Reduced Eye Blinking During Sentence Listening Reflects Increased Cognitive Load in Challenging Auditory Conditions

Abstract

While blink analysis was traditionally conducted within vision research, recent studies suggest that blinks might reflect a more general cognitive strategy for resource allocation, including with auditory tasks, but its use within the fields of Audiology or Psychoacoustics remains scarce and its interpretation largely speculative. It is hypothesized that as listening conditions become more difficult, the number of blinks would decrease, especially during stimulus presentation, because it reflects a window of alertness. In experiment 1, 21 participants were presented with 80 sentences at different signal-to-noise ratios (SNRs): 0,  + 7,  + 14 dB and in quiet, in a sound-proof room with gaze and luminance controlled (75 lux). In experiment 2, 28 participants were presented with 120 sentences at only 0 and +14 dB SNR, but in three luminance conditions (dark at 0 lux, medium at 75 lux, bright at 220 lux). Each pupil trace was manually screened for the number of blinks, along with their respective onset and offset. Results showed that blink occurrence decreased during sentence presentation, with the reduction becoming more pronounced at more adverse SNRs. Experiment 2 replicated this finding, regardless of luminance level. It is concluded that blinks could serve as an additional physiological correlate to listening effort in simple speech recognition tasks, and that it may be a useful indicator of cognitive load regardless of the modality of the processed information.

Introduction

Understanding speech in everyday conversation involves working through acoustic challenges such as background noise, underarticulation, foreign accents, and potentially more so when experiencing hearing impairments (Peelle, 2018; van Engen & Peelle, 2014). Such factors can increase the cognitive demand placed on the brain, which refers to the amount of resources recruited for comprehension (Defenderfer, 2022). The distribution of these cognitive resources across different brain structures to overcome such challenges in pursuit of a listening goal is known as listening effort (Pichora-Fuller et al., 2016). To yield the highest task performance with the lowest energetic cost, the brain engages in cognitive resource allocation. The Framework for Understanding Effortful Listening (FUEL) interprets Kahneman's Capacity Model of Attention in the context of psychoacoustics, which demonstrate how listening effort is influenced by various elements such as task demands, hearing difficulties, the listener's motivation, momentary intentions, and more (Pichora-Fuller et al., 2016). Since speech comprehension cannot be predicted solely by audiometric measures (Humes et al. 2013), it is crucial to study it in the context of listening effort.

Within hearing research, pupillometry is a tool commonly used as an index for listening effort (Granholm et al., 1996). Generally, as an auditory task like speech recognition becomes more demanding, listeners exhibit greater pupil dilation (Piquado et al., 2010; Wang et al., 2018; Zekveld et al., 2010) up to a point where this relation breaks down (often indicating a disengagement from the listener). Since pupil diameter cannot be assessed when the pupil is occluded (Siegle et al., 2008), typical analyses usually treat blinks as annoyances, removing them completely from the data and interpolating throughout their duration (Mathôt & Vilotijević, 2023). This common practice aims to remove blink-related artifacts (i.e., transient changes in pupil diameter due to the absence of light for a short period) and preserve the integrity of the pupil traces over time (Grootjen et al., 2024). However, blinks themselves could be of primary interest if they somehow reflected listening effort, and we surmised that this might be the case if blinks were involved in cognitive resource allocation during an auditory task. But is there any precedence for this surmise?

The study of blinking behavior was originally and primarily investigated within the visual system. Early research with visual tasks revealed that endogenous blinks are controlled by a central information processing system based on cognitive states (Ponder & Kennedy, 1927). Experiments based on reading tasks and simulated surgery tasks demonstrated an inverse relation between blinks and cognitive load, where blinks were progressively suppressed as cognitive load increased (Drew, 1951; Wong, 2002). The blink rate modulation in these experiments was initially thought to result from the brain's attempt to minimize visual loss, but later studies involving only auditory stimuli revealed a similar phenomenon (Holland & Tarlow, 1972; Oh et al., 2012) suggesting that this is not a modality-specific phenomenon.

Research on blinks within the auditory literature is limited and has conflicting findings. A study investigating the functional assignment of blinks in complex auditory speech perception tasks in normal hearing adults showed a similar blink suppression pattern to the one uncovered in visual tasks (Kobald et al., 2019). That is, participants may suppress their blinking as a way of minimizing information loss, even if this information is of auditory nature and consequently not directly impacted by the act of blinking. Additionally, these researchers found that blink rates increased significantly immediately after a signal indicating a no-go trial that rendered all the subsequent trial information irrelevant. The elevated blink count observed right after a no-go trial may suggest that blinks function as part of a broader information chunking strategy rather than merely serving to prevent the loss of visual input. It may also signify a distinct information processing endpoint and a shift towards attentional disengagement (Kobald et al., 2019). Opposingly, normal hearing adults were also found to increase their blink rates as a function of cognitive load during an auditory oddball paradigm (Magliacano et al., 2020). These findings suggest instead a positive relationship between cognitive load and blinks, which was attributed to a shared neurobiological mechanism. More specifically, a central dopaminergic activity was proposed as instrumental in this blink-to-attention relation, where higher attentional demands imply increased dopaminergic activity leading to elevated blink rates (Magliacano et al., 2020).

These conflicting pieces of evidence may stem from protocols capturing distinct attentional processes. The FUEL model identifies four key contributors to the allocation policy: enduring dispositions which govern involuntary attention, momentary intentions, demand evaluation, and arousal effects (Pichora-Fuller et al., 2016). Hence, the allocation of resources may vary depending on the task properties and its associated requirements to enable goal achievement. Additionally, it may be that different blink parameters reflect distinct cognitive processes. For instance, spontaneous blinks may signal shifts in attentional demands, while blink suppression may reflect deep engagement in a task. This would seem reasonable given that the pupil response itself is known to be sensitive to different facets of the protocols, for example, speech intelligibility (Zekveld et al., 2010), noise type (Koelewijn et al., 2012), auditory features (Kramer et al., 2013), and divided attention (Koelewijn et al., 2014). Another possibility, quite likely, is that the discrepancies in the literature could be reconciled once accounting for different time windows in the analytical methodologies. For example, it is conceivable that blinks would be temporarily suppressed during stimulus presentation (requiring timely attention) and subsequently increased after stimulus offset or during the inter-trial interval (a bit like an overshoot phenomenon in response to the suppression).

Additionally, most studies did not report the procedures related to luminance levels. In many cases, there was no mention of lighting conditions at all. It may be implicitly assumed that experiments were conducted under “normal” or moderate indoor lighting; however, due to the lack of explicit reporting, consistency across studies cannot be confirmed. Therefore, variations in luminance levels in the testing room could be another reason for the inconsistencies in the literature. Indeed, it remains unclear whether specific blink patterns (while eye gaze is kept controlled) hold in very dark or very bright environments. On one hand, more blinks occur in bright light compared to medium light, due to a physiological response aimed at shielding the eye from light-induced discomfort or potential damage (Keane, 1979). On the other hand, blinks also increase potentially in darkness, because with a reduced need for sustained visual attention, the regulatory control over blinking relaxes, allowing for more frequent blinks (Ziv & Bonneh, 2021). On this basis, blink suppression may perhaps be easier to observe in bright or dark environments since the baseline blinking rate is overall higher. Surprisingly, to our knowledge, no blinking study has purposefully manipulated overall luminance level to assess potential changes in blinking strategies under various levels of auditory-induced cognitive load. Lack of information around this topic may contribute to confounding interpretations about cognitive load and attentional engagement in prior blinking studies.

We believe it is valuable to investigate blinking under clearly defined luminance conditions, as outlined in Zhang et al. (2022). This choice addresses the gap in the existing literature, where luminance conditions are often underreported or overlooked. Systematically examining (1) different time windows and (2) different luminance levels may provide a more comprehensive understanding of blinking behavior in effortful listening conditions.

This paper reanalyzes pupillometric data from a standard speech recognition paradigm conducted by Zhang et al. (2022), with a renewed focus on blinking behavior. The primary objective of the reanalysis is to demonstrate that further investigations of such existing pupillometric data can be performed relatively easily and can yield additional information to the researchers, either on their own or in conjunction with other pupillometry measures. Furthermore, this article aims to bring more clarity to the field by investigating fluctuations in blink rates under various degree of auditory-induced cognitive load. Based on the previous literature, we hypothesized that (1) blinks would be suppressed as task difficulty increased, that is, at lower signal-to-noise ratio (SNR). We also hypothesized that (2) blink occurrence would not be uniform throughout the duration of trials; rather, blinks would be specifically suppressed during sentence presentation, when more attention is required. Note that throughout the article, we operationalized blink suppression as a reduction in blink rate during sentence presentation relative to prior/post windows (not referring to any internal mechanism of neural or ocular behavior inhibition). Lastly, (3) we hypothesized that blink suppression would be more pronounced in bright and dark environments than in medium light, as the eyes may exhibit heightened sensitivity to these luminance extremes, leading to greater fluctuations in blinking behavior.

Methodology

Participants

For experiment 1, the sample consisted of 21 listeners (10 men; 11 women) between the ages of 18 to 49 years old, with a mean (SD) of 27.3 (8.6) years. For experiment 2, the sample consisted of 28 listeners (21 women; 7 men) between the ages of 18 to 51 years old, with a mean (SD) of 27.7 (9.6) years. Participants were either French (n = 7 in experiment 1; n = 10 in experiment 2) or English (n = 14 in experiment 1; n = 18 in experiment 2) native speakers; the study was always run in each participant's native language, and no investigation of language effects were conducted (materials being validated in each language). All participants were administered a pure tone audiometry to ensure they possessed binaural thresholds at or better than 25 dB HL at 0.25, 0.5, 1, 2, 4, and 8 kHz. All participants reviewed the experimental protocol, signed the written informed consent form, and received 15 dollars as compensation for their involvement in the study. As this study is a reanalysis of Zhang et al. (2022), we aligned our sample as closely as possible with theirs to facilitate comparison with their pupillometric data. Three participants were excluded from experiment 2 in the original study due to excessive missing data, including blinks and other factors. While examining these cases could be informative, doing so would result in a non-equivalent sample. For full inclusion and exclusion criteria, see Zhang et al. (2022). The McGill University Faculty of Medicine Research Ethics Board provided ethical approvals for this research under the reference A05-B11-18B.

Materials

The speech materials originated from the Institute of Electrical and Electronics Engineers for the English stimuli (Rothauser, 1969) and from the Hearing in Noise Test sentences for the French stimuli (Vaillancourt et al., 2005). Both were recorded by a male native speaker. All conditions except quiet contained a speech-shaped noise masker generated from white noise filtered with the long-term spectrum of the whole material (respectively either French or English). In experiment 1, conditions varied in terms of SNR: 0,  + 7,  + 14 dB, and a quiet condition in fixed illumination (75 lux). In experiment 2, sentences (same materials) were presented at 0 and 14 dB in darkness (0 lux), medium light (75 lux), and bright light (220 lux).

SNR manipulations (with the exception of the quiet condition) were implemented by raising the target level rather than the masker level to prevent listeners from anticipating the block difficulty before hearing the target. The audio stimuli were presented through the Beyer Dynamics DT 990 Pro headphone via an external soundcard (Edirol UA), calibrated at 65 dB sound-pressure level. The light level was measured and fixed with a luxometer (TES-1335) sensor positioned near the left eye of the participant and facing the screen to give an approximation of the luminance entering the participant's eye.

Apparatus

In both experiments, the infrared binocular eye tracker Tobii Glasses Pro2 was used to monitor eye behavior as seen in the left panel of Figure 1. They are equipped with four cameras, two dedicated to each eye, and an additional camera to capture the participant's field of view. Each eye camera operated at a sampling frequency of 50 Hz, interleaved by 10 ms, resulting in an effective sampling frequency of 100 Hz. The experiment was run in Matlab 2016b, using the Psychtoolbox and custom software.

Figure 1.

Experimental Setup Showing the Eye-Tracking Equipment and the Testing Environment.

Procedure

The experiment was conducted inside a sound-attenuated room at the Center for Interdisciplinary Research in Music Media and Technology in Montreal (https://www.cirmmt.org/), where participants sat 2 m in front of a 35-inch screen monitor as seen on the right panel of Figure 1. Participants were instructed to fixate on a black cross centered on the screen all throughout the experiment. Participants wore the infrared binocular eyetracker and headphones, and they were first presented with a demo of the experimental procedure, listening to five sentences at +14 dB SNR to familiarize themselves with the speech materials (excluded from the test).

In experiment 1, participants were instructed to listen to 20 sentences, all presented at the same SNR, and verbally repeat what they understood to the experimenter. The experimenter transcribed their responses to prevent participants from shifting their gaze between the fixation cross on the screen and the keyboard. This was repeated four times (with four different blocks of sentences) across the four SNRs (0,  + 7,  + 14 dB, and a quiet) summing up to a total of 80 sentences (or trials) completed by each participants. Each sentence was heard only once for each participant. The sequence of blocks was randomized across participants to prevent materials’ influence and order influence on the conditions (SNR or/and light). In contrast, the sentence order within each block (20 sentences presented at a given SNR) remained fixed across participants. The overall luminance level (combining room light fixtures and screen projection) was fixed at 75 lux throughout the experiment.

In experiment 2, participants received the same instructions as experiment 1, but with the additional note that luminance levels would be manipulated throughout the experiment and that they would have a one-minute light adaptation period before each block. The experiment followed a 2 × 3 orthogonal design, which included two levels of noise (0,  + 14 dB) and three luminance conditions (dark, medium, and bright, fixed at 0, 75, and 220 lux, respectively) summing up to a total of 120 sentences (or trials) completed by each participants. Prior to each block, the room and screen luminance levels were adjusted based on the randomization settings from the Matlab program. Luminance levels were kept constant within each block to prevent non-task-related responses. The sentence order within each block remained fixed, while the block sequences were fully randomized from one participant to the next for the same reasons as in experiment 1.

During each trial of both experiments, the speech-shaped noise masker began 3 s before the sentence onset, allowing time for pupils to stabilize (recovering from the previous trial) and establish a baseline pupil diameter measurement (1 s prior to sentence onset). Then, the sentence was presented alongside the continuous noise, and lasted on average 2.5 (SD = 0.3) seconds. The masker noise continued for 2 s after the sentence ended to allow the pupil response to peak to support the initial goals of the pupillometric investigation by Zhang et al. (2022). Following the masker offset, participants saw the black cross change into a circle at the center of the screen, prompting them to verbally repeat the sentence. The experimenter scored the participants’ verbal responses by counting the number of correctly repeated keywords before moving on to the next trial. Thus, each trial was initiated by the experimenter and separated by about 10–15 s.

Analysis

This paper's reanalysis used the unfiltered raw pupil data (one for each eye), as illustrated in Figure 2, where one trial is displayed. In the top panel, the blue and red lines correspond to the data for the left and right eye, respectively. Starting from the left side, the two bold vertical lines and annotations correspond to the window of time associated with the sentence presentation, where the onset is fixed, but the offset varies depending on the sentence's length. Similarly, the third and fourth bold vertical lines and annotations represent the onset of the verbal response window, initiated 2 s post sentence offset, and the end of the trial. Each trial was manually screened for blinks, defined as instances where the pupil trace dropped to zero or fell approximately three standard deviations (SDs) below the pre-blink baseline pupil size. This identification process allowed for a nuanced assessment, considering variations that might not reach zero but still indicated a blink. The screening relied on the expertise of the first author, who developed familiarity with typical blink patterns after reviewing numerous trials. Note that all trials were video-recorded, so any case of uncertainty could be watched closely to verify the blink behavior, its precise timing, and the gaze direction when needed (Figure 3).

Figure 2.

Data Analysis Methodology. Pupil Diameter Plotted as a Function of Time to Illustrate Four Instances of Blinks in a Given Trial (top). Blinking Probability Within Individual Trials Displayed as a Binary Vector Derived From the Onset and Offset of Each Blink Occurrence (Bottom).

Figure 3.

Example of a Blink Within a Given Trial Captured by the Video-Recording Device Within the Tobii Glasses.

In experiment 1, each of the 21 participants was presented with 80 trials, and in experiment 2, each of the 28 participants completed 120 trials. Each trial consisted of a single sentence, preceded by 3 s of baseline and followed by several seconds of post-sentence recording. For each trial, the onset and offset times of blinks were recorded in Excel. Blink durations within each trial were then converted into a binary vector system using MATLAB, where “1” indicated the presence of a blink and “0” indicated the absence of blinks as illustrated in Figure 2 (bottom panel). This process involved identifying the precise start and end times of each blink within a trial and translating that temporal information into a time-series vector. With these vectors, we calculated the total number of blinks per trial as well as the trial blink rate per second. All subsequent statistical analyses were performed using JASP.

To better understand how blink behavior varies in relation to auditory-induced cognitive processing demands, each trial was divided into three distinct time windows: (1) before sentence onset, (2) during sentence presentation, and (3) after sentence offset. Segmenting the trials in this way allowed us to isolate blinking activity during periods of auditory-induced cognitive load (during sentence presentation) and compare it to periods without such load (before and after the sentence). When a blink spanned two windows, it contributed proportionally to each window according to the degree of overlap in each. Blink rates per second for each time window were then averaged across trials within each noise condition (SNR) and for each participant, allowing for condition-level comparisons of blink dynamics across individuals.

Following this, a repeated-measures analysis of variance (ANOVA) was conducted to examine the effect of the independent variables, time window and SNR, on the dependent variable, blink rate (expressed as the number of blinks per second). Mauchly's test of sphericity was conducted systematically for all main effects and interactions with more than 2 levels, and degrees of freedom were adjusted with Greenhouse-Geisser when the sphericity assumption was violated. Post-hoc tests were performed to uncover the main effect of time window (pairwise comparisons across the three windows with Bonferroni correction) as well as the interaction between noise (SNR) and time window (simple effect of noise [SNR] in each window). Statistical analyses were conducted using an alpha level of 0.05 to determine significance. Effect size values (η²) were calculated and interpreted based on conventional benchmarks, with values of 0.01, 0.06, and 0.14 indicating small, medium, and large effects, respectively. The data screening, processing, and analysis were conducted in the same way for both experiments, with the addition of the luminance factor (three light conditions) for experiment 2. Additional correlational analyses were conducted to investigate whether the blink rate or its suppression could be related to (1) behavioral performance (i.e., intelligibility of the sentences) and (2) the pupil metrics taken from the original study (Zhang et al., 2022). These additional analyses were exploratory in nature and no Bonferroni correction was applied.

Results

Results in Experiment 1

A summary of the descriptive statistics, including mean and SD values for blink rates per second across conditions and time windows, is presented in Table 1. In experiment 1, all sphericity assumptions were respected (window: p = .105; SNR: p = .468; window × SNR: p = .279). The ANOVA revealed a main effect of window: F(2,40) = 25.1, p < .001, η²=0.236. This is because listeners blinked less when the sentence was presented, compared to before or after its presentation (0.13 or 0.13 blink per second less than pre and post, on average across SNRs, p < .001 in both cases, while pre vs. post did not differ). This pattern is depicted in Figure 4, where downward triangles represent the pre-sentence window, circles indicate the sentence presentation window, and upward triangles correspond to the post-sentence window. There was also a main effect of SNR: F(3,60) = 4.92, p = .004, η²=0.063, and an interaction between window and SNR: F(6,120) = 7.3, p < .001, η²=0.069. This is illustrated in Figure 4, where each color represent a different noise condition. The gap between the downward triangles (pre-sentence) and the circles (sentence presentation) is wider in more adverse SNR conditions such as seen in black at 0 dB. This gap progressively shrinks as the conditions become less difficult, like in the quiet condition depicted in green. Simple effects revealed no effect of SNR (i.e., no difference between 0,  + 7,  + 14 dB and quiet) in the prior window F(3,60) = 1.3, p = .285, η²=0.061 or in the post window F(3,60) = 0.367, p = .777, η²=0.018, but did reveal a significant effect during sentence presentation F(3,60) = 17.3, p < .001, η²=0.463.

Table 1.

Descriptive Statistics of Experiment 1 Including the Mean Blink Rate Per Second and Standard Deviation (SD) for Each Time Window and Signal-to-Noise Ratio (SNR).

Experiment 1—Descriptive statistics
Window	SNR	Mean	SD
Prior	0 dB	0.466	0.316
	7 dB	0.512	0.361
	14 dB	0.530	0.359
	Quiet	0.509	0.322
During	0 dB	0.260	0.256
	7 dB	0.364	0.311
	14 dB	0.399	0.331
	Quiet	0.468	0.340
Post	0 dB	0.505	0.312
	7 dB	0.489	0.318
	14 dB	0.518	0.337
	Quiet	0.499	0.323

Figure 4.

Blink Rates Across Temporal Windows for Each Noise Conditions of Experiment 1. Symbols Represent the Means and Error Bars Represent one Standard Error From the Mean.

Results in Experiment 2

A summary of the descriptive statistics, including mean and SD values for blink rates per second across conditions and time windows, is presented in Table 2. The sphericity assumption was violated for the effect of window (p = .034), light (p = .013), and SNR × light (p = .003) but not for anything else (p > .275). The ANOVA revealed a main effect of window: F(1.6, 43.9) = 24.1, p < .001, η²=0.124, which replicated the observation of experiment 1: there were fewer blinks during sentence presentation than in other time windows (0.12 or 0.14 blink per second less than pre and post windows, p < .001 in both cases, while pre vs. post did not differ). This pattern is depicted in Figure 5, where the downward triangles represent the pre-sentence window, circles indicate the sentence presentation window, and upward triangles correspond to the post-sentence window, with the black and red colors representing the SNR levels of 0 and 14 dB respectively.

Table 2.

Descriptive Statistics of Experiment 2 Including the Mean Blink Rate Per Second and Standard Deviation (SD) for Each Time Window, Signal-Noise-Ratio (SNR), and Light Level (Dark, Medium, and Bright Light Fixed at 0, 75, 220 lux, Respectively).

Experiment 2—Descriptive statistics
Window	SNR	Light	Mean	SD
Prior	0 dB	Dark	0.600	0.520
		Medium	0.527	0.527
		Bright	0.614	0.516
	14 dB	Dark	0.586	0.425
		Medium	0.657	0.591
		Bright	0.627	0.483
During	0 dB	Dark	0.452	0.510
		Medium	0.393	0.516
		Bright	0.454	0.572
	14 dB	Dark	0.522	0.482
		Medium	0.537	0.581
		Bright	0.531	0.581
Post	0 dB	Dark	0.644	0.484
		Medium	0.605	0.496
		Bright	0.650	0.527
	14 dB	Dark	0.603	0.464
		Medium	0.618	0.536
		Bright	0.584	0.481

Figure 5.

Blink Rates Across Temporal Windows for Each Noise Conditions of Experiment 2. Symbols Represent the Means and Error Bars Represent one Standard Error From the Mean.

Consistent with the results of experiment 1, an interaction between window and SNR was also replicated: F(2,54) = 16.3, p < .001, η²=0.023. The gap between the downward triangles (pre-sentence) and the circles (sentence presentation) is wider in more adverse SNR conditions as seen in the black symbols as compared to red. Simple effects revealed no effect of SNR (i.e., no difference between 0 or +14 dB) in the prior window F(1,27 = 4.1, p = .052, η²=0.021) or in the post window F(1,27) = 1.2, p = .277, η²=0.014, but was significant during sentence presentation F(1,27) = 9.0, p = .006, η²=0.091. There was no main effect of luminance and no interaction involving luminance. In other words, blink suppression (during sentence presentation and exacerbated at lower SNR) disregarded whether participants were in a dark, medium, or bright environment, resulting in a lack of window × luminance interaction (p = .580) or window × SNR × luminance interaction (p = .288).

Collateral Findings

Additional analyses were conducted to investigate whether the blink rate or its suppression could be related to (1) behavioral performance (i.e., intelligibility of the sentences) and (2) the pupil metrics taken from the original study (Zhang et al., 2022).

In terms of behavioral performance, at a SNR of 0 dB, no significant trend was observed in experiment 1 between intelligibility and blink rate in any of the windows, whereas consistent significant correlations were observed in experiment 2 for all windows (prior r² = .49, p < .001; during r² = .42, p < .001, post r² = .54, p < .001), whereby the listeners with the most difficulties in speech recognition also happened to blink more (Figure 6, left panels). It is not clear why this occurred in the second but not the first experiment. As this SNR level was replicated three times (in dark, medium, and bright luminance) in experiment 2, it may perhaps reflect a stronger statistical power. Performing similar correlations between intelligibility and blink rate at higher SNR levels revealed no significant trends, likely because recognition performance was too close to ceiling levels. Additionally, both experiments revealed no significant association between intelligibility and the size of the blink reduction during sentence presentation relative to pre and post windows (Figure 6, right panels). Thus, while an excessive amount of blinks could perhaps be indicative of a listener struggling, the blink rate reduction does not appear directly related to performance Figure 6 (right panels).

Figure 6.

Scatterplot of Blink Rates in Each Window (Left) and Blink Suppression Rate (Right) Across the 21 and 28 Individuals of Experiments 1 and 2, as a Function of Their Sentence Recognition Performance. The Blink Suppression Rate is Calculated From the Average of the Blink Rate Before and After the Sentence Presentation (top and Bottom Left) Minus the Blink Rate During Sentence Presentation (Middle Left); More Positive Values Indicate More Blink Suppression. All Values in Experiment 2 Were Averaged Across the Three Luminance Levels (Since This has no Effect on Intelligibility or Blinking Behavior).

General Discussion

This study reanalyzed pupillometric data from Zhang et al. (2022) to extract additional insights into blinking behavior, demonstrating that this practice can be done relatively easily while providing valuable information to researchers. Two experiments assessed blinking behavior during sentence listening under varying levels of noise and luminance: experiment 1 tested 21 participants across four SNRs in controlled medium lighting (75 lux), while experiment 2 tested 28 participants at two SNRs across three luminance conditions (dark, medium, and bright fixed at 0, 75, 220 lux, respectively). In experiment 1, fewer blinks occurred when the sentence was presented compared to before its onset and after its offset. Additionally, this phenomenon was exacerbated in more adverse SNR, when auditory-induced cognitive load was higher. Experiment 2 replicated these observations, and demonstrated that luminance levels did not impact the blink suppression phenomenon. Therefore, these findings confirmed our first two hypotheses postulating an inverse relation between blinking rates and auditory induced cognitive load, at least during sentence presentation (and see further details in the Appendix about the fact that this window is actually wider than sentence presentation per se). However, these findings rejected our third hypothesis to the extent that luminance level played no role (noting that baseline blink rate did not even increase in dark or bright luminance as was initially expected—and therefore we could not properly assess within the same individual whether the blink suppression would be exacerbated when the baseline blinking rate was higher). Additionally, analyses of behavioral performance and blink metrics were conducted and revealed significant correlations between intelligibility and blink rate across all windows in experiment 2, suggesting that listeners who have more difficulty recognizing speech also tend to blink more frequently. Interestingly, no significant associations were found between intelligibility and the degree of blink reduction during listening. These analyses of intelligibility and blink rate suggest that while an overall elevated blink rate might signal listener difficulty, the blink rate reduction during sentence processing does not appear directly related to performance. Overall, this study provides evidence supporting the role of blinks (but not blink suppression) as indicators of listening effort and cognitive engagement, robust to various luminance effects, making its use ecologically valid. Blinking could serve as a complementary measure to investigate effortful listening.

Several potential explanations could account for the blink suppression phenomenon and its exacerbation in more adverse listening conditions. The first explanation relates to the mechanism within the visual system that regulates the strategic timing of blinks to reduce visual disruption and ensure continuous information processing (Nakano et al., 2009). This mechanism has been observed in multiple studies involving various visual task structures, all showcasing a decreased blink activity at salient times (Fukuda, 2001; Ichikawa & Ohira, 2004; Pitigoi et al., 2024). This indicates a temporal assignment of blinks according to the needs of the visual environment. Given this, it is possible that a similar regulatory mechanism is at play within the auditory system, adapting the principles observed in vision to support auditory processing by optimizing blink timing to reduce disruptions in auditory attention. This internalization may have developed as a cross-modal generalization (El-Boustani et al., 2024) where the brain applies the same strategic blink control used in vision to auditory tasks, ensuring uninterrupted cognitive processing during critical listening moments. Notably, it is important to consider that the tasks conducted here were (perhaps) not purely auditory. The presence of a gaze-fixation component could have introduced a degree of visual engagement, potentially activating similar blink-regulating processes and further blurring the boundaries between visual and auditory mechanisms of blink suppression. Regardless, it is plausible that the auditory system employs a similar strategy to the visual system, adapting blink suppression to prevent attentional lapses and maintain optimal listening performance, particularly in challenging environments.

In line with this idea, other researchers have explored the neural correlates of blink rate modulations in visual tasks, investigating how activation or inhibition of certain brain regions accounts for the functional assignment of blinks. During a video-watching task, blinks are linked to increased activity in the default mode network, alongside a corresponding reduction in activity within the dorsal attentional network (Maffei & Angrilli, 2019). These findings suggest that spontaneous eye blinks may serve as brief disengagements from external tasks, reflecting a temporary shift in cognitive state towards an endogenous response. Such findings would explain why blinks are strategically suppressed during stimulus presentation, when sustained attention is required. As hinted by the inverse correlation between intelligibility and blink rate at least in experiment 2 (Figure 6), listeners who did blink excessively had poorer performance, hence the blink suppression could indeed be an effort to minimize momentary lapses in attention that would otherwise hinder performance.

When interpreting these findings, one might draw the conclusion that auditory attention and task performance are inherently better with eyes open than closed. However, this would be an oversimplification. This idea was challenged by empirical evidence from a study that examined auditory attention in tone-identification tasks where participants either had their eyes open or closed (Wöstmann et al., 2020). The findings revealed no significant effect of eye closure on sensitivity or response criterion for auditory attention and tone detection. Hence, we cannot assert that the cause of decreased blink activity is because attention is better when the eyes are open compared to closed. The underlying factor driving the blink suppression phenomenon is the momentary inhibition of brain regions associated with external attention that occurs during a blink. The brief disengagement from task-relevant processing that accompanies each blink presents a potential risk to performance, particularly in cognitively demanding tasks. As a result, the strategic suppression of blinks appears to be an adaptive response aimed at minimizing these transient lapses in attention, ensuring continuous engagement with the auditory stimuli.

The connection between blink suppression and attentional processes is further supported by studies with populations having attentional deficits. A study focusing on visual tasks revealed that unmedicated attention-deficit hyperactivity disorder (ADHD) participants demonstrated significantly higher blink rates than control groups, with blink rate decreasing notably after medication (Fried et al., 2014). Interestingly, another study on auditory-driven visual perception found that ADHD participants exhibited a larger attentional blink compared to control participants (Khetrapal, 2007). Attentional blinks are slightly different than eye blinks. They are defined as temporal cognitive limitations implying that resources are captured by a first stimulus which only gradually decays in a sensory store and interferes with the perception of a second stimulus. To clarify, attentional blinks do not involve eye closure, but do reflect a similar task disengagement. These findings solidify the notion that blink suppression is closely tied to general attentional control, rather than being a peculiarity of the visual system. The higher blink rates observed in individuals with attentional deficits emphasize the role of blink suppression as a compensatory mechanism to maintain focus and prevent momentary disengagement across sensory domains.

Different Measures to Auditory-Induced Cognitive Load

In this study, additional analyses were conducted to investigate whether the blink rate or its suppression could be related to the pupil metrics taken from the original study (Zhang et al., 2022). Among pupil metrics, one might naturally want to choose peak pupil dilation (PPD). In simple listening tasks such as this one, PPD would generally be informative of listening effort, while the pupil baseline would be less relevant. No significant trends were observed between blink metrics (whether overall blink rate or blink suppression) and either PPD amplitude or PPD latency. However, asillustrated in Figure 7, a surprising association emerged between baseline pupil diameter and blink rate (averaged over the pre- and post-sentence windows, but not during listening). In experiment 1 (fixed at 75 lux), individuals with larger baseline pupil diameters tended to blink more. Experiment 2 did not replicate this finding under the same luminance condition, nor in darkness (0 lux), but did reveal the same pattern in bright light (220 lux). Conversely, baseline pupil diameter was never related to blink suppression. These findings add to the complexity of interpreting the meaning of blinking behavior and pupil behavior.

Figure 7.

Scatterplot of Averaged Blink Rate (top) and Blink Suppression (Bottom) in Experiment 1 (Most-Left) and Experiment 2 at Each Luminance Level (Middle-Left, Middle-Right, Most-Right) as a Function of Baseline Pupil Diameter Measured Just Before the Sentence Presentation. Here, Values are Averaged Across SNR Levels (Since They had Little Effect on Baseline Pupil).

This pattern is puzzling for several reasons. First, the baseline measurements of the pupil are not supposed to be very informative in simple tasks (in contrast to complex tasks where long-term changes in pupil across items in a list can be revealed as baseline differences—see Zhang et al. (2021) and (2024) for such examples in otherwise similar setups). Therefore, it is surprising to observe associations between blinking behavior and pupil baseline, but none with PPD. We cannot offer any convincing explanation as to the current lack of association between blinking behavior and PPD. Both are theoretically related to effort, but it is possible that blinking act (also) as a cause rather than merely a consequence of task difficulty (and we are currently in the process of testing whether blinks cause short information losses which would potentially make the task even harder).

The second puzzling aspect is the dependency of this correlation on luminance level. If it was purely due to the abundance of light, rendering the eyes drier for example, one might have expected (1) a higher blink rate in bright than in dark luminance (but this was not the case) and (2) an inverse relation between blink rate and pupil diameter in each panel (but this was also not the case). If it was purely due to cognitive reasons (i.e., internally generated), both blink rate and pupil dilation would be expected to show a positive relation, reflecting the degree of cognitive effort in a given listener. In that case, the association would expectedly also be evident, and perhaps even more pronounced, in darker conditions. Thus, this pattern results from an interaction between light settings, pupil dilation, and blink rate which is not trivial. Zhang et al. (2022) had concluded that pupil metrics are less sensitive in very dark or bright environments. In this study, we conclude that blinking behavior is relatively more robust to such luminance differences. Overall, the blink suppression phenomenon was observed similarly in dark, medium, and bright environments and it was never correlated with any pupil metric. This gives further support to the notion that blinking behavior is worth investigating on its own as it could reveal independent information from pupil behavior. The complexity in how blinking relates to luminance and pupil measures underscores the need to consider blinks as more than just noise in pupillometry.

With this study, we have provided evidence that blinks (or more precisely, a reduction in blink frequency during specific time windows) can serve as an index of effort, akin to pupil dilation. Consequently, the common practice of interpolating through blinks in pupillometry may limit researchers’ ability to capture effort-related dynamics at a global level. However, whether this should be considered a confound is debatable. Rather than interfering with pupil measurements directly, blink-related and pupil-based indices of effort may operate on different scales; for instance, the pupil would not necessarily have dilated had the blink not occurred. Thus, interpolation may not inherently reduce statistical power. On the contrary, in some cases, it may lead to an overestimation of cognitive load effects.

To understand this more clearly, it is worth considering the full dynamics of pupil behavior during and after a blink. Specifically, the pupil constricts and re-dilates within a few seconds after each blink, a phenomenon known as blink-locked pupillary responses (BPR) (Yoo et al., 2021). When blinks occur more frequently at specific time points, such as approximately 0.5 s after trial onset in our study (see Figure 8 and 9), and given that it takes several seconds for the pupil to return to its pre-blink diameter following the BPR, pupil size can be underestimated within this window. For example, had our sentences begun 1 s after trial onset, the baseline pupil size would have been underestimated, resulting in an inflated PPD estimate. This would not have been due to increased cognitive load, but simply because the pupil had not yet recovered from the preceding blink. In our design, sentences were presented 3 s after trial onset, which is likely sufficiently delayed from the blink rate peak to minimize this potential confound. Nonetheless, the degree to which BPR contaminates our findings, or the broader pupillometry literature, remains uncertain. This influence is highly contingent on task structure, the pacing of stimulus presentation relative to blink timing, and individual variability across participants (Daniels et al., 2012).

In addition to these methodological considerations, it is important to recognize the distinct trade-offs between blink-based measures and pupillometry. A key strength of blink metrics is their robustness to eye movements as compared to pupillometry which is susceptible to the foreshortening effect, where changes in gaze angle distort the apparent pupil area, leading to systematic underestimations even when actual size remains constant (Laeng & Mathôt, 2024). This makes blink metrics particularly advantageous in paradigms involving frequent eye movements, free-viewing conditions, or populations that struggle to maintain stable gaze, such as infants, young children, or certain clinical groups. Contrastingly, a primary shortcoming of blink metrics is their relatively low temporal resolution compared to pupillometry. While pupil diameter can be measured continuously with millisecond-level precision (Zhang & Emberson, 2020), allowing researchers to detect rapid event-related changes in cognitive load, blink rate typically requires aggregation over longer time windows or across trials as seen in this study. Given the strengths and weaknesses of blink measures and pupillometry, a multimodal approach that incorporates both blink and pupil measures may offer a more comprehensive and nuanced assessment of auditory-induced cognitive effort. In fact, even other ocular biomarkers (such as micro-saccades, see Cui & Herrmann, 2023) could be useful for capturing listening effort. Leveraging all methods would allow researchers to capture a broader spectrum of cognitive processing characteristics, improving the interpretability and ecological validity of psychophysiological studies.

Potential Limitations

Finally, certain limitations should be acknowledged. First, one limitation is the manual screening of the data that relied on the expertise of the first author. While manual screening allows for nuanced decision-making and the application of domain-specific knowledge, it introduces the possibility of subjective bias. Differences in judgment, even among experienced researchers, could impact the consistency and reproducibility of the results. Additionally, the manual approach is time-consuming. It would be valuable to conduct a systematic comparison of different tools for automatic blink detection and what they require as inputs (only the pupil diameter signal, or combined with gaze, or the full pupil videos such as OpenCV see Pimplaskar et al., 2015; or even the whole face video recordings such as MediaPipe, see Delgado Gomez et al., 2024) and compare them with a manual approach like we did here, to understand what may be preferentially used in any given situation. However, this falls beyond the scope of the current study, which was designed primarily as a proof of concept. Second, the present conclusions hold for a simple task, where participants only had to listen and repeat the sentences they heard. It remains to be seen whether this effect persists when tasks become more complex, for example, involving short-term memory (e.g., Zhang et al., 2021, 2024 using otherwise very similar paradigms and apparatus). If blink metrics were revealed to be informative across a range of task complexities (in contrast to PPD for example), it would strengthen their potential utility in cognitive load assessment. This would open up practical applications, based on machine learning techniques, to analyze blink data more efficiently. Given the tiny fluctuations in pupil size due to cognitive demands (e.g., a fifth or even a tenth of a millimeter) compared to the distinct and measurable nature of blink patterns, blinks indeed offer a promising alternative to pupillometry. Expanding this research to specific populations, such as cochlear implant users or individuals with hearing aids, would provide valuable insights into how blink dynamics vary with cognitive demands. Understanding whether these populations exhibit more or fewer blinks under effortful listening conditions could enhance assistive technologies and interventions.

Conclusion

The present study demonstrated that blinks reflect the listening effort experienced by participants in auditory tasks. In particular, blinks were significantly suppressed during a sensitive information time window, and especially more so in adverse listening conditions. The lack of luminance effects highlighted the robustness of this measure in simple tasks, and emphasized its potential use in ecological situations. However, further research is required to uncover blink strategies in more complex tasks, and with diverse populations.

Appendix

In this appendix, we demonstrate that the results stated in the main article are actually an underestimation of the true effect size of blink suppression, since the prior and post windows encompassed some edges close to the sentence where blinks were still actively suppressed. Typically in pupillometry, a 1-s baseline is recommended just before sentence onset, and similarly a wait-peak period is recommended after the sentence offset as the dynamics of pupil dilation are rather sluggish. Drawing from this practice, we were curious to investigate whether the blink suppression phenomenon applied to the sentence duration exclusively, or whether it extended on both edges of the sentence. To do so, we derived five time windows, which were (1) inter-trial, (2) baseline period, (3) during presentation, (4) wait-peak period, (5) repeat period (when participants verbally repeated the words to the experimenter). These time windows and their respective symbols are visible in Figure 8 (experiment 1) and Figure 9 (experiment 2). The upper panels of both Figures 8 and 9 demonstrate a diminished blink rate during sentence presentation, but also during the baseline and the wait-peak periods corresponding to the edges of the sentence. Blinks occurred more within the first second of a trial and 8–9 s after trial onset, as indicated by the spikes around 0.5 and 8.5 s relative to trial onset for both experiments, as seen in the bottom panels (Figures 8 and 9). This suggests that the results presented in the main article actually represent an underestimation of the effect size of blink suppression, as the prior- and post-sentence windows included times near the edges of the sentence during which blink suppression was still actively occurring.

Figure 8.

Blink Rates Across Additional Temporal Windows for Each Condition of Experiment 1.

Figure 9.

Blink Rates Across Additional Temporal Windows for Each Condition of Experiment 2.