Exploring the Benefits of Conversing with a Digital Voice Assistant during Automated Driving : 1 A Parametric Duration Model of Takeover Time 2

1 The current study investigated the role of an in-vehicle digital voice-assistant (VA) in 2 conditionally automated vehicles, offering discourse relating specifically to contextual factors, 3 such as the traffic situation and road environment. The study involved twenty-four participants, 4 each taking two drives: with VA and without VA, in a driving simulator. Participants were 5 required to takeover vehicle control following the issuance of a takeover request (TOR) near 6 the end of each drive. A parametric duration model was adopted to find the key factors 7 determining takeover time (TOT). Paired comparisons showed higher alertness and higher 8 active workload (mean NASA-TLX rating) during automation when accompanied by the VA. 9 Paired t-test comparison of gaze behavior prior to takeover showed significantly higher 10 instances of checking traffic signs, roadside objects, and the roadway during the drive with 11 VA, indicating higher situation awareness. The parametric model indicated that the VA 12 increased the likelihood of making a timely takeover by 31%. There was also some evidence 13 that demographic factors influenced the TOT of drivers. Male drivers likely to resume control 14 1.72 times earlier than female drivers. The study findings highlight the benefits of adopting a 15 futuristic in-car voice assistant to keep the drivers alert and aware about the recent traffic 16 environment in partially AVs. 17


INTRODUCTION 1
The advances in vehicle automation allow the drivers to disengage from driving and become a 2 passive monitor of the system. However, the shift in driver responsibility from an active 3 operator to a passive observer in an automated vehicle (AV) leads to the loss of active task 4 engagement, thereby compromising drivers' alertness required to intervene at such critical 5 moments (1)(2)(3)(4). Such a decline in alertness, caused by 'task-disengagement' or 'low workload ' 6 conditions during automated driving, induces a passive fatigue state, resulting in loss of 7 awareness of the current traffic situation (1,5,6). 8 Passive fatigue is specifically a problem in partially autonomous vehicles (level 2 or level 3 9 automation according to SAE International, (7)), in which the system will issue a takeover 10 request (TOR) to the driver in situations that fall outside its capability. Ensuring a safe takeover 11 of vehicle control is therefore one of the major challenges for highly automated driving. Hence, 12 it is suggested that some form of system feedback is required during periods of automation, to 13 maintain driver alertness with appropriate situation awareness, prior to taking over vehicle 14 controls (8). 15

Automation and alertness 16
Automation in vehicles can lead to long periods of driver inactivity leading to loss of alertness. 17 Some studies showed a significant increase in symptoms of fatigue after only 15-20 minutes 18 of automated driving in an autonomous vehicle (AV) (2, 3,8). A study by Neubauer et al. (2) 19 found that 30 minutes of automated driving increased proneness to fatigue, indicated by the 20 mean Driver Stress Inventory (DSI). Gonçalves et al. (10) noticed that after only 15 minutes 21 of automated driving there was significant increase in Stanford Sleepiness Scale (SSS) rating 22 indicating increased sleepiness among drivers, leading to poor driving performance. Further, 23 Neubauer et al.
(2) found significant correlations between the fatigue ratings and lower task 24 engagement using the Dundee Stress State Questionnaire (DSSQ). Saxby et al. (6) observed 25 that the automation led to lower workload ratings on NASA-TLX scale(11) indicating low-26 workload. Wu et al. (8) reported a significant increase in eye blink duration and subjective 27 sleepiness captured through Karolinska Sleepiness Scale (KSS) after 30-minutes of automated 28 driving indicating fatiguing effects, irrespective of drivers' age. Vogelpohl et al. (3) noticed 29 that 20 minutes of automated driving led to same level of fatigue among drivers as experienced 30 after 40-50 minutes of manual driving, where fatigue was indicated using KSS sleepiness rating 31 and indicators such as yawning, blinking, eye-closure etc. Collectively, these studies suggest 32 that a period of automated driving lasting 20 minutes or longer is sufficient to significantly 33 lower the task engagement and cause a significant decline in driver alertness. 34

Takeover time 35
The need for human intervention in partially automated vehicles is conveyed through pre-36 defined alerts, called as a takeover request (TOR). The takeover time (TOT) is the response 37 time of drivers to the TOR. It includes both, the time it takes a driver to make an assessment of 38 the traffic situation, i.e. regain their situation awareness (SA) (12), and demonstrate their 39 readiness to drive by re-engaging with the driving controls (13,14). Many studies have used 40 the driver's response time to a hazard scenario at TOR as takeover time (8,13,(15)(16)(17).
However, in the case of a potential collision event, the response is often reflexive, typified by 1 sudden, emergency braking (14,17). Therefore, it would seem prudent to avoid a reflexive  2  braking response to measure the absolute time to motor readiness when a TOR is issued, and  3 therefore avoid a hazard scenario. This approach was taken by Merat et al. (18), who reported 4 that an average time to respond to resume steering and brakes in response to TOR was between 5 10 and 15 seconds, where the drivers responded at their ease in the absence of any hazard 6 detection event at TOR. 7 Gaze behavior and situtation awareness 8 Passive fatigue due to mental underload situations can reduce the visual attention of drivers 9 resulting in additional time required to regain self-alertness before building their SA in 10 response to TOR (3,17). Visual, or gaze behavior can be a useful method to explore the process 11 of allocating attention (19). Factors such as the duration and frequency of glances spent 12 checking the speedometer, the road ahead, and side and rear-view mirrors, which are all 13 associated with building situation awareness (12,13,16 help the driver to effectively regain situation awareness during a takeover (12,18,33). 22 However, providing traffic narrative at the point of takeover is likely to distract the driver and 23 could influence their driving performance. An alternative would be to provide additional traffic 24 information intermittently throughout a long journey. This is likely to improve driver alertness 25 and awareness may additionally reduce the takeover time. 26

Study motivation and hypothesis 27
This study aimed to examine how a digital voice assistant can help in mitigating passive fatigue 28 induced by automation and in improving situation awareness at takeover through traffic-related 29 information. Therefore, it is hypothesized that intermittent, traffic-related conversation with a 30 voice assistant (VA) will reduce delays in takeover time caused by passive fatigue and 31 disengagement from driving during highly automated driving. Secondly, we hypothesize that 32 the traffic status updates provided by VA prior to the TOR, will help redirect drivers' attention 33 to the road, traffic or traffic signages, as per the messages, which can be confirmed through 34 their gaze behavior. Thirdly, previous studies show the possible influence of various factors on 35 TOT (age, experience, gender, involvement in secondary tasks, etc.) in isolation. However, 36 they act as covariates to influence the takeover process. Further, the effectiveness of a VA 37 proposed here, also depends on the ease and interest of drivers to use such technology. 38 Therefore, this study focuses on modelling the TOT to determine how the presence of VA, 39 (either directly or indirectly by influencing the gaze behavior or other factors gender or 40 accustomed to use of voice assistants), can be effective in assisting takeover or resuming 41 manual control after automated driving. 42 1 The study was conducted using a medium fidelity, fixed-base driving simulator at the 2 University of Nottingham Human Factors laboratory ( Figure 1a). This driving simulator 3 comprises an Audi TT car located within 270 degrees field of view. Three inobtrusive cameras 4 were installed at different positions inside the car to capture drivers' hand and feet movements 5 in response to TOR, and any physical signs of sleepiness during the experiments (yawning, 6 extended blinks etc.). This simulator has been used in various previous studies (27, 31), and is 7 capable of providing an experience of driving level 3 automated vehicle. VA mainly conveyed 8 driving-related information to the drivers, and thus, aimed to enhance the process of (re)gaining 9 situation awareness, which was subsequently assessed using gaze behavior i.e. checking 10 mirrors, road in front or traffic signs/signals etc.factors that reflect the takeover time (TOT)  11 among drivers (14,17). The TOT was modelled using a parametric duration approach 12 illustrating the influence of VA and related situation awareness (as covariates) on TOT. 13

Participants 14
Participants were invited using flyers displayed around the University campus. A pre-driving 15 questionnaire was used to identify and exclude participants reporting any sleep-related 16 disorders and/or an Epworth Sleepiness Score (ESS) >16 (34). In addition, participants were 17 specifically instructed to refrain from consuming caffeine, mint or alcohol for a few hours 18 before the study and to take adequate sleep prior to the day of study. Thirty-one eligible 19 participants volunteered for the study. Three participants did not turn up for the experiment and 20 four dropped out partway through due to simulator sickness. The final data is reported from the 21 remaining twenty-four participants (Table 1). Each participant received due compensation for 22 their time. The study protocol was approved by the Faculty of Engineering Research Ethics 23 Committee, University of Nottingham, UK. 24

Experimental Set-up 25
The study involved a within-subject design with two driving sessionsone with and one 26 without the voice assistant. A bespoke scenario was created using STISIM Drive 3 software. 27 The route represented a transition from an initial two miles of two-lane urban road to a standard 28 UK dual carriageway ( Figure   In the drive with VA (employing vocal interactions), VA introduced its capabilities such as 9 providing surrounding traffic feedback, route navigation, event reminders or operating music 10 or radio to the driver prior to the start of the drive. Drivers could either respond to or initiate 11 Observer's set-up Driving simulator conversation with VA using natural, conversational language. To achieve this, a 1 comprehensive set of pre-recorded spoken messages were embedded in the STISIM scenario. 2 These were played by the experimenter as per evolving scenario conditions, for example, 3 changing speed limits, gap from leading vehicle, suggestions for rest/ refreshment spots, 4 expected traffic congestion etc. The first message was played after 5-minutes from the start of 5 the drive, based on the expected onset of fatigue symptoms after 5 to 7 minutes of automation 6 (8). Each participant received the same opening gambits, however, the follow up statements 7 differed slightly based on individual's response. For example, 8 VA: "There is a pedestrian crosswalk ahead. Please be engaged in the drive or would you like 9 to slow down?" 10 Driver: "yes", 11 VA: "Reducing speed. You are now driving at 'x' miles per hour.", 12 However, this part of conversation will end if the driver responds "no". 13 In situations where an appropriate reply was not available from the pre-recorded messages, VA 14 responded with error messages such as "sorry! no network connectivity at the moment to 15 perform this task", "sorry! This function is not available currently" (although in practice, these 16 were rarely used). VA initiated a new conversational exchange or topic at approximately every 17 3 minutes, and these would last for at least 30-60 seconds (29). There was no conversation 18 initiated during 60 seconds prior to the TOR, although VA had already informed drivers about 19 the upcoming change in the posted speed-limit and the approaching pedestrian crosswalk. 20

Takeover event 21
Prior to the test drives, participants undertook a practice drive involving multiple instances of 22 takeovers, so that they become familiar with switching controls from 'manual' to 'automation', 23 and vice-versa. The instruction to transfer control was intimated with a voice message followed 24 by three consecutive beeps indicating the precise moment of the transfer of control, thereby 25 avoiding any visual distraction. During the test drives, participants were pre-informed that they 26 may be required to resume manual control at a certain point, but otherwise, should relax (35). 27 After completing 1-minute into the drive, automation was engaged at a fixed point at 0.75 miles 28 (1.2km) in each test scenario. After approximately 25 minutes into the drive or at 21.6 miles 29 (35 km)participants received a takeover request (TOR). To emphasize the need of a timely 30 response to the prompted TOR without imposing a critical hazard, a construction zone was 31 created on the roadside at about 61m (200ft) from the onset of TOR with a gravel pile spilling 32 on the roadside (Figure 3). In addition, a red-light traffic signal with a pedestrian cross-walk 33 was presented at 152 m (approximately 500ft) from the point of TOR, and drivers were 34 naturally expected to apply brakes in response to the signal. During the drive with VA, drivers 35 received a notification few minutes prior to TOR, that they were approaching a pedestrian 36 cross-walk. In order to determine drivers' awareness of the road environment (i.e. their 'situation 9 awareness'), immediately prior to the TOR, glance duration and frequency were calculated for 10 defined areas of interest (AOIs) using semantic gaze mapping (with BeGaze 3.7 software). The 11 following AOIs were selected: external mirrors (side-view mirrors), rear mirror, road ahead of 12 the driver (windshield), roadside objects (obstacles), speedometer, traffic signs and signal in 13 line with similar studies (12,13,18,36). Visual behavior was analysed during the 60 seconds 14 prior to the red-light stop signal, which appeared approximately 195m (640ft) following the 15 onset of TOR voice message (17). 16

Takeover time(TOT) 17
The TOT or response time to takeover request (TOR) was calculated as the time from the start 18 of the stimulus (i.e. end of TOR voice message and start of beeps) to the time at which drivers 19 acquired motor readiness i.e. hands on steering, feet on pedals and looking ahead (or eyes on 20 the road). The time to resume steering and pedals were determined using frame-by-frame 21 analysis of videos captured during the drive, whereas the time to resume glances on road was 22 noted from the eye tracking. The maximum of the three times was noted as the TOT. 23

Subjective sleepiness and workload ratings 24
In addition to the visual indicators of fatigue, drivers rated their level of alertness using the 25 Karolinska Sleepiness Scale (KSS) (ranging from '1-very alert' to '9-very sleepy') and 26 cognitive workload using NASA-TLX workload index (increasing scale of 1 to 21), (2, 11, 27, 27 28). The subjective ratings were collected on three occasions during each drive: firstly, prior 28 to each drive, secondly, towards the end of automated drive (prior to TOR) and finally, after 29 resuming manual drive. To avoid any interference during the drive, the latter two ratings were 30

Construction zone
Signage collected at the end of each drive. Also, the experimenter manually noted the relevant 1 symptoms of sleepiness e.g. frequency of yawning and incidents of 'nodding off', when 2 automation was engaged. 3

Post-drive Questionnaire 4
Finally, a questionnaire was used to collect data such as driver demographics, exposure to 5 various in-car driver assistance systems (DAS) and voice-assistants, such as Google, Siri, 6 Alexa etc. At the end of experiment, drivers were also asked to provide their subjective 7 feedback on usefulness of VA. 8

ANALYSIS AND RESULTS 9
Dataset 10 The participant characteristics are summarized in Table 1. Most of the drivers already used 11 voice-based assistants for either route navigation or as a music player ( Figure 5), but did not 12 use these to stimulate any conversations, for example, voice-based web search etc. In this study, 13 individual responses to each suggested use (as listed in Figure 5) of a voice assistant (rated on 14 a 1 to 5 Likert scale) were summed. This provided a single covariate indicating the frequency 15 of using VAs rated on a linear scale varying from 1-20 (mean in Table 1). 16  4

Alertness measures 5
The mean KSS and NASA-TLX workload ratings were compared using paired-samples t-tests 6 ( Figure 6 and Table 1). The t-tests showed a significant increase in mean KSS scores during 7 automation (prior to TOR) from pre-drive in both the drives (Figure 6a). This indicates the 8 fatiguing effects of automation, which sustain post-takeover as indicated by the KSS ratings, 9 specifically in the absence of VA. However, the mean KSS rating during automation was 10 significantly lower in the drive with VA (t(23) = 3.391, p<0.005) indicating higher alertness in 11 presence of VA. Further, the NASA-TLX workload ratings were significantly higher during 1 automation when the drivers were accompanied by VA suggesting higher alertnessduring 2 the drive with the VA, confirming that automation for long periods lowers cognitive workload 3 and makes the drivers vulnerable to symptoms of passive fatigue. The paired comparisons of 4 visual indicators of fatigue during automation i.e. average pupil diameter was significantly 5 larger during the drive with VA, suggesting higher alertness with VA compared to the drive 6 without VA. The AOI data are compared in Figure 7 for the two drives. The percent time spent for each AOI 10 was calculated from the total glance duration of each participant. The time spent and glance 11 frequency data for each AOI is then compared using a non-parametric Wilcoxon signed rank 12 test across the two drives. Only four pairs showed significant differences as indicated in Figure  13 7. Participants spent significantly more time glancing at the rear-view mirrors in the drive 14 without VA (Mean= 3.55%, SD =±5.74%) compared to the drive with VA (Mean = 0.73%, SD 15 = 0.85%). Comparison of glance frequency showed that drivers directed significantly more 16 glances to the road in front, roadside objects (construction zone or parked vehicles on roadside) 17 3 suggesting that they were more alert and engaged with the driving task when accompanied by 2 the VA. 3 The comparison of the alertness indicators and AOI statistics provide preliminary evidence of 9 engaging effects of VA, which is likely to influence the takeover time. Therefore, a parametric 10 duration model, or survival analysis approach was adopted to quantify the contribution of these 11 factors on takeover time (26, 38). 12 3 Parametric duration modelling is a probabilistic approach to analyse the conditional probability 2 of the elapsed time until the event of interest, provided the event continues to time, t (38). In 3 this study, the event is defined as "gaining motor readiness as shown by hands-on-wheel, feet 4 on pedals and eyes on road" and the length of time to gain complete motor readiness in response 5 to TOR is the duration variable (T). The probability of resuming manual control after the 6 time 't' (i.e. after the construction zone appeared) is called the survivor function, S(t). The 7 hazard function, h(t) which is also called the instantaneous failure rate, gives the conditional 8 probability that the event will occur between the time t and (t+dt) provided the event has 9 continued for 't' or more duration (38). In this study, accelerated failure time (AFT) model was 10 used. An AFT model allows the covariates to rescale (accelerate) time directly in the baseline 11 survivor function (38). Here, as the probability of completing the takeover is likely to increase 12 over time, it indicates a monotone hazard rate that increases exponentially with time. Thus, 13 Weibull distribution is suitable to model the takeover time data, with scale-parameter (P > 0) 14 and location-parameter (λ > 0) is given by (38): 15 In the Weibull duration model, the hazard function and survival function are expressed as: 17 Here, the repeated observations were collected across the two drives with the same participants, 20 which can cause intra-group heterogeneities. To account for such heterogeneities, Weibull AFT 21 model with clustered heterogeneity and gamma frailty were developed and compared using 22 Stata SE-16 (at 95% significance level). Among all comparable models with the covariates 23 (variables related to glance behavior, driver demographics, drive condition, workload and 24 frequency of using voice assistants), the final model with clustered heterogeneity, with 25 minimum Akaike's information criteria (AIC) and Bayesian information criteria (BIC) values 26 (38) is reported in Table 2. The scale parameter p = 4.53 (>1) confirms that the hazard rate 27 increased with time. Table 2 summarizes the estimated exponential of coefficients (hazard 28 ratio) which directly represents the relative change in survival time duration with unit 29 increment in the covariates. The model results show that participants were likely to gain motor 30 readiness 31% quicker in the drive with VA compared to the other drive. There was a slight 31 influence of cognitive workload, which was relatively higher during drive with VA. Also, the 32 model results show that male drivers are likely to resume control 1.72 times earlier than female 33 drivers. In addition, individuals who indicated that they frequently used VAs are likely to take 34 3% less time to resume control. Higher annual mileage and checking rear mirror did not 35 influence the takeover time significantly. 36 The takeover probabilities were calculated for the two driving conditions (with and without 5 VA) at different TOTs in Figure 8(a, b and c). All other variables were either kept at their 6 reference category or corresponding means were substituted using Table 1 Automation relieves the driver from the task of assessing the traffic scenario and taking 8 required physical actions for driving, which results in low-workload conditions (1,5,6). 9 Therefore, as hypothesized, the intermittent conversational exchanges with VA during 10 automation led to a significant increase in driver workload as indicated by NASA-TLX ratings 11 (Figure 6b). The lower KSS ratings and larger pupil dia in the drive with VA, indicated 12 efficiency of VA in maintaining driver alertness during automation (27,39,40). Also, as noted 13 by the experimenter, none of the drivers were observed sleeping during the drive with VA, 14 whereas six drivers had short episodes of "nodding off" during automation in the drive without 15 VA, and were notably startled by the takeover request. It is concluded that the regular 16 conversational interludes made by the VA interrupted the monotony of the automated drive. In 17 combination with this, providing traffic-related conversations, such as informing drivers of the 18 speed limit, upcoming intersections etc. kept the drivers more engaged with the driving 19 environment. 20

Traffic feedback and situation awareness 1
The paired comparison of AOI statistics between the two drives showed the differences in 2 allocation of visual attention in response to the traffic feedback provided by VA near the TOR, 3 supporting our second hypothesis. The higher glances associated with checking exterior 4 mirrors, concentrating on the road ahead and checking traffic signs during the drive with VA 5 (Figure 7), relate to the verbal message delivered by VA (about the new speed limit and an 6 approaching pedestrian crossing), a few minutes prior to TOR which shows that verbal cues 7 can direct drivers' visual attention in the driving scene. Secondly, information about the 8 pedestrians might have led to increased mirror-checks and additional focus made by drivers on 9 the road ahead to prepare themselves for any remedial action that might be required. Increased 10 mirror checks are generally associated with the cognitive processes to gain situation awareness 11 at takeover (14,17,21). Drivers tended to shift their gaze to the speedometer, immediately  12 after the posting of a new speed limit, resulting in no difference in glances at the speedometer 13 between both driving conditions. However, it may also suggest a lack of trust and acceptance 14 by drivers (for automation and the digital assistant), although this may change over time, as 15 drivers' experience with such systems will increase. Another interesting finding was the 16 increase in glances towards the rear-view mirror in the absence of VA. During the one-minute 17 period of AOI analysis, there was no vehicle or event to reserve drivers' attention in the rear-18 view mirror which is otherwise a positive step in gaining SA (17, 36). As mentioned by a few 19 participants during a post-drive discussion, they were curiously observing the simulated objects 20 in the scenario, indicating a potential distraction. Furthermore, without any vocal alerts to 21 redirect their gaze at this time, drivers may have remained distracted by the virtual environment 22 in rear-view mirror. 23

VA and takeover 24
For a timely takeover, the drivers were expected to check the surrounding environment prior 25 to the construction zone to gain situation awareness (SA) and motor readiness by resuming 26 driving controls, to avoid the risk of the car heading into the construction zone. The survival 27 graph in Figure 8a shows the probability of a longer TOT is relatively higher during the drive 28 without VA, compared to the drive with VA. During the drive with VA, drivers were more 29 alert, to notice the construction zone following the TOR. Moreover, VA pre-informed them 30 about an intersection signal ahead and the new speed limit, to engage them with driving 31 environment, even in absence of any TOR. As shown by comparative AOI analysis (Figure 7), 32 this information appears to have influenced drivers' gaze behavior, encouraging them to check 33 for traffic signs, their speed etc., prior to TOR, thereby improving their ability to regain SA 34 and reducing the TOT by 3% to 9% (see 'glance frequency' in Table 2). Zeeb et al. (17) also 35 claimed that gaze behavior is a significant indicator of cognitive process at TOR, influencing 36 the TOT. However, during the drive without VA, drivers were not only fatigued and sleepy, 37 but had been provided with no such traffic information. Therefore, it is suspected that the 38 process of becoming alert and building SA would have been responsible for delaying the 39 takeover process during this drive. 40 According to the model results, female drivers are likely to take longer to takeover (i.e. to 41 demonstrate motor readiness) than male drivers (Figure 8b). Such a finding is interesting and could reflect a more cautious approach amongst female drivers, who may spend more time 1 exploring and assessing the driving scenesimilar results were reported by (24, 25). Among 2 the various non-driving activities that drivers could perform during automation, sleeping might 3 also be a voluntary action rather than just induced by the automation (35). Therefore, in this 4 study, the drivers who expressed that they would be likely to sleep in an automated vehicle 5 could suffer an increase in the probability of delayed takeover by 13% (Table 2). Nevertheless, 6 willingness to sleep also suggests high trust and acceptance in the technology. 7 It was apparent that drivers who frequently used other voice-based digital assistants felt more 8 comfortable using VA, and this may have encouraged them to engage more in conversations 9 (which might also be in terms of attentive listening). The drivers expressed their intention to 10 use similar voice-based driver assistant systems in the future, which is likely to have a positive 11 impact on driver alertness and SA (17). A hypothetical increase in rating from 0 (drivers who 12 have never used any voice assistant) to 20 (very often or always using different types of voice 13 assistants) as shown in Figure 4, suggests an increase in adoption of such technology. The 14 survival curves plotted in Figure 8c also suggest that higher use of these systems (indicated by 15 increase in rating) could potentially increase their effectiveness in assisting the drivers during 16 takeover after automation. 17

CONCLUSION 18
Extended periods of highly automated driving can disengage drivers from the driving task and 19 reduce their alertness. Therefore, the AOI analysis and model findings show clear advantages 20 of conversing with VA: 21 i. to counter the effects of passive fatigue. 22 ii. traffic-related information by VA can direct driver's cognitive process through 23 relocating visual attention to traffic signs, mirrors or road-ahead. 24 iii.
VA could effectively assist the drivers in a timely takeover. 25 Further, the parametric model of takeover time highlighted the gender-based differences in 26 takeover time of drivers. The younger drivers are expected to be more tech-savvy and 27 therefore more likely to use voice-based technologies than older driverswho may 28 subsequently not receive the benefits highlighted in the study. However, the current study 29 did not explore the effect of factors such as driver age, exposure to various in-car driver 30 assistance systems due to limited sample size. The findings highlight the need of VA 31 systems to maintain appropriate alertness and SA, especially for the drivers who may choose 32 to sleep in highly automated vehicles. However, the positive effects of conversing with VA 33 are likely to be transient, and therefore more research is required to investigate the lasting 34 effects of such interventions. 35 this manuscript are the sole responsibility of the authors of this paper and can in no way be