Whispers in the Dark: Effects of Voice Type and Dialog Context in Likability and Tension

2. 1. User Interface in the Nighttime Context

While a substantial body of HCI literature has examined interface design in nighttime or presleep contexts, most of this work has concentrated on visual modalities. Common strategies—such as screen dimming, blue-light filtering, and simplified layouts—are typically intended to lower physiological arousal and facilitate relaxation (Oshima et al., 2020; Nagare et al., 2019). For example, Chang et al. (2015) demonstrated that evening exposure to blue light from electronic devices can suppress melatonin production, delay circadian rhythms, and reduce alertness the following morning. Such findings underscore the physiological and cognitive rationale for minimizing sensory stimulation in night-oriented UX design. However, in practice, most voice agent interactions still show limited adaptation to the sensory constraints and expectations of low light or nighttime environments. As visual cues diminish in low light environments, the auditory system assumes a more prominent role in spatial and situational awareness (Zheng et al., 2024), potentially increasing users’ sensitivity to sound intensity and quality during voice based interaction.

Although these strategies were developed for visual interfaces, their underlying goals are potentially transferable to auditory modalities, particularly in dark or low-light environments. Auditory design becomes particularly relevant in nighttime UX due to a combination of environmental and physiological factors. From an environmental perspective, domestic nighttime settings impose unique social constraints, such as the need to minimize acoustic leakage to adjacent spaces (e.g., inter-floor noise) or to avoid disturbing cohabitants. From a psychological perspective, users preparing for sleep typically exhibit reduced tolerance for sensory stimulation and heightened sensitivity to auditory cues, making abrupt or salient sounds more likely to disrupt relaxation and sleep readiness.

Despite these characteristics, existing nighttime UX research has paid limited attention to how such constraints translate to auditory or voice-based interaction. While prior work has begun to explore nighttime UX more broadly, far fewer studies have examined how principles established in visual nighttime interface design could inform the design of auditory interfaces or shape the behavior of conversational agents in nocturnal settings (Nißen et al., 2022).

Moreover, current research on nighttime auditory interfaces remains narrow in scope, focusing largely on biometric tracking or clinical sleep interventions (Karlgren et al., 2022; Aarts et al., 2022; Semertzidis et al., 2023; Karlgren & McMillan, 2022). As a result, a significant knowledge gap persists regarding how conversational agents should adapt their social and acoustic behaviors to harmonize with everyday domestic nighttime environments. Addressing this gap, the present study extends context-sensitive design principles established in visual nighttime UI research to the auditory domain. Focusing on domestic after-dark interactions, we examine whispering as a context-aware vocal strategy that responds to both the environmental constraints (e.g., low stimulation, noise sensitivity) and the psychological expectations (e.g., calmness, reassurance) characteristic of nighttime interaction.

2. 2. Voice Characteristics of the VAs

Individuals can anthropomorphize VAs using the “voice” element, thereby attributing human-like characteristics and behaviors to them (Appel et al., 2012; Wagner & Schramm-Klein, 2019). This serves as a core feature that distinguishes computers (Nass et al., 1994) and allows all devices sharing a common voice to be considered to possess shared intelligence (Nass & Brave, 2005). Due to these characteristics, significant research has been conducted in the field of VAs with a focus on voice. Despite the central role of voice in user experience, most commercial VAs continue to utilize standardized voices (Cambre & Kulkarni, 2019), and academic inquiry has focused predominantly on factors such as gender (McGinn & Torre, 2019) and pitch (Tolmeijer et al., 2021; Goodman & Mayhorn, 2023).

More recent work has begun to acknowledge the role of usage context(Cambre & Kulkarni, 2019), but this line of inquiry has largely concentrated on spatial settings such as automobiles or public environments (Park & Lee, 2022; Jestin et al., 2022). Temporal context—particularly nighttime use—remains comparatively underexplored, despite its distinct social and physiological constraints.

Auditory interfaces must be calibrated for the distinct socio-environmental constraints of nighttime. In interpersonal communication, individuals naturally modify their speech patterns after dark, adopting softer tones and slower pacing to maintain quietness and avoid disturbing cohabitants. While this behavior is partly a response to environmental factors like noise sensitivity, it also reflects a shift toward a lower-arousal communicative state. Instead of merely reducing volume, humans often utilize whispering—a specialized vocal register—to facilitate private and intimate sharing in the dark.

Nighttime interaction therefore calls for a vocal strategy that goes beyond incremental stylistic adjustments within voiced speech.

While parameters such as pitch, prosody, or speech rate modulate expressiveness along a continuous dimension, whispering constitutes a distinct phonatory mode characterized by the absence of vocal fold vibration (Wilkinson, 1982). Because this shift represents a categorical change in vocal production rather than a gradual prosodic adjustment, whispering carries acoustic and social meanings that are difficult to achieve through volume or pitch modulation alone. For this reason, the present study treats whispering as a theoretically distinct vocal strategy, rather than as an extreme point on a continuum of vocal intensity. Importantly, this nighttime shift in speech is not limited to quantitative adjustments such as reduced volume or slower speech rate, but often involves the adoption of whispering as a socially learned communicative convention.

Prior research suggests that whispering is commonly used to reduce speech audibility (Li, 2011) and to signal intimacy by decreasing perceived interpersonal distance (Andersen, 2015). In mediated contexts, studies on ASMR and related auditory content indicate that whispering can effectively reduce physiological arousal, such as anxiety and heart rate (Barratt & Davis, 2015; Seifzadeh et al., 2023; Hardian et al., 2020). However, whispering may also evoke negative impressions, such as being perceived as “creepy,” when deployed without contextual appropriateness (Parviainen & Søndergaard, 2020). This underscores the critical importance of alignment between vocal register and interaction context.

It is worth noting that commercial voice assistants have already begun to incorporate whispering as a design feature. For example, Amazon Alexa’s ‘whisper mode’ enables the system to respond softly when it detects a whispered input (Erika, R., 2023). However, such implementations remain largely heuristic and reactive, relying on simple acoustic triggers rather than a principled understanding of when and why whispering is most appropriate. To date, little empirical research has examined how whispered interaction modulates conversational context, user affect, or perceived social appropriateness, particularly during the sensitive nighttime period. By systematically testing these variables, the present study seeks to provide the empirical evidence necessary to move beyond reactive features toward context-aware, human-centered vocal design.

At the same time, whispering is not universally appropriate. When misaligned with conversational intent or situational expectations, it may evoke discomfort or be perceived as eerie or inappropriate (Parviainen & Søndergaard, 2020; Li, 2011). This dual nature underscores the importance of contextual alignment in vocal design. Accordingly, this study examines whispering as a theoretically grounded yet context-sensitive vocal strategy, comparing a whispering voice agent with a conventional voice agent to investigate how this phonation mode shapes user experience in nighttime interactions.

H1: Interactions with a whispering VA in a nighttime context will lead to lower tension than interactions with a normal VA in the same context.

The voice of a VA is closely tied to its likability, which refers to how users perceive the system as useful, enjoyable, and friendly (Hone & Graham, 2000). Likability has long been recognized as a key factor in motivating technology adoption (Davis et al., 1992). Prior research also shows that human-like traits in AI VAs can significantly enhance user perceptions and interactions (Kühne et al., 2020). Building on these findings, we expect that whispering voices that align more closely with nighttime contexts will further enhance user likability.

H2: Interactions with a whispering VA in a nighttime context will lead to higher likability than interactions with a normal VA in the same context.

2. 3. Dialog Context in VAs

Building on the established importance of nighttime-appropriate voice design, interactions with VAs in these contexts are shaped not only by voice characteristics but also by the purpose and content of conversation. Cassell (1999) emphasized that VAs should integrate both task-oriented and social aspects, and subsequent work has stressed the importance of strategies that go beyond task fulfillment to enrich user experience (Sciuto et al., 2018; Gaver et al., 2008; Barber et al., 2014).

Despite research on voice characteristics, the role of conversational content in voice-based interactions remains underexplored. Conversational behavior is often classified as social- or task-oriented, depending on purpose and relationship (Chattaraman et al., 2019; Keeling et al., 2010). In task-oriented conversations, dialog is goal-driven and formal; however, in social-oriented conversations, it is characterized by relational interactions, including exclamatory feedback and small talk (Chattaraman et al., 2019; Bickmore & Cassell, 2001).

Other work has highlighted how voice interacts with dialog type. Female voices, for instance, are often rated more positively in social interactions (Lee & Park, 2019), and preferences in counseling contexts vary depending on user demographics (Kim et al., 2021). Moreover, while social small talk with VAs has been effective in domains such as customer service, it has proven less effective in serious domains like military training (Bickmore & Cassell, 2005).

These findings collectively suggest that both dialog type and context shape how users evaluate VA voices. However, few studies have explored how these factors interact in nighttime settings. To address this gap, we examine whether whispering voices, which are often linked to intimacy and relaxation, enhance social nighttime interactions compared to task-oriented.

H3: There will be an interaction effect between voice type and dialog context on the user’s level of tension toward VAs.

H4: There will be an interaction effect between voice type and dialog context on user’s likability evaluation of VAs.

3. 1. Study Design

3. 1. 2. Experimental environment

The experiment was conducted in a quiet, noise-controlled laboratory space on campus that was deliberately configured to approximate an after-dark, pre-sleep interaction context. The room was dimmed to a low-light level to reflect nighttime conditions, and external light sources were blocked using blinds and opaque materials. Partitions were installed behind and beside the participant to reduce visual distractions and create a stable, enclosed listening environment. Although the space was not a fully furnished bedroom, the setup was designed to reproduce key perceptual characteristics of nighttime in-home voice interaction. The laboratory environment remained quiet throughout all sessions, minimizing potential interference from ambient noise.

To support the intended nighttime framing, the physical environment was complemented by contextual instructions that encouraged participants to mentally situate themselves in a pre-sleep interaction context. Prior to each trial, participants were provided with a brief overview of the upcoming scenario and asked to imagine interacting with the voice agent at home before going to sleep (e.g., lying in bed or preparing to sleep in a quiet room). This cognitive framing was intended to compensate for the limitations of a laboratory setting by approximating psychological characteristics of nighttime interaction in conjunction with the low-light physical environment.

A Kakao Mini speaker was used to deliver whispering and non-whispering voice stimuli at consistent volume levels. Participants were provided with a brief adaptation period and a short practice interaction before the main trials to acclimate to the environment and reduce potential novelty- or laboratory-induced tension.

Figure 1

Experimental Environments

3. 2. Participants

We recruited participants with experience using VAs through an online advertisement. A total of 31 participants were recruited via online advertisements, with a mean age of 25.5 years (SD= 3.82, Male (n = 12), Female (n = 19)). All participants were native Korean speakers without hearing problems. Additionally, all participants had experience with VAs. Each participant received $10 as compensation. There were instances where a participant provided double responses to a single survey question, for which the average of the two responses was used in the data analysis. The study was approved by the University’s IRB (IRB No. 7001988-202404-HR-2074-04).

3. 3. Stimuli

To understand the user experiences, we created the treatments (voice and scenarios) required for this experiment.

3. 3. 2. Scenario design

Four dialog scenarios were developed to examine user experiences across different conversational contexts: two task-oriented (weather and news recommendations) and two social-oriented (concerns about the future and stress from work). Scenario topics were selected to reflect after-dark, in-home voice-agent use and routines. Prior research on in-home conversational agents reports that the bedroom is a frequent device location and that users describe routines that include “getting ready for bed,” alongside clear time-of-day patterns with usage peaking again in the late evening. In addition, information-seeking requests such as weather checks are among the most popular voice commands in everyday usage (Sciuto et al., 2018). Based on this evidence, we designed task-oriented scenarios to represent routine information-seeking and planning interactions in after-dark home settings. For social-oriented scenarios, we focused on topics such as future anxiety and work stress. This choice was motivated by psychological findings that repetitive negative thinking (e.g., rumination and worry) is closely associated with sleep-related difficulties(Takano et al., 2012, Takano et al., 2014). Accordingly, these topics were selected to reflect emotionally supportive conversations that can occur near bedtime.

The initial drafts of the experimental scenarios were generated using ChatGPT-4o and were subsequently revised and restructured by the authors to align with the study’s objectives. To ensure methodological consistency, all scripts were standardized using a fixed four-turn exchange format (User–Agent–User–Agent) with agent responses strictly limited to 2–3 sentences to maintain consistent information density and interaction duration. Average interaction times were as follows: weather recommendation (M = 03:18, SD = 00:13), news recommendation (M = 03:10, SD = 00:20), worries about the uncertain future (M = 03:12, SD = 00:22), and stress about work (M = 03:32, SD = 00:20). Participants experienced all four conditions in a counterbalanced order, as determined by the Latin Square design described above.

To support internal validity and comparability across conditions, we employed scripted scenarios rather than open-ended dialog. This design choice enabled the standardization of interaction duration, information density, and conversational structure across participants, thereby reducing variability associated with personal conversational content. To prevent specific topics (e.g., weather vs. news) from confounding the results, the four scenarios were decoupled from the voice types and counterbalanced across participants via the Latin Square design described above. For instance, the ‘weather’ scenario was presented in a whispering voice for some participants and in a non-whispering voice for others. This cross-assignment ensured that the observed differences in user experience were attributable to the voice type and dialog context, rather than the inherent nature or emotional load of the specific topics.

Table 1

presents one task-oriented and one social-oriented dialog as representative examples of the four scenarios used in this study.

3. 4. Procedure

The experiment was conducted in a dimly lit on-campus laboratory designed to simulate nighttime conditions. Upon arrival, participants were briefed on the study, provided informed consent, and given a five-minute rest period to adapt to the environment. They then completed a baseline activation questionnaire before the experiment began. An introductory voiceover presented the agent’s voice, followed by a short practice dialog to ensure familiarity with the system. Before each scenario, participants were instructed to imagine that they were interacting with the voice agent at home after dark, in a quiet, low-light environment as part of their bedtime routine.

Participants engaged in four scenario-based conversations with the voice agent “Nua” each followed by a paper-based questionnaire assessing their experience. A two-minute break was given between sessions. The entire procedure lasted approximately 50 minutes(Figure 2).

Figure 2

Experimental Procedure

3. 5. Measurement

To evaluate user experiences, we collected quantitative data on likability and tension using the Activation-Deactivation Adjective Checklist (AD-ACL) scale (Thayer, 1986).

The tension subscale of AD-ACL was used to evaluate participants’ activation or arousal levels, capturing their momentary mood and arousal states. AD-ACL is a self-reported scale that provides a subjective assessment of arousal, serving as an alternative to physiological measurements by reflecting the participant’s immediate perception of their overall activation state. In neuropsychological contexts, arousal is linked to awareness of emotional experiences (Bhandari et al., 2017). In this study, we measured the element of stability using tension.

The AD-ACL scale comprises four primary categories: energy (general activation), tension(high activation), tiredness (general deactivation), and calmness (deactivation-sleep). Among these, tension was selected to examine psychological comfort in the nighttime context, particularly after establishing the dark setting of the experiment. Tension represents tense arousal, which ranges from tense to calm, and is a key indicator of external stress responses (Thayer, 1986; Bhandari et al., 2017). This subscale included five descriptive terms that participants could rate based on their immediate mood, measured on a 4-point Likert scale (1-Strongly Disagree, 4-Strongly Agree).

Likability refers to a positive emotional response toward an object. In this study, likability is defined as an interpersonal liking that motivates participant to continue engaging with VA(Kang et al., 2021). The subjective assessment of speech system interfaces scale was based on the translated version used by Kang et al. (2021) to assess the degree to which the agent exerts influence on the participant (Hone & Graham 2000). All measurements were conducted using a 5-point Likert scale.

3. 6. Statistical Analysis Method

First, paired t-tests were performed to analyze differences in tension scores between the baseline and each condition. A two-way repeated measures ANOVA was then conducted to examine the main and interaction effects of voice type and dialog context on tension. Finally, likability was also analyzed using a two-way repeated measures ANOVA.

4. 1. Descriptive Statistics and Measurement Validation

The reliability of the measurement tools was verified since the data collected in the experiment were based on the participants’ subjective measurements. To assess the reliability of the survey items for the two dependent variables, likability and tension, Cronbach’s ⍺ coefficients were calculated. The results showed that likability (⍺ = .83) and tension (⍺ = .72) demonstrated Cronbach’s ⍺ values above 0.7, thus confirming the reliability of the dependent variable(Fornell & Larcker, 1981).The means and standard deviations (SDs) of the dependent variables for each condition are presented in Table 2.

Table 2

Mean and SD scores (type of voice x context of dialog, N = 31).

4. 2. Hypothesis Testing

Baseline–Compare Tension for each condition

To investigate the effects of whispering and non-whispering voice conditions on tension, we conducted paired t-tests using the baseline as the reference for each condition. In the whispering voice conditions, we observed a significant difference in tension between the baseline and task-oriented whispering (TW) condition (t(30) = 2.295, p < .05, SD = 0.412). The tension score for the TW condition (M = 7.35, SD = 2.274) was significantly lower than that of the baseline (M = 8.58, SD = 2.754). We also observed a significant difference in tension between the baseline and social-oriented whispering (SW) conditions (t(30) = 2.730, p < .05, SD = 0.366). The tension score in the SW condition (M = 7.23, SD = 1.927) was significantly lower than that in the baseline.

In the non-whispering voice conditions, we observed a significant difference in tension between the baseline and task-oriented non-whispering (TNW) conditions (t(30) = 2.062, p < .05, SD = 0.370). The tension score in the TNW condition (M = 7.55, SD = 2.461) was significantly lower than that in the baseline. However, no significant difference was observed between baseline and social-oriented non-whispering (SNW) conditions (t(30) = 0.172, p = .864, M = 8.48, SD = 3.150). Therefore, the whispering condition showed lower levels of tension than the other conditions, with the SW condition exhibiting the lowest scores overall. Hypothesis 1 predicted that interactions with a whispering voice would result in lower tension compared to interactions with a non-whispering voice. Paired t-tests supported this prediction: both task-oriented and social-oriented whispering conditions showed significantly lower tension than the baseline.

Figure 3

Comparison of tension levels between baseline and experimental conditions* p < .05, ** p < .01, TW (task + whispering), SW (social + whispering), TNW (task + non-whispering), SNW(social + non-whispering) 

Within-Subjects Analysis of the Tension

A two-way repeated measures ANOVA revealed that tension did not have a significant effect of either voice type or dialog context (Table 3). However, the interaction effect of voice type and dialog context (F(1, 30) = 4.430 p < .05, η_p² = .129) was significant (Figure 4). In other words, non-whispering voice and social-oriented dialogs had the highest tension levels, and whispering voice and social-oriented dialogs had the lowest tension levels. Thus, Hypothesis 3, which stated that there would be an interaction effect between the voice type and dialog context on the user’s level of arousal toward VA, was supported. However, the repeated-measures ANOVA revealed no significant main effect of voice type on tension. Instead, a significant interaction between voice type and dialog context was found, indicating that the effect of whispering voices on tension was context-dependent rather than uniform across all dialog types. Thus, H1 was partially supported.

Table 3

Results of Repeated Measures Two-Way ANOVA

Within-Subjects Analysis of Likability

Repeated measures two-way ANOVA results showed that the main effect of voice type (F(1,30) = 4.358, p < .05, η_p² = .127) was significant; however, no significant effect was found for the dialog context (Table 4). According to Bonferroni’s post-hoc analysis results, the whispering voice had a higher mean (M = 3.782, SD = .119) than the non-whispering voice (M= 3.508 SD = 0.151), suggesting that the whispering voice was more likely to be favorable in a nighttime context after dark. The interaction effect of voice and dialog type (F(1, 30) = 6.731, p < .05, η_p² = .183) was significant (Figure 5). Whispering voice and social-oriented dialog had the highest likability ratings. Thus, Hypotheses H2 and H4 were supported.

Table 4

Results of Repeated Measures Two-Way ANOVA

Figure 5

Interaction effect between voice type and dialog context under likability

4. 3. Summary of Hypotheses Testing Results

To provide a clear overview of the experimental findings, summarizes the results of the hypothesis testing. Below, a detailed interpretation of each hypothesis in relation to the observed data is provided.

• H1: Partially supported. Whispering voices reduced tension in t-tests relative to baseline, but no main effect emerged in ANOVA.

• H2: Supported. Whispering voices were generally rated as more likable than non-whispering voices.

• H3: Supported. A significant interaction revealed that the calming effect of whispering was strongest in social-oriented dialogs.

• H4: Supported. Likability was highest when whispering voices were paired with social-oriented dialogs.

H1, which predicted that whispering voices would reduce user tension compared to non-whispering voices in nighttime interactions, was partially supported. While paired t-tests showed that both task-oriented and social-oriented whispering conditions significantly reduced tension relative to the baseline, the repeated-measures ANOVA revealed no significant main effect of voice type on tension. Instead, the effect of whispering on tension was contingent on dialog context. H2 was supported, as whispering voices were generally perceived as more likable than non-whispering voices, evidenced by a significant main effect of voice type on likability. H3, which proposed an interaction effect between voice type and dialog context, was supported for both tension and likability. Specifically, whispering voices were most effective in reducing tension and enhancing likability in social-oriented dialogs, highlighting the importance of contextual congruence in nighttime voice interactions. Finally, H4 was supported, showing that likability was highest when whispering voices were paired with social-oriented dialogs.

5. 1. Central Contribution: Vocal-Contextual Congruence in Nighttime UX

The central contribution of this study is the articulation of vocal–contextual congruence as a primary driver of nighttime voice interaction UX. Rather than positioning whispering as a universally calming strategy, our findings demonstrate that its effectiveness depends on the alignment between vocal delivery (the “how”) and dialog intent (the “what”) within low-arousal nighttime contexts. Across both affective and evaluative measures—tension and likability—the interaction between voice type and dialog context consistently shaped user responses. In line with this pattern, Hypotheses 2, 3, and 4 were fully supported, demonstrating that whispering voices—specifically in social-oriented dialog contexts—significantly enhanced likability and reduced tension. User experience in nighttime interactions was shaped by the degree of congruence between vocal delivery and dialog intent: when these were aligned (SW condition), both likability and tension improved; when mismatched (SNW condition), tension increased and likability declined.

Importantly, Hypothesis 1 was only partially supported, as whispering did not produce a uniform main effect across all dialog types. Although whispering generally reduced tension relative to baseline conditions, the absence of a broad main effect in the ANOVA reinforces a nuanced insight: nighttime voice UX is driven less by a single “optimal” voice type and more by achieving harmony between vocal tone and conversational purpose. For instance, the significant interaction effect reveals that psychological relief is optimized only through contextual alignment. This suggests that tension arises from vocal-contextual incongruity rather than from the whispering voice per se, effectively countering concerns that novel whispering tones might provoke “uncanny valley” discomfort (Moore, 2017; Carr et al., 2017). Together, these results extend prior work on the calming and intimate qualities of whispered speech (Barratt & Davis, 2015; Poerio et al., 2018; Zarazaga & Malisz, 2023) by demonstrating that likability and tension reduction are fundamentally optimized through situational appropriateness.

Ultimately, this study demonstrates that voice is not merely a medium of information delivery but a dynamic design resource that must be calibrated to the user’s temporal and social context. Our findings underscore the necessity of moving beyond “one-size-fits-all” vocal models toward multidimensional context-awareness that considers when, where, and why an interaction occurs (Sciuto et al., 2018; Luger & Sellen, 2016). The core takeaway is that nighttime voice UX is not driven by a single optimal voice type, but by the harmony between vocal intensity and conversational dimensions. By prioritizing situational appropriateness over uniform vocal strategies, designers can transition toward adaptive, context-sensitive voice assistants that better align with the user’s emotional and temporal state (Purington et al., 2017).

5. 2. Design Principles for After-Dark Voice Interaction

Based on the empirical evidence of this study, we propose two core design principles for developing voice agents in after-dark routines. These principles move beyond our specific implementation and articulate reusable design knowledge at a general level.

• Principle 1: Prioritize Tone–Intent Congruence. Intimacy-forward, low-intensity vocal delivery should be selectively aligned with supportive or socially oriented dialog in low-arousal contexts. The benefits of whisper-style delivery emerge most clearly when it resonates with users’ emotional expectations and conversational goals.

• Principle 2: Treat Vocal–Intent Mismatch as a Potential Source of Tension. In nighttime settings, the absence of vocal adaptation is not a neutral design choice. A high-intensity, standard conversational tone used in emotionally sensitive interactions may function as a mismatch that increases tension or diminishes likability. Designers should therefore evaluate vocal strategies based on contextual appropriateness rather than acoustic clarity alone.

These principles are expected to generalize across low-arousal, domestic nighttime contexts where users seek emotional attunement over task efficiency. However, they may not directly apply to daytime productivity interactions, high-urgency alert systems, or contexts where acoustic clarity takes precedence over intimacy. The boundary conditions of these principles should be tested in future work across diverse temporal and functional interaction contexts.

5. 3. Theoretical and Practical Implications

Academic contributions. Prior research on voice assistants has largely focused on daytime or productivity-related contexts, leaving nighttime interactions underexplored. By examining how whispering voices shape user experiences in nighttime settings, this study fills an important gap in the literature on context-aware voice interaction. In doing so, it extends earlier conceptual work on nighttime UX (e.g., Parviainen & Søndergaard, 2020) with empirical findings that go beyond theoretical propositions. Critically, whereas prior voice interaction research has treated vocal characteristics as context-independent design variables, this study introduces vocal–contextual congruence as an empirically grounded construct—demonstrating that the perceived intimacy, comfort, and social engagement afforded by whispering are contingent on dialog intent rather than inherent to the voice itself. This reframes how researchers and designers should theorize the role of voice in emotionally sensitive, low-arousal interaction contexts.

Practical implications. Our findings suggest that whispering may be more effective as a situational choice rather than a default nighttime voice, with its benefits emerging most clearly when the vocal tone aligns with socially oriented interaction goals. In socially oriented dialogs, whispering voices enhanced comfort, reduced tension, and increased likability, indicating that vocal strategies should be selected in relation to both dialog type and temporal context rather than applied uniformly. By aligning voice tone with the appropriateness of context, designers can foster more engaging, supportive, and emotionally resonant interactions, especially in intimate settings such as the home. From a design perspective, whispering Vas can be applied in everyday contexts such as supporting bedtime routines, guiding stress management practices, or serving as sleep companions that provide reassurance during nighttime. In healthcare or elder-care settings, whispering voices could deliver unobtrusive nighttime support, enhancing comfort without disturbing rest.

Our study focused on user-initiated interactions in which participants explicitly called the agent to initiate a conversation. In this context, whispering was associated with enhanced social presence during socially oriented nighttime interactions. However, system-initiated interactions (e.g., proactive notifications) may require more cautious use of whispering. Because whispering functions as a strong signal of intimacy, an unsolicited whispered response from an agent could be perceived as intrusive or eerie, particularly when the user has not yet engaged in a social dialog. These findings suggest that interaction initiative should be considered as a key factor when applying whispering as a nighttime vocal strategy.

In addition to temporal context, our observations during the study suggest that physical and environmental factors may also shape nighttime voice interactions. For example, several participants spontaneously adjusted their posture or moved the device closer in order to better hear whispering voices in the low-light setting. While not systematically measured, this behavior highlights how vocal intensity can interact with spatial proximity and listening effort. From a design perspective, this suggests that whispering voices may benefit from complementary cues—such as adaptive volume scaling, proximity awareness, or clarity enhancement—to maintain intimacy without increasing cognitive or physical effort for users. Designers should therefore consider the embodied conditions of nighttime use—where listening effort and physical comfort are paramount.

Finally, adaptive voice agents may benefit from incorporating user-specific signals. Rather than relying solely on time-of-day heuristics, future systems could adjust vocal intensity based on users’ recent interaction patterns or speech characteristics. Treating whispering as a situational response—rather than a fixed design choice—allows voice agents to better accommodate individual preferences and evolving nighttime contexts.

5. 4. Limitations and Future works

Limitations. This study focused on voice-based interactions in after-dark contexts, examining how whispering voices shape user experiences in low-arousal environments. While this scope enabled a detailed exploration of nighttime-specific design needs, it limits the ability to determine whether the effects observed are unique to nighttime or reflect broader user preferences. Our single-context approach aligns with prior studies that examined nighttime or dark settings without daytime comparisons, such as Oshima et al. (2020) on smartphone use in dark environments and Himschoot et al. (2024) on feelings of safety during nighttime outdoor recreation under different artificial lighting conditions.

Additionally, the controlled laboratory environment minimized distractions but may not fully capture the complexity of real-life bedroom contexts, including the situational nuances and emotional states that can shape users’ experiences and affective responses during bedtime routines. Furthermore, the use of predetermined dialog scenarios, while necessary to ensure experimental control and comparability across conditions, may have constrained participants’ sense of personal relevance and spontaneity compared to fully user-initiated interactions. As a result, the observed responses reflect reactions to standardized interaction structures rather than participants’ own naturally occurring nighttime conversations.

Furthermore, a methodological limitation concerns the operationalization of the whispering condition. In this study, whispering was implemented as a low-intensity vocal delivery consistent with how whispering naturally occurs in nighttime contexts. As such, phonatory characteristics were inherently accompanied by reduced loudness, making it difficult to attribute the observed effects exclusively to whisper phonation alone. The findings should therefore be interpreted as reflecting an intimacy-forward, low-intensity voice configuration rather than an isolated phonatory manipulation. Future research should systematically disentangle phonation, loudness, and vocal identity to clarify their independent contributions to user experience.

Moreover, this study focused on short-term interactions, limiting our findings to initial interactions. Long-term studies are needed to explore how ongoing dialog with whispering agents affects user experience and relationship-building over time. Finally, our sample was restricted to adults aged 20 to 30, who were experienced voice-agent users, which limits the generalizability of the findings. Experienced users in their 20s may be more receptive to novel voice tones and social-oriented conversations than other demographic groups or novice users.

Future research. Future research should (1) include daytime and other contextual comparisons to test the boundary conditions of whisper-style voice effects, (2) replicate the study in more naturalistic settings and through in-the-wild and longitudinal designs to capture everyday routines and contextual variability (e.g., ambient lighting and background noise), (3) recruit broader demographic groups and users with varying levels of VA experience, and (4) disentangle additional vocal parameters (e.g., pitch, prosody, speech rate) to identify which acoustic components drive comfort and trust in low-stimulation contexts. In summary, this study establishes that voice is not merely a communication tool but a contextual cue that fundamentally shapes user perceptions and experiences. By situating voice design within nighttime contexts, it advances both academic discourse on voice interaction and the practical development of adaptive, context-aware VAs that can support richer and more natural interactions.