shared usage, there is a significant positive correlation between Google Home users and PM (.225 at p < 0.05). This follows from the significant positive correlation at NC/P (.244 at p < 0.05). Own usage does not show a significant correlation. However, Google correlates to C/P (.120). There is also a positive correlation between “other SSp” and CombNP (.102).

Table 7. Pearson correlation table for SSp manufacturers and concerns/protection
Full size table

After asking for general concerns in SSp, the survey asked about their opinion on the use of emotion recognition in SSp. Those users who oppose its use (Table 8), are significantly more likely to have concerns and take PM (.371 at p < 0.01). Those who opposed its use and who have no concerns and do not take any PM (−.105) or take PM (−.144) are negatively correlated. However, the number of respondents to NC/P is small (61 respondents). Users having concerns but not taking PM does not give an indication whether they are opposed to it or not.

Table 8. Pearson correlation for those opposed to emotion recognition and having concerns/protection
Full size table

5 Discussion

While SSp offer convenience, they pose sec/priv threats which lead to concern on the part of users and to the adoption of various PM. Our results reveal that these PM vary by context of use, gender, usage duration and opinion towards emotion recognition. However, the speaker model and manufacturer is not relevant in this aspect. Further, education, linguistic region and whether users live in urban or rural areas also show no influence on the PM taken.

Users interact with a speaker in different contexts (own, shared or as bystanders). Those who actively purchase and set up a SSp show significantly less concerns and PM. However, considering PM without concerns (CombNP), own users adopt slightly less PM than the other groups. This could be due to the possibility that own users are (or at least believe they are) sufficiently informed to have no concerns and therefore do not need to take PM [2]. Bystanders are less concerned than own users and hold a similar degree to which they show and to not show PM. This could be attributed to their different familiarity with the technology and awareness that a SSp is present [1]. As own users adopt slightly less PM, we see H1 as partly confirmed.

Our results confirm that gender has significant impact on whether or not PM are taken. This is not a new observation in general sec/priv research [17]. However, our results show that sec/priv in SSp must be considered differently than in other computer systems [2, 5]. Women are significantly more concerned about sec/priv in SSp contexts than men [17]. Although age is highly significant with respect to the adoption of SSp, results show no notable influence on PM. This is interesting, as younger users tend to have less privacy concerns [23]. Furthermore, education, linguistic region and whether users live in urban or rural areas also show no influence on the PM taken. This is somewhat surprising, as particularly educated users could have shown greater awareness and thus more concerns and a need for protection [1, 23]. Thus, our results indicate differences between sec/priv research in SSp compared to other domains. Therefore, we consider H2 to be only partially confirmed.

Novice users (usage for less than one month) have less concerns and therefore take PM to a lesser extent. This effect is completely reversed after this initial period within the first year of use. Explanations for increasing concerns and PM after one year could be that users start with a playful instead of purposeful use, first familiarize themselves with the technology, learn to take PM over time [6] and perform more automation tasks (e.g., with IoT devices) over time [24]. Another explanation could be the increasing dissemination of media news about data leakages [2]. The analysis revealed a negative correlation between a usage time of 1 to 3 years and C/P. This means that participants who use a speaker between 1 and 3 years do not show a higher level of concerns or PM. This could be due to a saturation of information flows or a rising awareness within the first year of usage, which makes taking further action obsolete for these users. Since we find significant changes in the three periods of use examined, we accept H3.

Literature suggested that emotion recognition is a strong reason for privacy concerns [28]. Our results can fully support this observation in the domain of SSp, which is why we consider H4 as verified. Opponents of emotion recognition are more likely to raise concerns and take PM.

When examining the influence of different reasons for acquisition, the results show that persuasion leads own users to take PM to a greater extent, but at the same time they have less concerns. Similar results are found for own users who buy a speaker because friends are using one (see C/NP in Table 6). This can be interpreted as passive influence and supports previous findings [5]. The group of NC/P seems contradictory at first glance. However, there could be several explanations: A general cautiousness, a great sense of privacy toward guests or the fact that users with high privacy literacy are largely less concerned [35]. This group however, has no significant correlation with age, gender or residential area in our analysis. The very different situation for shared usage requires further research. At this stage of research, one possible explanation could be that there are only half as many cases of shared usage compared to own use in our data set. The low price of a speaker significantly causes users to have concerns and take PM. This confirms prior studies found in [36], as users believe audio recordings to be used as a monetization strategy. All other reasons for acquisition show substantially lower correlation values. Since the values for PM (CombP) in Table 6 are very different between the different reasons for acquisition (with some being significant and others not), we provisionally confirm H5. However, a more detailed investigation of the relationships with regard to this would certainly be useful.

The literature review found that manufacturers follow different data protection practices, although their SSp show very similar privacy settings options for PM. Whereas a survey of UK users finds Echo users to have slightly more concerns, the results of our study show little correlation between manufacturers for the sample of all users. Shared usage, however, significantly increases the amount of PM for Google Home. This could be due to the higher number of Google devices in our dataset. These observations lead us to reject H6.

Our results have implications for research and practice. They investigate factors that influence taking PM in the field of SSp and contribute to this relatively new research field. With our results, we want to contribute to the ongoing discussion about the factors which influence sec/priv concerns and PM, using the SSp domain as an example. We confirm prior findings that SSp sec/priv is not completely transferable to other computer systems.

Our results provide two insights for manufacturers. First, in regards to what factors influence the purchasing of SSp and second, what makes users concerned and take PM. As gender, age and education level influence the acquisition of SSps significantly, more targeted marketing is advisable, especially for females, as they are more concerned than males. To enhance global sales, manufactures should focus on the French population, as they use speakers less often. Furthermore, the results show that shared use is quite common, but also leads to concerns and PM. Thus, manufacturers should develop (priv/sec) features that allow for separated usage with different settings/accounts and also for a guest mode [1]. To do so, a conversation with the SSp at setup-stage could be advisable to ask specific demographics to automatically suggest pre-defined privacy settings which are then reviewed together with the user. Our results show that price is not only a reason for purchase [1], but also that a cheaper price leads users to take more PM. Because data privacy is a major concern for some users, manufacturers could focus on more privacy features as a unique selling point. On a regulatory level, it is advisable to decide on a common standard of priv/sec default settings that users can rely on. Further, a common standard for PM could also be beneficial for manufacturers, as more trust leads to fewer concerns and thus more purchases.

Finally, our work is subject to two limitations. First, there are only weak correlation values for several questions. This is due to the low number of subsamples n we had to work with for several cases, despite the representative sample of N = 1000. Second, a separate and quantitative consideration of concerns and PM, e.g., using Likert-Scales, in the survey design would have allowed for a more differentiated investigation, which could also include an analysis of specific PM. Consequently, future work could investigate additional influencing factors (e.g., trust, regulations and personality) and should distinguish between concerns and PM. All of this could be done additionally with specific PM (e.g., muting or behavioral change).

6 Conclusion

This paper has investigated which factors influence sec/priv concerns and the adoption of PM for SSp. To this end, we first evaluated the literature on the areas of sec/priv threats posed by SSp and the PM taken by users. From this analysis, we then derived six explanatory hypotheses regarding potentially relevant factors influencing the taking of PM. We then conducted a representative survey with N = 1.000 respondents in Switzerland. Our findings reveal that taking PM is affected by five of the six factors: Context of use (partially), usage duration, gender (but not age and education level), reasons for acquisition, and opinion towards emotion recognition. Different speaker models, however, have no influence. Our results indicate that taking PM and sec/priv concerns in the context of SSp are different from the predictions found in the sec/priv literature with respect to other contexts. This calls for further investigation of the relationships between PM in these contexts.