When eyes beat lips: speaker gaze affects audiovisual integration in the McGurk illusion

The eyes of another person are of fundamental importance to human behavior from early infancy on (Farroni et al., 2004; Striano & Reid, 2006), shaping attentional, perceptual, and affective processing. If someone directs their gaze towards us, our attention is immediately captured. Direct gaze (as opposed to averted gaze or closed eyes) is preferentially detected and processed (e.g., Baron-Cohen, 1995; Coelho et al., 2006; Senju & Hasegawa, 2005; von Grünau & Anston, 1995; also see Böckler et al., 2015). Being looked at by another person may elicit self-referential processing in the addressee (Conty et al., 2016; Hietanen & Hietanen, 2017) and boost prosocial behavior (Izuma et al., 2010, 2011) as well as positive appraisal of others (see Kleinke, 1986, for a review). Another’s eyes convey information about their affective and cognitive states (e.g., Emery, 2000; Kleinke, 1986; Schilbach, 2015) and can signal aggression (Nichols & Champness, 1971) as well as attraction (Mason et al., 2005). Critically, the eyes not only signal information to others but they simultaneously encode information from the environment, thus serving a dual function (Cañigueral & Hamilton, 2019; Gobel et al., 2015; Kendon, 1967; Risko et al., 2016).

Not surprisingly, eye contact plays a major role in human communication. In particular, direct gaze serves as an ostensive communicative signal (Csibra & Gergely, 2009; Sperber & Wilson, 1986): A speaker typically directs her gaze towards the addressee, thereby making it manifest to the addressee that she is the designated recipient of the upcoming message and that this message is being intentionally communicated to her (Csibra & Gergely, 2009; Lanthier et al., 2019, 2021). Moreover, the gaze is used in a conversation to control turn-taking behavior such that a speaker ends her turn with a direct gaze at the addressee and the addressee then begins her turn with averted gaze (Kendon, 1967; recently replicated by Ho et al., 2015). When responding to questions, preferred responses are produced while gazing at the questioner whereas the gaze is averted for dispreferred responses (Kendrick & Holler, 2017). To signal understanding in conversation, addressees systematically use eye blinking (Hömke et al., 2017). It has further been found that the stronger the coupling between speakers’ and addressee’s eye movements, the better the addressee’s comprehension (Richardson & Dale, 2005).

While the processing of a speaker’s gaze is important in a conversation, the processing of the speaker’s speech is naturally paramount. During speech perception, addressees integrate the incoming auditory and visual signals (i.e., the speaker’s vocal sounds and lip movements) into a unitary percept through a process known as multisensory integration (Ernst & Bülthoff, 2004). Generally, whether our brain binds multiple incoming sensory signals together or processes them separately depends on whether these signals are perceived to be causally related. When we hear another person’s words while seeing their lips move, the visual signal (the lip movements) and the auditory signal (the vocal sounds) have the same causal origin (the other person) and are integrated into a single percept rather than perceived as two separate signals. Typically, the integration of signals from multiple sensory modalities boosts perceptual performance because several sources of redundant information are combined. For instance, speech in a noisy environment is understood better if the speaker’s lip movements can be observed by the addressee (Ma et al., 2009; MacLeod & Summerfield, 1987; Ross et al., 2007; also see Altieri et al., 2016).

If, however, the signals from multiple sensory modalities provide inconsistent information, the process of multisensory integration can impair or bias perceptual performance and can lead to sensory illusions (e.g., Shams et al., 2000). In the case of speech perception, if the visual and auditory signals do not match, this can give rise to auditory percepts that neither correspond to the visual nor the auditory signal. In particular, McGurk and MacDonald (1976) have shown that when participants were presented with an auditory syllable (e.g. “Ba”) while observing the speaker’s lips uttering a different syllable (e.g. “Ga”), they reported hearing an illusory syllable (e.g. “Da”). This striking multisensory illusion (often referred to as the “McGurk illusion” or “McGurk effect”) demonstrated, for the first time, the powerful influence of vision upon auditory speech perception. The McGurk illusion has since been extensively investigated with regard to the mechanisms underlying it and the factors influencing it (for a comprehensive review, see Alsius et al., 2018). Note that the illusion has also been reliably found in studies that were run online (e.g., Brown et al., 2018; Karas et al., 2019; Magnotti et al., 2018, 2020), with highly similar results between lab-based and online studies (Magnotti et al., 2018). Previous research suggests that the McGurk illusion can be explained using a causal inference model of multisensory perception (Magnotti & Beauchamp, 2017) that has also been applied successfully to a variety of other multisensory phenomena, such as the ventriloquist effect (Körding et al., 2007; Rohe & Noppeney, 2015) and the sound-induced flash illusion (Shams et al., 2005).

In terms of factors that influence the McGurk illusion, previous studies have shown that attentional demands play a critical role. In particular, if participants were presented with the typical McGurk stimuli and asked to indicate what they heard while at the same time performing a secondary (visual or auditory) task, the perception of the McGurk illusion decreased (Alsius et al., 2005). This finding indicates that audiovisual integration of speech is weakened under a high attentional load. Relatedly, another study (Munhall et al., 2009) showed that the McGurk illusion is perceived only if addressees consciously attend to the visual signal, i.e., the speaker’s lip movements. Together, these findings suggest that one needs to pay (a sufficient amount of) attention to the presented stimuli in order for the integration process to take place.

Whereas dual tasks like the above (Alsius et al., 2005) explicitly shift and divide participants’ attentional resources, a recent study by Gurler et al. (2015) investigated the natural (i.e., uninfluenced) distribution of people’s visual attention while they observed the face of the “McGurk speaker” uttering syllables. Via eye tracking, the authors measured which region on the speaker’s face participants tended to fixate on, using eye fixation as a proxy for attentional focus. The results showed that the distribution of participants’ eye fixations predicted the degree to which participants perceived the McGurk illusion. Specifically, the more participants tended to look at the speaker’s mouth, the more they perceived the McGurk illusion—presumably because they were more strongly influenced by the visual signal provided by the lip movements (for a replication and additional manipulations, see Stacey et al., 2020). Hence, observing the speaker’s lip movements—a strategy that would usually improve speech comprehension (Ma et al., 2009; MacLeod, & Summerfield, 1987; Ross et al., 2007)—has a negative and misleading effect because the visual signal coming from the lips is inconsistent with the auditory signal.

Interestingly, Gurler et al. (2015) also observed that there were several participants who did not look primarily at the speaker’s mouth, choosing instead to look at the speaker’s eyes or explore multiple regions of the face. These participants perceived the McGurk illusion to a lesser extent. Thus, this study suggests that there is no commonly shared focus of attention in the McGurk paradigm, i.e., participants tend to look at the mouth, or at the eyes, or at multiple face regions. Hence, one cannot predict in advance where a particular participant will focus her attention when confronted with the McGurk stimulus.

In sum, previous research by Alsius et al. (2005), Gurler et al. (2015), Munhall et al. (2009), and Stacey et al. (2020) suggests that the degree to which people perceive the McGurk illusion depends (1) on their attention in general (with less attention leading to a decrease of the illusion) and (2) on their attentional focus on the speaker’s mouth versus elsewhere (with a focus on the mouth leading to an increase of the illusion).

The eyes of another person have—amongst other things—the function and the power to spontaneously direct, divert, or capture our attention. In particular, another’s direct eye gaze and motion onset (e.g., switching from direct to averted gaze, or from closed to open eyes) are two powerful cues that capture our attention (e.g., Abrams & Christ, 2003; Böckler et al., 2014; Senju & Hasegawa, 2005). In turn, what we attend (and to what extent we attend it) determines how we process incoming multisensory signals (e.g., Alsius et al., 2005; Munhall et al., 2009) – and thus affects whether we can be “tricked” by our senses and subjected to multisensory illusions (Gurler et al., 2015; Stacey et al., 2020).

During a conversation, a speaker’s eyes and the audiovisual signals coming from their lips both typically provide relevant information. The addressee processes the speaker’s gaze while at the same time processing her audiovisual speech signals. To date, it has not been systematically investigated whether these two processes interact, i.e., whether the speaker’s gaze affects how the addressee integrates the speaker’s audiovisual speech signals. In the present study, we aimed to address this question using the classic McGurk illusion. Building on previous research, we examined if a speaker’s gaze behavior (i.e., motion and eye contact) would dynamically capture an addressee’s attention and would thereby influence how the addressee integrates the speaker’s vocal sounds and lip movements.

To test the effect of motion, we manipulated whether the speaker moved his eyelids up/down (i.e., open/closed his eyes) prior to speaking or did not show any eye motion (factor “Motion”). To test the effect of eye contact, we manipulated whether the speaker spoke with open eyes or closed eyes (factor “Eyes”). The factor Motion was manipulated between-subjects, the factor Eyes was manipulated within-subjects.

We determined our target sample size of 2 × 70 participants by running an a priori power analysis using G*Power (Faul et al., 2007, 2009) targeting moderately sized effects (Cohen’s d = 0.34 for paired samples t-tests; Cohen’s d = 0.48 for independent samples t-tests; alpha = 0.05, Power = 0.80). Thus, we recruited 70 participants for each level of our between-subjects factor Motion, i.e., 70 for the “Static” condition (motion absent) and 70 for the “Dynamic” condition (motion present). Data was collected through the online participant recruitment service Prolific (https://​www.​prolific.​co/​).¹ All participants had normal or corrected-to-normal vision and hearing; they were between 18 and 45 years old and fluent in English.² Only participants whose performance in previous Prolific studies had been reliable (approval rates of at least 75%) were admitted to this study. Moreover, participants who showed below 60% accuracy in an auditory baseline condition of our study were excluded because we needed to ensure intact hearing capabilities.

The participant samples for the Static and the Dynamic conditions consisted of 33 females, 36 males, and 1 other (M = 25.47 years, SD = 6.16 years), and of 21 females, 48 males, and 1 other (M = 24.30 years, SD = 6.33 years), respectively. All participants gave written informed consent and received monetary compensation for their participation (1.50 GBP for the Static condition (~ 12 min) and 1.88 GBP for the slightly longer Dynamic condition (~ 15 min)).

Videos were recorded with a MacBook (early 2016) using its internal microphone and its 480p FaceTime camera. Each video showed a headshot of a man (see Fig. 1) uttering one out of five syllables (“Ba”, “Pa”, “Ga”, “Ka”, “Na”). The speaker’s eyes were either open (“Eyes open”) or closed (“Eyes closed”) while he was speaking. In the Dynamic condition, the speech act was preceded by motion: the speaker either opened or closed his eyes before uttering the syllable (with his eyes remaining open or closed). Each video clip lasted 3.5 s. To create a second version of the videos that did not contain the preceding motion (Static condition), the respective first part of each clip was cut off, resulting in a shorter (2 s) version of each clip. Thus, the long video clips showed the speaker first closing or opening his eyes and then uttering one of the syllables with either open or closed eyes (Dynamic) whereas the shorter clips only included the speech act without the preceding motion (Static), see Fig. 1.

These video clips naturally included matching auditory and visual signals (e.g., the speaker’s lips produced the word “Ga” and the auditory signal was “Ga”). In addition to these “congruent” versions, we created “incongruent” versions with mismatching auditory and visual signals. To this end, the actual sounds in the videos were muted and dubbed either with the auditory syllable “Ba” or “Pa” (using iMovie, version 10.1.9) such that, for example, the speaker’s lips produced the word “Ga” yet the auditory signal was “Ba” or “Pa”. The final set of audiovisual stimuli included 6 incongruent versions and 6 congruent versions (see Table 1 for an overview). Additionally, we used only the audio tracks for the syllables “Ba” and “Pa” as an auditory baseline (see “auditory only” in Table 1).

All video material was recorded in one continuous take to ensure similar intonation and loudness of the uttered syllables and to keep visual conditions constant. The speaker opened and closed his eyes and produced the syllables in sync with a metronome beat (40 bpm; not audible to participants) to ensure that the temporal sequences were comparable across videos. This way, we also attempted to exclude potential tempo differences between speaking with open vs. with closed eyes. All videos are publicly accessible (under CC-BY license) via the Open Science Framework (https://​osf.​io/​vjw6k/​?​view_​only=​7c4c9831aa6a4dbb​8fcfbd5a8ef7e501​).

The experiment was programmed in PsychoPy 3 (Peirce et al., 2019) and run online via Pavlovia (https://​pavlovia.​org/​).

To test whether motion and eye contact affect multisensory speech processing, we used a 2 × 2 mixed factorial design. As a between-subjects factor, we varied whether motion (i.e., opening or closing of the speaker’s eyes) preceded the speech act or not (Motion: Dynamic/Static). As a within-subjects factor, we varied whether the speaker’s eyes were open or closed while he was uttering the syllable (Eyes: Open/Closed).

Participants were presented with three different trial types (incongruent, congruent, auditory only) in randomized order, see Table 1. In the incongruent trials, the syllable that the speaker produced with his lips (either “Ga”, “Ka”, or “Na”) did not match the auditory syllable that was presented (“Pa” or “Ba”). In the congruent trials, the syllable that the speaker produced matched the auditory syllable (“Pa” or “Ba”). In the auditory only trials, participants were presented with a black screen (instead of the speaker’s face) and either heard the syllable “Pa” or “Ba”. Note that we did not include “visual only” trials.

After stimulus presentation, participants were asked to indicate which syllable—out of four options presented on the screen—they heard (see Table 1, for a list of all syllable combinations and response options). The four response options were adopted from Stropahl et al. (2017) and comprised the presented auditory syllable, the presented visual syllable, and two fusion responses. The two fusion responses were the two responses that showed the highest fusion percentage (i.e., the most commonly reported illusory percepts resulting from the fusion of mismatching auditory and visual signals) in the original study by McGurk and MacDonald (1976). The response options were the same for all three trial types (see Table 1).

Participants responded by pressing the corresponding number of the response option (1, 2, 3 or 4) on the keyboard. There was no time limit for responses. Participants were told that if they were unsure about what the speaker said, they should simply choose the response option that seemed most likely to them. They were ensured that this task was not about accuracy but about their individual perception.

Participants performed a total of 120 trials. The order of response options for each trial type was randomly chosen out of two possible orders (either “Visual, Auditory, Fusion1, Fusion2” or “Fusion1, Fusion2, Visual, Auditory”). The options were presented in rows, one word beneath the other, and numbered consecutively from 1 to 4. The 120 trials were composed of 24 auditory only, 24 congruent, and 72 incongruent trials. Critically, in half of the congruent and in half of the incongruent trials, the speaker’s eyes were open (first and third row, Fig. 1); in the other half they were closed (second and fourth row, Fig. 1).

The only difference between the two experimental conditions was that in the Dynamic condition, the speaker either opened or closed his eyes prior to uttering the syllable (third and fourth row, Fig. 1). In the Static condition, there was no motion preceding the utterance (first and second row, Fig. 1). Exemplary trial sequences for all four-factor combinations (Static + Eyes open, Static + Eyes closed, Dynamic + Eyes open, Dynamic + Eyes closed) are shown in Fig. 1.

As part of the general study instructions, participants were asked to turn off all distractions (e.g., music, TV, phone) and to always look at the screen and listen to the voice. They were asked to concentrate fully on the task and not to perform any other tasks at the same time. The instructions pointed out that they should read the response options carefully because the available options and the order of options would change continuously. Finally, participants were asked to wear headphones, if possible, while completing the study.

Before starting the actual experiment, participants were familiarized with the trial logic by performing three training trials. The first two training trials were congruent trials and the third trial was an auditory-only trial.

As dependent variables, we recorded response accuracies (for congruent and auditory-only trials), response choices (for incongruent trials), and response times³ for all trials. Response accuracy was computed as the proportion of trials in which participants chose the auditory syllable (i.e., the vocal sound that is actually presented) out of the four response options. Response choice for incongruent trials (where auditory and visual signals mismatched) was computed as the proportion of trials in which participants chose the auditory syllable, the visual syllable (i.e., the sound originally produced by the lips), and the fused syllable (i.e., the illusory percept), respectively. Response time was computed as the time between stimulus offset (i.e., end of video and appearance of the response options) and response selection.

For statistical inference, we used permutation-based ANOVAs and post-hoc tests. That is, the null distribution of the test statistics was estimated by repeatedly sampling permutations of the actual data under the assumption that there are no differences between the levels of our experimental factors (Kherad-Pajouh & Renaud, 2015). All post-hoc tests were Bonferroni-corrected. As effect size measures, we report generalized eta squared (η_G²; Bakeman, 2005) for the ANOVAs and Cohen’s d for the post-hoc tests. Data were analyzed using customized R scripts.

All raw data are publicly available via the Open Science Framework (https://​osf.​io/​vjw6k/​?​view_​only=​7c4c9831aa6a4dbb​8fcfbd5a8ef7e501​).

When looking at the main experiment, the effect of motion seems to highlight the power of motion cues to capture people’s attention, converging with previous research on the effects of sudden onset (eye) motion on attentional capture (e.g., Abrams & Christ, 2003; Böckler et al., 2014, 2015; van der Wel et al., 2018). In particular, the sudden opening/closing of the speaker’s eyes presumably directed participants’ attention to the eyes (and away from the lips). Thus, the misleading visual signal provided by the speaker’s lips had a smaller influence on the audiovisual integration process such that participants perceived fewer illusory percepts and instead perceived the actual auditory syllables. This finding is in line with previous research showing that the degree to which people perceive the McGurk illusion depends on their attentional focus on the speaker’s mouth vs. elsewhere, with a focus on the mouth leading to an increase of the illusion (Gurler et al., 2015; Stacey et al., 2020). It has also been shown that people only experience the McGurk illusion if they consciously attend to the speaker’s lips (Munhall et al., 2009). Thus, the effect of motion may be interpreted in terms of attentional capture.

However, when considering the control experiment (“Static matched”), this interpretation is called into question, as the results show that a static stimulus that matched the motion stimulus in length also led to a reduced McGurk illusion. From this result, one might conclude that the motion itself did not actually play the main role, but it was rather the video length that caused the difference between the Static (motion absent, shorter video) and the Dynamic condition (motion present, longer video) in the main experiment. Some additional differences between the Static matched and Dynamic conditions should be considered, however. Specifically, participants in the Static matched condition (1) experienced the McGurk illusion to an even smaller extent than participants in the Dynamic condition and (2) responded more slowly than in the Dynamic condition. These two aspects in combination suggest that different processes might be at work in the two conditions, yet leading to similar outcomes. In particular, it is possible that the reduced McGurk illusion in the Dynamic condition is caused, at least partially, by the motion cue functioning as attentional capture (as discussed above). In the Static matched condition, however, the reduced McGurk illusion might be caused by the fact that participants were generally less attentive because the still frame at the beginning of each video is rather boring. This general reduction of attention might lead to (1) a reduction in audiovisual integration (cf. Alsius et al., 2005; Munhall et al., 2009; Talsma & Woldorff, 2005) and (2) a slow-down of responses; just as observed in the Static matched condition.

Our results suggest that the McGurk illusion is smaller when the eyes of the speaker were closed compared to open–provided that people reliably perceive the basic McGurk illusion (regarding the latter constraint, please see Sect. Control condition: Static matched - Results). First of all, it is possible that when the speaker spoke with closed eyes, participants did not consider themselves the intended recipient of the message and thus paid less attention overall to the stimuli. Since sufficient attention is a prerequisite for audiovisual integration to occur (Alsius et al., 2005; Munhall et al., 2009; Talsma et al., 2007), a general reduction of attention might have impaired the integration process, thus resulting in fewer illusory percepts when the speaker’s eyes were closed.

The effect of eye contact might also suggest that a speaker’s closed eyes capture the addressee’s attention more strongly than open eyes in the setting of the present study. One reason could be that closed eyes in a conversation are very unusual and thus salient for an addressee. Typically, interlocutors converse with open eyes following specific gaze patterns (Ho et al., 2015), yet the speaker in the present study spoke with closed eyes. Presumably, the fact that speaking with closed eyes is not consistent with common social norms led to increased salience and thus caught and captured participants’ attention. Apart from the fact that closed eyes might be more salient than open eyes in a conversation context, it is also possible that participants focused more on the speaker’s closed compared to his open eyes because feeling another’s direct gaze elicits self-referential processing (“Watching Eyes model”, see Conty et al., 2016; Hietanen & Hietanen, 2017) and self-awareness (e.g., Baltazar et al., 2014; Hazem et al., 2017; Pönkänen et al., 2011), increases arousal (Helminen et al., 2011), and invites for social interaction (Ho et al., 2015). Thus, in line with the finding that people avoid long eye contact with strangers (Ellsworth et al., 1972; Laidlaw et al., 2011) and look longer at faces with averted than direct gaze (Helminen et al., 2011), participants in the present study might have preferred to attend to the speaker’s eyes when those were closed and they were not feeling watched, as this creates less self-involvement. Thus, the effect of eye contact can be interpreted in terms of the saliency of closed eyes in a conversation context as well as in terms of the Watching Eyes model (cf. Conty et al., 2016), as both mechanisms would result in fewer illusory percepts when the speaker’s eyes were closed.

A future lab-based study could use eye tracking to measure gaze fixations as a proxy of overt attention and pupil size as a proxy of attentional processing to disentangle the two interpretations. Specifically, the interpretation that a speaker’s closed eyes reduced the illusion due to generally reduced attention would be supported by gaze patterns that reflect disengagement of the visual scene (e.g., fewer fixations on the face) in the Eyes closed condition. By contrast, the interpretation that the reduced illusion is due to attention capture by closed eyes would be supported by earlier, more and/or longer fixations at the eye region in the Eyes closed condition. Finally, it is also possible that both interpretations are partially correct, as they are not mutually exclusive.

Some of our findings warrant additional research before stronger conclusions can be drawn. Regarding our interpretation of the motion cue in terms of attentional capture, it is possible that apart from the eyes, other motion cues in the speaker’s face (e.g., wrinkling the forehead) might have similar effects. It has been shown, for instance, that an external moving object (i.e., a falling leaf in front of a speaker’s face) reduces the McGurk illusion (Tiippana et al., 2004). It is noteworthy that in this study, the leaf motion occurred during the speaker’s utterance whereas the motion in the present study occurred prior to the utterance. Future studies are needed to determine the specificity and boundary conditions of the motion effect, especially regarding the ‘identity’ and the timing of the motion.

More generally, regarding the reduced McGurk illusion in the Dynamic condition (motion present, longer video) compared to the Static condition (motion absent, shorter video), we cannot argue conclusively as to whether it was the motion or the length of the video that caused this effect. The results of the control condition (“Static matched”: motion absent, longer video) resembled those of the Dynamic condition, suggesting, at first glance, that the effect can be attributed to the video length. However, as pointed out above, it is possible that different processes were at work in the Dynamic condition (motion cue guides addressee’s attention to the speaker’s eyes) and the Static matched condition (still frame leads to addressee feeling bored and thus paying less attention overall). Both processes would result in an impaired integration and thus in fewer illusory percepts. Further control experiments with systematic and orthogonal manipulations of motion and video length are needed to resolve this issue.

The effect of eye contact (i.e., a smaller McGurk effect when the speaker’s eyes are closed) in the present study is rather small and not significant in all pairwise comparisons. However, the effect is significant when analysed with an ANOVA with higher statistical power (see Sect. Control condition: Static matched - Results) and the effect (or a tendency, at least) occurs in all of the three conditions, indicating that it is reliable. Future research could identify potential factors that might increase the size of the effect or abolish it altogether.

When it comes to the generalizability and applicability of findings from the McGurk illusion, one should bear in mind that the processing of the McGurk illusion does not necessarily generalize to natural forms of audiovisual speech processing (for reviews, see Alsius et al., 2018; Rosenblum, 2019) as, for instance, audiovisual sentence recognition abilities do not predict one’s susceptibility to the McGurk illusion (van Engen et al., 2017) and distinct brain regions are active during the perception of audiovisual speech and the McGurk illusion (Erickson et al., 2014). Further systematic studies are thus needed to investigate in how far findings from the McGurk illusion extend to audiovisual speech perception more generally.

The present findings contribute to an ongoing debate on whether multisensory integration is an automatic process or whether it can be affected by attentional processes (for reviews, see Talsma et al., 2010; Ten Oever et al., 2016; Spence & Frings, 2020). In line with earlier work (Alsius et al., 2005; Tiippana et al., 2004), our findings support the view that multisensory integration is susceptible to attentional manipulations in the case of audiovisual speech processing. In particular, our findings suggest that attentional focus on the speaker’s eyes affects the integration process in the McGurk illusion. Future studies could expand this work by investigating if attention-capturing stimuli in the auditory domain (e.g., a word spoken prior to the McGurk syllable) also affects the integration process or if the observed effects are specific to the visual modality.

In sum, the present study shows that speaker gaze affects audiovisual integration in the McGurk illusion. Our results thereby demonstrate that a speaker’s gaze behavior can dynamically capture an addressee’s attention, influencing how the addressee processes the speaker’s audiovisual speech signals. These findings reaffirm the power of speaker gaze to guide attention, showing that its dynamics can modulate low-level processes such as the integration of multisensory speech signals.