"Persuasive Vibrations": Studying the influence of vibration parameters on speech persuasion

HAL Id: hal-05372456 https://cnrs.hal.science/hal-05372456v1 Submitted on 19 Nov 2025 HAL is a multi-disciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. Distributed under a Creative Commons Attribution 4.0 International License “Persuasive Vibrations”: Studying the influence of vibration parameters on speech persuasion Sabrina Toofany, Anatole Lécuyer, Ferran Argelaguet, Justine Saint-Aubert To cite this version: Sabrina Toofany, Anatole Lécuyer, Ferran Argelaguet, Justine Saint-Aubert. “Persuasive Vibrations”: Studying the influence of vibration parameters on speech persuasion. IEEE Transactions on Haptics (ToH), In press, pp.1 - 12. �10.1109/toh.2025.3600579�. �hal-05372456� JOURNAL OF L A T EX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 1 “Persuasive Vibrations”: Studying the influence of vibration parameters on speech persuasion Sabrina Toofany, Anatole Lécuyer, Ferran Argelaguet, and Justine Saint-Aubert Abstract—This paper investigates the notion of “Persuasive Vibrations”, which showed that augmenting a person’s speech with vibrotactile feedback could artificially increase persuasion. However, while the initial paper has shown the effect, the underlying reasons why vibrations enhance persuasion remain unknown. Through two different user studies, this paper aims to study how the underlying parameters of the vibratory feedback (e.g., frequency, amplitude, or audio-vibration synchronization) influence persuasion. The first study aimed to identify the parameters of vibrotactile feedback that can positively influence persuasion. The second study evaluated vibrotactile feedback that might impair the persuasive effect. In a nutshell, the first experiment suggests that the isolation of different properties of the vibratory signal could tend to provide higher persuasion compared to no vibratory feedback. A lower frequency at 100 Hz seems the most efficient way to generate a persuasive effect. In contrast, the second experiment suggests that some alteration of the vibratory signal ( e.g., latency) does not decrease the levels of persuasion compared to the no-vibration condition. All in all, the results suggest that using lower frequencies could have a better effect on persuasion. These results could serve as a basis for haptic design in applications like videoconferencing, virtual meetings, and training systems where supporting user speech is essential. Index Terms—Audio, Haptic, Vibrotactile feedback, Persuasion, Social Communication, Speech I. INTRODUCTION With the rise of social media and the widespread use of remote communication, human interactions are increasingly taking place at a distance. It raises challenges related to the sharing of emotions and intentions between individuals. Studies have shown that remote communication, primarily through text or video, often results in a loss of emotional nuance and clarity of intentions [1]–[3]. In particular, persuasion and influence during social interactions tend to decrease in remote settings [4], [5]. Touch is a fundamental sense that plays a crucial role in social bonding [6], and the integration of haptic technologies has emerged as a promising way to enhance interactions [7]. A relevant approach in this area is the concept of “Persuasive Vibrations” [8]. It suggests that adding vibrotactile feedback that corresponds to a speaker’s voice can enhance the persuasion of their message. Vibrations help emphasize the speech, making the message more engaging. In virtual meetings where individuals with varying levels of leadership participate, integrating the vibrotactile feedback could help Sabrina Toofany, Anatole Lécuyer and Ferran Argelaguet are with Inria, Univ Rennes, IRISA, 35000 Rennes, France (e-mail: sabrina.toofan[email protected]). Justine Saint-Aubert is with CNRS, Univ Rennes, Inria, IRISA, 35000 Rennes, France. Manuscript received December 13, 2024; revised May 1st, 2025. balance power dynamics. However, while the initial results are promising, the underlying reasons why vibrations support persuasion still need to be explored. In this paper, our objectives were to identify which parameters of the vibrotactile feedback (e.g., frequencies, envelope, or synchronization) positively or negatively influence persuasion. For instance, audio research has shown that speech characteristics such as rate, intonation, or pitch can significantly affect persuasion [9], [10]. We aimed to evaluate if similar effects occur with tactile feedback. To achieve these goals, we conducted two user studies. The first study aimed to identify the parameters of vibrotactile feedback that can positively influence persuasion. The second study evaluated vibrotactile feedback that might impair the persuasive effect. II. RELATED WORK The following section discusses previous work related to how haptic feedback, particularly vibrations, can influence social interactions. We explore the impact of haptic on components related to persuasion, such as emotions, speech comprehension, and persuasion itself. A. Influence of Haptic Feedback on Speech Perception The combination of haptic and audio have shown some benefits in augmenting speech perception and focus. Bernstein et al. [11] conducted experiments showing that in environments with high background noise, haptic cues helped listeners differentiate important auditory signals from noise, allowing them to stay engaged with the audio content. Similarly, HnathChisolm et al. [12] found that adding vibrotactile feedback to audio in noisy settings enhanced the listener’s ability to focus on speech, helping them stay emotionally connected to the conversation. Reed’s work on the Tadoma method, in which individuals perceive speech by feeling the vibrations and movements of the speaker’s face, confirms the role of tactile input in communication. This method illustrates the effectiveness of touch in conveying complex information and influencing perception [13]. Vibrotactile feedback can improve speech perception in practical applications by providing tactile cues that supplement missing or unclear auditory information. Some studies [14]– [16] demonstrated that introducing haptic feedback to support auditory signals enhances speech comprehension. Vibrotactile cues help fill perceptual gaps, especially when crucial audio information is missing, improving overall understanding in challenging auditory environments. In summary, past studies suggest that vibrotactile feedback This article has been accepted for publication in IEEE Transactions on Haptics. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TOH.2025.3600579 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: INRIA. Downloaded on November 19,2025 at 09:11:35 UTC from IEEE Xplore. Restrictions apply. JOURNAL OF L A T EX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 2 can enhance audio clarity in noisy settings, supporting speech comprehension and focus. B. Influence of Haptic Feedback on Emotions Understanding the link between haptics and emotions is essential when studying haptic-based persuasion since studies indicate that emotionally engaging communications can enhance trust and compliance, thereby increasing the effectiveness of persuasion [17]. Haptic feedback plays a significant role in conveying and amplifying emotions [6], [18]. Touch communicates distinct emotions like anger, happiness, and sadness, effectively complementing verbal cues [19], [20]. Sawabe et al. [21] emphasize that combining robotic touch with speech boosts positive emotions in human-robot interactions, suggesting that tactile feedback reinforces emotional responses. In the context of vibrotactile feedback, slower vibrotactile stimuli tend to evoke positive emotional responses, while faster stimuli are linked to negative valence [22]. Therefore, a lower frequency might receive a more positive response from the user. Additionally, vibrotactile feedback replicating real-world perceptions can elicit emotions, with specific tactile patterns intensifying feelings by mimicking physical experiences like a steady, heartbeat-like pulse or cat purring [23]. Research has shown that haptic or vibrotactile feedback, combined with other cues or modalities, results in a more complete experience. For instance, when combined with thermal haptic feedback, vibrations still enhance persuasion, but warm feedback increases the friendliness of voice messages [24]. Integrating haptic feedback with audio offers a multisensory experience that enhances perception and emotional engagement. For example, in a study by Perrotta et al. [25], participants exposed to synchronized haptic and audio stimuli reported heightened emotional experiences compared to audioonly conditions. Specifically, tactile cues were synchronized with audio elements , such as rhythmic beats or voice inflections, allowing participants to experience the audio dynamics physically . These interactions enabled participants to feel sound intensity and rhythm variations , enhancing immersion and emotional engagement. Similarly, Gunther et al. [26] explored how vibrations synchronized with musical beats can create a deeper immersion for listeners, making the auditory experience feel more personal and physical. Collectively, these studies underscore how haptic feedback can shape emotional experiences and, when combined with audio, enrich perception and intensify emotional engagement. C. Influence of Haptic Feedback on Persuasion 1) Potential of Haptic Feedback: Some studies reveal the potential of haptic feedback to enhance persuasion and trust in both face-to-face and remote interactions. Concerning trust, mediated touch, such as vibrotactile feedback, enhances emotional closeness and trust between individuals even in remote settings [27]. Similarly, haptic feedback strengthens users’ confidence in the robot’s actions in human-robot interactions , increasing perceived reliability and safety [28]. A line of research, the “Midas Touch” effect, highlights how physical contact, such as a light touch on the arm, can significantly enhance persuasion by providing positive feelings [7], [29]. Similar behavior can be found with a robot as the touch origin [30], or in a virtual environment [31], highlighting the potential of haptic feedback to impact remote social interaction. Based on these findings, some studies suggest that people are more likely to be persuaded by persuasive messages that include a tactile element [32], [33]. Further exploring this area, Saint-Aubert et al. [8] show that vibrations combined with speech can subtly and effectively increase the persuasion of a virtual agent speaking or of some people when speaking. This concept is known as “Persuasive Vibrations”. Al Taha et al. [34] suggested reducing decision-making biases favoring lower-pitched voices by haptically enhancing higher-pitched voices with a vibrotactile signal that mirrors the voice pitch. However, a pilot study shows no significant improvement with the vibrotactile feedback. Altogether, these studies suggest that the type of vibrotactile parameters could influence persuasion. 2) Influence of Vibrotactile Parameters: In this paper, we wonder how vibrotactile parameters affect persuasion when combined with speech. Some works focused on the link between audio and vibrotactile feedback on leadership and trust. A haptic device called the BarryWhaptics [16], [35] was introduced to address voice pitch biases by modulating the intensity of the haptic response based on the pitch of the speaker’s voice. Specifically, high-pitched voices, which are generally perceived as less dominant, were augmented with higher vibration amplitude to balance perceived dominance. A user experiment showed that female participants tended to perceive voices as stronger when the haptic feedback was displayed, making them appear more dominant. However, no significant effect was found for enhancing trust or was noted for male participants though. While studies indicate that haptic feedback can influence persuasion, the effects of vibrotactile parameters are poorly understood. III. USER STUDY 1: VIBROTACTILE PARAMETERS THAT CAN ENHANCE PERSUASION A. Overview The objective of the first study was to understand which parameters of the vibrotactile signals cause the “persuasive” effect when combined with speech. Each parameter was isolated under different conditions. Participants were seated in front of a screen that displayed a remote communication interface (Fig. 1). They received audio feedback through headphones and held an actuator in their hand, which provided vibrotactile feedback related to the speech signal. The study consisted of two phases: a forced-choice phase and a questionnaire phase. In the forced-choice phase, two virtual speakers labeled “1” and “2” were presented to participants . Both speakers had the same voice and delivered the same sentence consecutively. However, the vibrotactile feedback differed between the two speakers, and participants had to choose which speaker they found more convincing. During the questionnaire phase, participants listened to one speaker This article has been accepted for publication in IEEE Transactions on Haptics. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TOH.2025.3600579 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: INRIA. Downloaded on November 19,2025 at 09:11:35 UTC from IEEE Xplore. Restrictions apply. JOURNAL OF L A T EX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 3 and then completed a questionnaire for each condition . The study protocol was approved by Inria’s ethical committee (COERLE). The experiment took approximately 60 minutes for a participant. Fig. 1. The setup of the user study, including a screen (displaying the videoconference-like), an audio headset (for audio feedback) and a vibrotactile actuator (held in the right hand). B. Hypotheses Previous research on “Persuasive Vibrations” indicates that vibrotactile feedback, when correlated with audio feedback, may positively influence persuasion [8]. Therefore, we expected this feedback to enhance persuasion compared to no vibrotactile feedback. Thus, our first hypothesis is: (H1a) Vibrotactile feedback corresponding to the speech (all frequencies and envelope) positively impacts persuasion compared to no vibrotactile feedback. Some studies suggest that a single-frequency vibrotactile signal could better enhance focus, hence persuasion, than a complex-frequency vibrotactile signal. Indeed, findings suggest that single-frequency signals allow for focused sensory engagement without straining cognitive processing, making them ideal for clarity and focus [36], [37]. Therefore, we set up the second hypothesis: (H2a) Vibrotactile feedback with only the fundamental frequency of the speech has the same impact on persuasion as vibrotactile feedback with all frequencies. The pitch of an audio speech seems to impact its persuasion: a voice with a low pitch will tend to be perceived as more persuasive than a voice with a higher pitch [38], [39]. The pitch is directly linked with the fundamental frequency of the voice [40], so tactile feedback composed of low frequency could positively impact persuasion. (H3a) Vibrotactile feedback with a low frequency positively impacts persuasion compared to vibrotactile feedback with a higher frequency. Other studies suggest that the envelope of a vibrotactile signal is better perceived than frequency variations [41]. A better perception of the signal can increase the focus [42]– [44], which in turn could enhance the persuasion [45], [46]. Therefore, the signal envelope could play an important role in persuasion. Based on that, we define the last hypothesis: (H4a) Vibrotactile feedback with a variable envelope positively impacts persuasion compared to vibrotactile feedback with a constant envelope. C. Vibrotactile Conditions A total of five conditions were investigated based on the previous hypotheses. Fig. 2 provides a schematic summary of the different vibrotactile conditions. •“No haptic”: The audio was not augmented by vibrotactile feedback. This condition was the control condition. •“Audio like”: The audio was enhanced with vibrotactile feedback replicating the same signal as the audio, including its frequency and envelope. The construction of the vibrotactile signal follows the same principle outlined in “Persuasive Vibrations”. In other words, the audio is directly used as the vibrotactile signal without any modification [8]. •“Mean frequency”: The vibrotactile feedback was based on a sinusoidal signal at the same frequency as the mean value of the fundamental frequency of the audio speech (167Hz). The audio envelope modulated the signal , which was determined with a Hilbert transform. •“Low frequency”: The vibrotactile feedback was based on a sinusoidal signal at a lower frequency (100Hz) than the mean value of the fundamental frequency of the audio speech. The choice of frequency was determined such that the difference between the mean value of the fundamental frequency and the lower frequency was higher than the just noticeable difference and respects the audio gap in the literature, that is to say, a frequency difference superior at five semitones (so frequency ratio of approximately 1.3 times the original frequency) [39], [47], [48]. The audio envelope modulated the signal , which was determined with a Hilbert transform. •“On-off”: The vibrotactile feedback was based on a sinusoidal signal at the same frequency as the mean value of the fundamental frequency of the audio voice (167Hz). The envelope was constant. The signal processing to create these vibrotactile feedback from the audio signal can be found in the appendixes (Tab. III). When vibrotactile feedback was rendered, it was synchronized with the speech: there was no feedback when the agent did not speak. During the forced choice phase, two conditions were compared at each trial. The order of the combinations was randomized. Each participant tested all 20 possible comparisons of the conditions 5 times, so a participant performed 100 comparisons. D. Experimental Setup The videoconferencing application ZOOM inspired the meeting interface. It was displayed on a screen (2560 x 1440 pixels) and implemented in Unity 3D (version 2023.2.3f1). Participants were seated at 1 meter in front of the screen. Virtual icons represented virtual speakers to avoid bias from physical appearance (Fig. 3). The icon was placed in the center of the screen to avoid bias from lateralization between visual and tactile feedback. Speakers spoke in a row and were identified by a label on the screen (“Listening 1” or “Listening 2”). An audio headphone provided the audio (HyperX Cloud II). This article has been accepted for publication in IEEE Transactions on Haptics. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TOH.2025.3600579 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: INRIA. Downloaded on November 19,2025 at 09:11:35 UTC from IEEE Xplore. Restrictions apply. JOURNAL OF L A T EX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 4 Fig. 2. Schematic of the temporal shape of haptic signal conditions for the first user study: “Audio like”, “Mean frequency”, “Low frequency”, “On-off”. More information about the signal frequencies and their temporal representation are available in the supplemental materials. The signal intensity attenuation can be found in the section D.2. An LRA actuator provided the vibrotactile feedback (Vybronics VG2230001H; cylinder of 22*30mm; resonance frequency at 70Hz; acceleration 2.5g) that allows the control of amplitude and frequency independently to some extent. Participants held it in their right hand. Depending on the actuator, the acceleration can vary according to the frequencies. We choose the vibrotactile actuator due to its constant acceleration value in an extensive range of frequencies. It was chosen so the resonance frequency is inferior to the fundamental frequency of the chosen voice and the frequency of the parameters chosen for the conditions. Therefore, the acceleration bandwidth on this range is linear so the frequency has less impact on the amplitude. The temporal shapes of the signals were measured with an accelerometer (LSM6DSV16BX 6-axis IMU model) and were also represented in frequency terms. The graphical representations obtained correspond to what was expected (see supplemental materials). Audio and vibrotactile feedback were synchronized using the audio stereo channels (left for tactile signal and right for audio signal) of an external audio interface (Focusrite Scarlett Solo 2nd Generation). Fig. 3. Experimental videoconferencing interface. The agent’s image is framed in yellow when audio is launched. The label varies (“Listening 1, Listening 2...”) depending on whether it corresponds to the first or the second signal in the comparison pair. 1) Audio feedback: Based on the literature, speakers always said the same sentence: “I urge you to vote for me” [35], [49]. We choose a male voice with a mean fundamental frequency (called here “Mean frequency”) at 167Hz. We adjusted the audio volume to ensure the message was clear to users without being excessively loud. We chose a voice with a mean frequency far enough from the resonance frequency of the vibrotactile actuator. During the signal creation, the audio signal was attenuated by -20 dB. The mean decibel level was at 52.3 dBA. The maximum value was at 57.0 dBA. These values were measured with a sonometer (01dB-Stell sonometer, SIP95S model) in A-weighting sound level. 2) Vibrotactile feedback calibration: The frequency perception of a signal is inherently linked to its amplitude perception and can vary among participants. Consequently, a calibration phase was conducted at the beginning of the study to determine the perceived amplitude of each signal relative to the reference signal, “Audio-like". Each signal was compared to the “Audiolike” signal, which was presented either first or second in a randomized order. During each comparison, participants were asked to adjust the amplitude of the signal, either by decreasing or increasing it, until they perceived intensity matched that of the reference. The direction of adjustment (increase or decrease) was also randomized. Each participant performed both type of adjustments for all signals, and the final perceived amplitude was computed as the average of the values obtained across four calibration trials. Participants calibrated 4 times per signal. The calibration was performed using a scale from 100% to 0% of the signal amplitude, a global volume at 100%, and a resolution of 1%. The calibration was made on a sample of 7 volunteers (7 males, between 20 and 40 years). They all gave written informed consent before testing and were not compensated for their participation. The calibration lasted 30 minutes. The decibel attenuation of each signal has the followed values: Audio like: 0 dB, Mean frequency: -16.5 dB, Low frequency: -24.4 dB, On-off: -27.9 dB. E. Data Collected During the forced choice phase, participants had to choose between the two speakers by choosing “Which agent was the more convincing?”. They answered on a keyboard: “1” for the first agent and “2” for the second. Each participant did 100 forced choices. A break was proposed every 20 forced-choice questions to reduce fatigue-related biases. During the questionnaire phase, participants completed a questionnaire for each condition about their haptic perception, based on questions proposed by Anwar et al. [50], and the This article has been accepted for publication in IEEE Transactions on Haptics. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TOH.2025.3600579 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: INRIA. Downloaded on November 19,2025 at 09:11:35 UTC from IEEE Xplore. Restrictions apply. JOURNAL OF L A T EX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 5 persuasion of the speaker based on a question suggested by Hanus et al. [51]: •The haptic feedback felt satisfying. •I felt that the agent was convincing. Participants answered these questions on 7-point Likert scales for each condition. After the first question, they were asked to develop their answer in a comment section. F. Participants A sample of 25 volunteers (16 males, mean age= 27.6 ± 8 years) participated in the first study. They all gave written informed consent before testing and were not compensated for their participation. All participants tested all the possible combinations between each condition. G. Results 1) Results of the Forced Choice Phase: The results were first analyzed to see the global effect on all the conditions (Fig. 4). The normality assumption was met for the data. For more accuracy, we applied a repeated-measures Friedman test to analyze differences across conditions (χ2= 16.47,df = 4, p < 0.01). This first test provides a compact overview of the results. We perform pairwise Wilcoxon signed-rank test posthoc comparisons using Bonferroni correction to control for multiple comparisons. Fig. 4. Global results of the first user study. Percentage of choices made for each condition in pairwise comparisons to the question “Which one of the two agents seems more convincing?”. The symbol “*” marks the significance for a p < 0.05. The “No haptic” condition was significantly less chosen than the “Low frequency” condition (p < 0.05). Then, we analyzed each condition’s influence on the participant’s binary outcome choice (Tab. I). For each comparison, we used a generalised linear mixed model (GLMM) to predict binary outcomes based on conditions while accounting for individual differences; and calculated probabilities of selecting one condition over another. The GLMM serves as a complementary model that estimates probabilities and confidence intervals for each pairwise comparison. We used a binomial family with a logit link function, based on R function (lme4 package, version 2024.12.1) as: glmer(Choice ~ Condition + (1 | Participant), family = binomial). The independent variable was the choice (binary) and the dependent variable was the comparison (method 1 vs method 2, 10 comparisons in total). We also included the participant as a random factor, and used a binomial function. If the probability and confidence intervals are <50%, the condition is less chosen; if they are >50%, the condition is more chosen. Regarding all the different pairwise conditions, two main results were observed. First, the “No haptic” has a lower level of persuasion than all the other conditions, as all the confidence intervals fall below 50% (first line of table I). Second, we observed the opposite effect for the “lowfrequency” condition. For all pairwise conditions, it has a higher level of persuasion (fourth line of table I). For the remaining pairwise comparisons, as the confidence intervals contain 50%, we cannot conclude that there is any difference. 2) Results for the Questionnaire Phase: The questionnaire results are displayed in Figs. 5 and 6. Most of the data did not meet the normality assumption . We applied a Friedman test for repeated measures and adjusted it with a Bonferroni correction. No significant difference between each condition was found. H. Summary of the User Study 1 The objective of user study 1 was to define which vibrotactile parameters can enhance persuasion. The (H1a) hypothesis (Vibrotactile feedback corresponding to the speech (all frequencies and envelope) positively impact persuasion compared to no vibrotactile feedback) is supported by the fact that all haptic signals proposed show a positive significant difference compared to “No haptic” condition. The (H2a) hypothesis (Vibrotactile feedback with only the fundamental frequency of the speech has the same impact on persuasion as vibrotactile feedback with all frequencies) is supported. The two conditions have no significant difference, and both are better at persuasion than speech without vibrotactile feedback. The (H3a) hypothesis (Vibrotactile feedback with a low frequency positively impacts persuasion compared to vibrotactile feedback with a higher frequency) is also supported. The data aligns with the (H3a) hypothesis in pairwise and global analyses, with low frequency outperforming all other effects. These results suggest that frequency may be a key factor in enhancing persuasion. The (H4a) (Vibrotactile feedback with a variable envelope positively impacts persuasion compared to vibrotactile feedback with a constant envelope) is not supported here. This result suggests that the effect of amplitude variations on persuasion is small.. Overall, the results of user study 1 suggest that vibrotactile feedback can positively influence persuasion, with lower frequencies appearing more effective, and vibrotactile persuasion could come from the fundamental frequency of the vibrotactile signal. However, these results should be further studied, to This article has been accepted for publication in IEEE Transactions on Haptics. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TOH.2025.3600579 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: INRIA. Downloaded on November 19,2025 at 09:11:35 UTC from IEEE Xplore. Restrictions apply. JOURNAL OF L A T EX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 6 No haptic Audio like Mean frequency Low frequency On-off No haptic 38 [29, 45] % 26 [19, 35] % 23 [17, 32] % 38 [29, 48] % Audio like 62 [55, 71] % 42 [33, 53] % 32 [24, 42] % 41 [32, 51] % Mean frequency 74 [65, 81] % 58 [47, 67] % 38 [29, 49] % 48 [38, 58] % Low frequency 77 [68, 83] % 68 [58, 76] % 62 [51, 71] % 64 [54, 73] % On-off 62 [52, 71] % 59 [49, 68] % 52 [42, 62] % 36 [27, 46] % TABLE I PAIRWISE PROBABILITIES BETWEEN CONDITIONS BASED ON GLMM ANALYSIS FOR THE FIRST USER STUDY. EACH CELL REPRESENTS THE PROBABILITY OF SELECTING THE CONDITION IN THE ROW COMPARED TO THE CONDITION IN THE COLUMN Fig. 5. Results for the first experiment, based on the Likert scale 7-point responses for the first question of the questionnaire phase: “The haptic feedback felt satisfying”. Fig. 6. Results for the first experiment, based on the Likert scale 7-point responses for the second question of the questionnaire phase: “I felt that the agent was convincing”. explore the role of specific vibrotactile parameters such as frequency and envelope in persuasive vibration. IV. USER STUDY 2: VIBROTACTILE PARAMETERS THAT CAN IMPAIR PERSUASION A. Overview The objective of the second user study was to determine whether the vibrotactile signals provided to participants might decrease persuasion compared to speech without any tactile feedback or other vibrotactile feedback. Vibrotactile conditions different from those in user study 1 were tested. These vibrotactile conditions were supposed to disturb the user’s speech. The experiment setup, protocol, and data collected were similar to those of the first user study. For the forced choice phase, participants had to choose the speaker that was the most of the two. The protocol was approved by Inria’s ethical committee (COERLE). The experiment took approximately 60 minutes. B. Hypotheses Research in audio suggests that speech with a low pitch will tend to be perceived as more persuasive than a speech with a higher pitch [38], [39]. Our first user study shows that speech with a low pitch will tend to be perceived as more persuasive than speech with a higher pitch, here the Mean frequency . We can then test this assumption with higher frequency so that vibrotactile feedback with higher frequency could be less persuasive than mean frequency. Thus, our first hypothesis is: (H1b) A vibrotactile feedback with high frequency is less persuasive than a vibrotactile feedback with low frequency. Previous research suggests that audio-haptic delays can be detected at intervals between 25 and 31 ms, though this threshold may extend up to 179.4 ms when using pulse feedback [52]–[55]. Humans are sensitive to the temporal congruency between feedback, so the perception of a delay could divert focus and, therefore, reduce persuasion. Therefore, we set up the second hypothesis: (H2b) A vibrotactile feedback delayed with the audio speech feedback is less persuasive than speech without vibrotactile feedback. Audio research investigating the impact of environmental noise on focus finds that it tends to reduce focus, hence persuasion [45], [56]. Based on that, we define the last hypothesis: (H3b) A vibrotactile feedback consisting of random noise is less persuasive than speech without tactile feedback. We can then assume that a similar effect could be elicited with a tactile signal consisting of random noise. We did not use white noise, as it is perceived much weaker through the actuator than the reference signal (Mean frequency condition). C. Vibrotactile Conditions A total of five conditions were investigated based on the previous hypotheses. Fig. 7 provides a schematic summary of the different vibrotactile conditions. •“No haptic”: The audio was not augmented by any vibrotactile feedback. This condition was the control This article has been accepted for publication in IEEE Transactions on Haptics. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TOH.2025.3600579 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: INRIA. Downloaded on November 19,2025 at 09:11:35 UTC from IEEE Xplore. Restrictions apply. JOURNAL OF L A T EX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 7 Fig. 7. Schematic of the temporal shape of haptic signal conditions for the second user study: “Mean frequency”, “High frequency”, “Delay”, “Pink noise”. More information about the signal frequencies and their temporal representation are available in the supplemental materials. The signal intensity attenuation can be found in the section D. condition. •“Mean frequency”: Similar to the first experiment. •“High frequency”: The vibrotactile feedback was based on a sinusoidal signal at a higher frequency than the fundamental frequency of the speech (231Hz). The choice of frequency was determined such that the difference between the fundamental frequency and the higher frequency was higher than the just noticeable difference and respect the audio gap in the literature, that is to say, a frequency difference superior at five semitones (so frequency ratio of approximately 1.33 times the original frequency) [39], [47], [48]. The audio envelope modulated the signal and was determined with a Hilbert transform. •“Delay”: The vibrotactile feedback was similar to the ”Mean frequency” condition but with a positive delay (+220ms) superior to the minimum delay perceptible between sound and vibrotactile feedback to ensure the user can perceive the delay without ambiguity. •“Pink noise”: The vibrotactile feedback was based on a signal with constant amplitude and frequencies following a pink noise frequency spectrum. Except for the “Pink noise” condition, when a vibrotactile feedback was displayed, it was synchronized with the speech: there was no feedback when the agent did not speak. The signal processing to create these vibrotactile feedback from the audio signal can be found in the appendixes (Tab. III). During the forced choice phase, a combination of conditions was compared. The order of the combinations was randomized. Each participant tested all 20 possible comparisons of the conditions 5 times, so a participant performed 100 comparisons. The signal processing to create these vibrotactile feedback can be found in the appendixes (Tab. III). D. Experimental Setup The same protocol was used for User Study 2, and the same voice was chosen for the audio part. Moreover, a calibration part similar to the first user study was also made. The calibration was made on a sample of 7 volunteers (6 males, between 20 and 30 years). They all gave written informed consent before testing and were not compensated for their participation. The calibration lasted 30 minutes. The experiment itself took approximately 60 minutes. The decibel attenuation of each signal has the followed values: Mean frequency: -16.5 dB, High frequency: -0.4 dB, Delay: -16.5 dB, Pink noise: -30.4 dB. E. Participants A sample of 25 volunteers (17 males, mean age: 29 +/- 6 years) participated in the second experiment. Participants were not the same as in the first user study. All participants gave written informed consent before testing and were not compensated for their participation. They all tested all the possible combinations between each conditions. F. Results 1) Results for the Forced Choice Phase: The results were first analyzed to see the global effect on all the conditions (Fig. 8). The normality assumption was met for the data. For more accuracy, we applied a repeated-measures Friedman test to analyze differences across conditions (χ2= 26.06,df = 4, p < 0.01). This first test provides a compact overview of the results. We perform pairwise Wilcoxon signed-rank test posthoc comparisons using Bonferroni correction to control for multiple comparisons. * ** * * 0 50 100 No haptic Mean frequency High frequency Delay Pink Noise Condition Total number of choices Fig. 8. Friedman Test for second experiment conditions. Bonferroni correction applied. ”*” marks the significance for a p < 0.05. ”**” marks the significance for a p < 0.001. This article has been accepted for publication in IEEE Transactions on Haptics. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TOH.2025.3600579 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: INRIA. Downloaded on November 19,2025 at 09:11:35 UTC from IEEE Xplore. Restrictions apply. JOURNAL OF L A T EX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 8 No haptic Mean frequency High frequency Delay Pink noise No haptic 18 [12, 28] % 37 [26, 49] % 30 [20, 41] % 62 [50, 73] % Mean frequency 82 [72, 88] % 79 [69, 87] % 71 [59, 81] % 76 [65, 85] % High frequency 63 [51, 74] % 21 [13, 31] % 30 [20, 41] % 59 [47, 71] % Delay 70 [59, 80] % 29 [19, 41] % 70 [59, 80] % 67 [55, 77] % Pink noise 38 [27, 50] % 24 [15, 35] % 41 [29, 53] % 33 [23, 45] % TABLE II PAIRWISE PROBABILITIES BETWEEN CONDITIONS BASED ON GLMM ANALYSIS FOR THE SECOND USER STUDY. EACH CELL REPRESENTS THE PROBABILITY OF SELECTING THE CONDITION IN THE ROW COMPARED TO THE CONDITION IN THE COLUMN Fig. 9. Results for the second experiment, based on the Likert scale 7point responses for the first question of the questionnaire phase: “The haptic feedback felt satisfying”. ”*” marks the significance for a p value inferior at 0.05. ”**” marks the significance for a p value inferior at 0.001. Fig. 10. Results for the second experiment, based on the Likert scale 7-point responses for the second question of the questionnaire phase: “I felt that the agent was convincing”. ”*” marks the significance for a p value inferior at 0.05. ”**” marks the significance for a p value inferior at 0.001. Whereas in Experiment 1, the difference between “No haptic” and “Mean frequency” was not significant, this time the difference is (p < 0.05). This result can be explained by a bias caused by multiple comparisons with other persuasive conditions. In addition, the “Mean frequency” was significantly more chosen than other conditions (p < 0.05). Then, we analyzed each condition’s influence on the participant’s binary outcome choice (Tab. II). For each comparison, we used the same GLMM model as the first user study. First, compared to other vibrotactile included conditions, the “No haptic” condition seems less chosen except for the “Pink noise” condition (first line of table II). However, the probability difference between the median of the “No haptic” condition and the “Pink noise” condition is over 60%. Second, the “Mean frequency” has a higher level of persuasion compared to all the other conditions, as all the confidence intervals are superior to 50% (second line of table II). Third, the “High frequency" and the “Pink noise" conditions have lower level of persuasion compared to the “Delay" condition. For the remaining pairwise comparisons, as the confidence intervals contain 50%, we cannot conclude that there is any difference. 2) Results for the Questionnaire: The normality assumption was not met for most of the data. We applied a Friedman for repeated measures test adjusted with a Bonferroni method to see the global effect on all the conditions (Figs. 9 and 10). There is a preference for the ”Mean frequency” condition compared to other vibrotactile conditions (Fig. 9). It follows the results seen in the global Friedman Test (Fig. 8). Moreover, the ”Mean frequency” condition is seen as more persuasive than other conditions except the ”High frequency” condition. G. Summary of the User Study 2 The objective of user study 2 was to define which vibrotactile parameters might impair persuasion. The results support the hypothesis (H1b) (A vibrotactile feedback with high frequency is less persuasive than vibrotactile feedback with a low frequency), indicating that some vibrotactile feedback parameters might affect persuasion compared to speech not synchronized with haptic. It goes along with the hypothesis that there is a congruence between vibrotactile feedback persuasion and vibrotactile frequency. That is to say, when the frequency is higher than the mean fundamental frequency of the voice, persuasion decreases. Conversely, persuasion increases when the vibrotactile signal’s frequency is lower than the mean fundamental frequency of the voice. (H1b) is also supported by the fact that “Mean frequency” condition is more persuasive than other conditions. (H2b) (A vibrotactile feedback delayed with the audio speech feedback is less persuasive than speech without vibrotactile feedback) is not supported. However, a difference seems to appear in the pairwise analysis. So, this suggests that even with a delay, a vibrotactile signal could enhance the persuasion of a speech more effectively than speech not haptically augmented. (H3b) (A vibrotactile feedback consisting of random noise is This article has been accepted for publication in IEEE Transactions on Haptics. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TOH.2025.3600579 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: INRIA. Downloaded on November 19,2025 at 09:11:35 UTC from IEEE Xplore. Restrictions apply.