Individual Differences in Cognition and Perception Predict Neural Processing of Speech in Noise for Audiometrically Normal Listeners

Flanker test

The Flanker test measures visuospatial interference, selective attention, and inhibitory control (Zelazo et al., 2013). We used a computerized version of the task using the PEBL software (Psychology Experiment Building Language; Mueller and Piper, 2014). In the task, five light gray arrows were displayed in the center of the screen against a black background. The arrows were either all pointing in the same direction (congruent condition) or with the center arrow pointing in the opposite direction to the flanking arrows (incongruent condition). Participants were asked to press the left or right shift keys on a keyboard corresponding to the direction of the center arrow. A fixation cross appeared for 500 ms before the arrows were presented, and participants had 2 s to respond before the trial timed out. A blank screen was shown for 2 s before the next trial began. Participants completed 8 practice trials, with 2 trials for each combination of conditions (left/right, incongruent/congruent), followed by 80 test trials (20 per condition). The interference effect was calculated as the median reaction time difference between incongruent and congruent trials.

Stroop

The Stroop test is a long-standing, well established assessment of linguistic interference and response inhibition (Stroop, 1935; Strauss et al., 2006). In our study, we utilized a computerized version of the Victoria Stroop Test (VST; Strauss et al., 2006), implemented using PEBL. The VST was selected for our study due to its shorter test administration time compared with other standardized Stroop tests. Evidence suggests that shorter test durations may be more sensitive to individual differences on this task (Klein et al., 1997). The VST has been proven effective with a wide range of adults, from 18 to 94 years old (Troyer et al., 2006).

As with other Stroop versions, the VST involves color identification of items presented to the participant, including a color word interference task. The VST is divided into three blocks always presented in the following order: colored dots (Part D), neutral words (Part W), and color words presented in contrasting text colors (Part C). Each block displays 24 items in a 6 × 4 grid on a gray background. Items are colored red, blue, green, or yellow, and participants press a corresponding number key to identify the color of each item. A square is shown around the current item trial, and it only moves to the next item with a correct response. The square flashes white when an incorrect response is made. Each color is used six times, and the item colors and the color-number keypad pairs are pseudorandomized for every participant (one of each color in each row). Part D serves as a practice block for color-number mapping and is not scored here. For Part W (neutral words) and Part C (color words), participants must name the colors in which the words are printed, disregarding the verbal content printed in lowercase letters (neutral words, when, hard, and, over; color words, red, green, blue, yellow). In Part C, the color name never corresponds to the text color, thus requiring individuals to inhibit an automatic reading response and only identify the text color itself. The interference effect is measured as the difference in median reaction times between Part C (i.e., the linguistic interference task) and Part W (i.e., the control task).

Reading span

Working memory ability was assessed using the reading span test (Daneman and Carpenter, 1980), as employed in previous research (Lunner, 2003; Rudner et al., 2012; Dryden et al., 2017). We administered a computerized version of the reading span via the PEBL program. The PEBL reading span test was developed for automated scoring as described by Unsworth et al. (2009). In this task, participants read sentences displayed on the computer screen and indicated whether each sentence made sense by pressing a key (i.e., in grammatical, coherent, or plausibility sense). Sentence lengths ranged between 9 and 14 words. After the sentence, a letter was presented on the screen for 1 s, and participants were required to memorize the sequence of letters presented across sentence–letter pairs. The task consisted of blocks of 3–7 sentence–letter pairs, with each block length presented three times in a random order throughout the test. At the end of each block, participants selected the correct sequence of letters from a 4 × 3 grid displayed on the screen and were instructed they could guess or leave a blank in the letter sequence if they were uncertain. They were then provided feedback regarding both sentence judgment and letter sequence accuracy.

Prior to the test, participants completed two practice examples each for sentence reading, letter memorization, and sentence–letter memorization pairs. The average reading time during these practice trials determined the timeout duration for the test blocks (i.e., timeout = 2.5 × average practice reading time; Unsworth et al., 2009). If participants did not respond to a sentence before the timeout period, the sentence trial was marked incorrect, and the subsequent letter was shown on the screen. Throughout the test, sentences, letters, and feedback were shown in white text on a black screen background.

Working memory span was determined by calculating the average accuracy of letter recall for three- and four-letter trials, when any letters were correctly recalled in the correct order (i.e., each correct letter contributing to the score). Though earlier research has suggested a working memory capacity of up to ∼7 items (Miller, 1956), recent studies indicate that this capacity is lower and can vary based on task demands and circumstances (Cowan, 2015). Consistent with findings indicating working memory capacity is ∼4 items (for reviews, see Cowan, 2001, 2015), our findings revealed that the accuracy for the three- and four-letter trials provided more stable and informative data for our model compared with the average accuracy across 3–7 letter trials (pilot data not shown). Therefore, we used the average accuracy from the three- and four-letter trials as our primary measure of working memory span.

TMT

The TMT (Reitan, 1958) was used to assess cognitive flexibility and executive function. In TMT-A, participants connect lines sequentially between numbers, while TMT-B requires them to alternate between numbers and letters (e.g., 1-A-2-B). Participants were given TMT-A followed by TMT-B, with a short practice at the beginning of each. The test was administered via PEBL, with white nodes displayed on a gray background and black text for letters and numbers. Participants clicked on the nodes in the correct sequence as quickly and accurately as possible, and lines were drawn on the screen to connect consecutive nodes upon correct selection. Given known limitations of the original Reitan-developed TMT-A/B (i.e., length differences between the nodes of the A and B versions; Gaudino et al., 1995), our version includes fixed node positions as mirrored images across TMT-A and TMT-B to ensure similar distance lengths for the two tests (Strelcyk et al., 2019). Any differences in reaction time are therefore presumed to reflect processing differences across the A and B versions. Executive function and cognitive flexibility were measured as the difference in median reaction times of TMT-B and TMT-A conditions.

TFS

The ability to localize low-frequency sounds along the horizontal plane relies on interaural time differences (ITD), which are important cues for distinguishing multiple sounds such as recognizing speech from a target talker amidst spatially separated background noise or other interfering sounds (Moore, 2021). For sinusoidal signals, ITDs are equivalent to interaural phase differences (IPDs), and the ability to discriminate changes in IPDs depends on binaural sensitivity to TFS. To assess IPD discrimination abilities, we used the TFS-LF (low-frequency) test (Hopkins and Moore, 2010), which is implemented in a computerized software program developed by Sęk and Moore (2012, 2021). In the TFS-LF test, participants are tasked with discriminating sinusoidal tone bursts with adaptively varying IPDs, where the tone envelopes are synchronized across both ears, ensuring that task performance relies on sensitivity to TFS (IPD). TFS-LF administration was consistent with previous work and is briefly described below (Hopkins and Moore, 2010; Sęk and Moore, 2012, 2021; Füllgrabe and Moore, 2018).

IPD discrimination was evaluated at test frequencies of 250 and 500 Hz, consistent with previous studies (for review, see Füllgrabe and Moore, 2018). The intensity level was set individually for each ear 30 dB above the participant's pure-tone thresholds at the tested frequencies (i.e., 30 dB sensation level at 250 and 500 Hz). The TFS-LF task consisted of a two-interval, two-alternative forced–choice design, where each interval included four successive tones. In one of the intervals, all four tones had a consistent 0° IPD, while in the other interval, the IPD alternated between 0° and ϕ across the four successive tones. A 0° IPD is perceived as a tone localized to the center of the head, whereas a large IPD is perceived as a tone lateralized (or as more diffuse) toward one ear. Participants were asked to identify the interval in which the tones sounded different or appeared to change in some way. Each tone had a duration of 400 ms, with 20 ms rise and decay times and a 300 ms interinterval gap between tone blocks. Feedback was provided after each trial. The IPD ϕ was adaptively varied, starting from 180° (maximally lateralized), using a two-down, one-up procedure to estimate a threshold corresponding to 71% correct responses. That is, the value of ϕ is divided by a factor k after two successive correct responses or multiplied by factor k following one incorrect response. Before the first turn point, k = 1.253; between the first and second turn points, k = 1.252; and k = 1.25 after the second turn point. The threshold was computed as the geometric mean value of ϕ at the last six turn points, and the final TFS-LF measure was then calculated as the average across both 250 and 500 Hz frequencies.

Additional tests

During the cognitive session, participants also completed two additional tests: (1) a pitch discrimination task and (2) a speech recognition in noise test, which assesses SiN perception. These measures were not included as predictors in the current analysis, as our primary focus was on evaluating variables that most closely align with existing research on predictors of SiN perception (Akeroyd, 2008; Shinn-Cunningham, 2008; Janse, 2012; Perrone-Bertolotti et al., 2017).

Stimuli

The stimuli consisted of short fairytale stories available in the public domain (available for listening on www.librivox.com). Stories were selected to have content that was unlikely to be widely recognized, reducing the chances of participants relying on modern renditions for comprehension recall (cf. stories popularized by children's books or movies). Additionally, each story was chosen to have a similar narrative arc—i.e., up through rising action, climax, or early falling action—within a 7.5 min block to ensure sufficient and consistent content for probing comprehension. Thirteen stories were piloted with four raters who evaluated them based on factors such as interest and engagement, ease of understanding, complexity of grammar or writing style, familiarity (i.e., low scores preferred), and suitability for an adult audience (not only for children). The seven highest-scoring stories across these criteria were selected for use in the experiment.

The short stories were revised to incorporate monosyllabic color words as targets for the experiment (e.g., red, green, black, white, etc.). The color words were integrated into the story itself rather than being inserted as unrelated, random target words (e.g., “…with a narrow blue ribbon over her shoulders”). A male native English speaker with an American Midwestern accent recorded the modified stories using a Shure KSM244 vocal microphone (cardioid polar pattern, high-pass filter at 18 dB per octave with an 80 Hz cutoff) and Adobe Audition [sample rate, 48,000 Hz; 32 bit depth (float)] in a soundproof, quiet room. Silent gaps in the recordings longer than 500 ms were shortened to 500 ms, following procedures similar to those in previous studies (Broderick et al., 2018; Teoh and Lalor, 2019; Teoh et al., 2022). Finally, the edited recordings were cropped to 7.5 min per story.

Cheech

The short story auditory stimuli were subsequently modified using a patented technique known as “Cheech” (“chirped-speech”), which is intended to elicit robust auditory evoked potentials from the brainstem to the cortex (Backer et al., 2019; Miller et al., 2020). In addition to providing insight into multiple levels of neural processing simultaneously (e.g., ABR, MLR, LLR), Cheech provides greater acoustic control over the specific acoustic parameters in naturalistic speech used to evoke the brain responses. We describe the details in this paper for completeness, but some of the details may not be relevant for LLRs. In Cheech, some of the glottal pulse energy within specified frequency bands is replaced with narrowband synthetic chirps (for more details, see Backer et al., 2019). These chirps are aligned with the natural glottal pulse energy in voiced speech, creating an acoustically fused auditory perceptual object. This modification maintains the naturalistic qualities and linguistic content of speech while introducing sufficient transient chirp activation to measure auditory brainstem, thalamic, and cortical responses.

The process begins by identifying voiced epochs where chirps will be inserted. Audio was filtered from 20 to 1,000 Hz, and voiced periods of at least 50 ms were defined when the speech envelope between 20 and 40 Hz exceeded a threshold of ∼28% of the overall speech root-mean-square (RMS) amplitude. The timing of the glottal pulses within these voiced periods was determined through a speech resynthesis process using a custom MATLAB code (The MathWorks Inc., 2021) and the TANDEM-STRAIGHT toolbox, which retains the original speech periodicity characteristics (Kawahara et al., 1999, 2008; Kawahara and Morise, 2011). The continuous speech was then refiltered into alternating, octave-wide frequency bands, 0–250, 500–1,000, 2,000–4,000, and 11,000–∞ Hz, with chirp energy constrained to frequency bands of 250–500, 1,000–2,000, and 4,000–11,000 Hz. The chirp and speech bands in the alternating, interleaved octave-wide bands were then added together. Each mono track was then duplicated to form stereo audio. Before the Cheech synthesis, audio files were normalized to ensure consistent chirp amplitudes relative to the speech levels across each story, and the RMS amplitudes were rechecked across all Cheech-modified stories to ensure equivalence after the Cheech process was completed.

The timing of the chirps varied with the natural fluctuations in the running speech, with special modifications to optimize measurement of auditory event-related potentials. Specifically, when voicing energy was detected, chirps were inserted at minimum 18.2 ms apart (55 Hz), skipping glottal pulses that occur within the 18.2 ms window. The first chirp in any sequence was always followed by a minimum interval of 48 ms before the next chirp was presented, skipping glottal pulses in between. This longer ISI was designed to enhance the measurement of middle-latency responses (MLRs); we use this “first chirp” to elicit LLR N1s as well.

In the dual-talker conditions, one story from each pair was pitch and vocal tract modulated using MATLAB STRAIGHT prior to the Cheech process. Specifically, stories intended to represent a female voice were pitch modulated to an average fundamental frequency (F0) of 180 Hz, while the original male voice audio had an average F0 of ∼128 Hz. By modulating the F0, key speaker characteristics such as phrasing, intonation, and pacing were maintained, as all stories were initially recorded by a single talker. Additionally, this pitch modification created a perceptual distinction between the two voices, introducing a combined voice gender and spatial release from masking effect. This effect, which is consistent with prior findings (Oh et al., 2021, 2022), facilitated better differentiation of the spatially separated voices and improved task performance during pilot testing.

Continuous multitalker spatial attention task

The Cheech stimuli were spatially filtered using head-related transfer functions (HRTFs) from the SADIE II database (Armstrong et al., 2018) to simulate audio at ±15°. This spatialized audio was mirrored by two symbols on two computer screens centered ∼15° to the left and right of the participant's midline. The symbols “<” and “>” indicated the target story's location on the left or right, respectively, while a “+” symbol denoted the location of the target story, as shown in Figure 1. The visual symbols were identical in size, color, and line lengths to ensure consistent luminance for eye-tracking measurements (not analyzed in this report). The custom MATLAB Psychtoolbox code (Brainard, 1997; Pelli, 1997) was used to program the spatial and visual stimuli to switch between left and right locations 75 times over a 7.5 min period, with each switch occurring approximately every 6 s (random distribution of 6 ± 1 s). The visual icons switched instantaneously at each switch cue, while the audio crossover fade time was set to 35 ms, starting at the onset of each switch cue. During piloting, this fade duration was found to effectively minimize audio artifacts, such as “pops,” while also reducing the perception of a gradual audio glide between spatial locations (i.e., as sudden of a jump as possible without introducing noticeable distortions).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

Continuous multitalker spatial attention task procedure. The figure illustrates the visual cues presented on screen to participants during the task while they listened to a series of short stories. The target and masker (when present) alternate between spatial locations. Only the target story is present in the mono-talker condition, but the spatial location of the talker still switches sides as shown by the red line. Participants are instructed to focus on the side of the screen indicated by the “+” symbol, which marks the location of the target story. They are also instructed to press the spacebar when they hear a color word spoken by the target speaker (red “X” symbols) and ignore any color words in the masker story (blue “O” symbols).

Seventy-five switches were placed in the story, occurring every 6 s on average. Twenty-five of those switches were specifically timed to precede the color words. The time intervals between these switches and the color words were 0.125, 0.25, 0.5, 1, and 2 s, with the intervals randomly assigned to the color words so that each interval duration occurred exactly five times. These intervals were determined based on preliminary piloting that aimed to evaluate whether brief lapses in attention following a shift in spatial focus—analogous to “attentional blinks”—affected participants’ ability to detect color words; these results will be presented in a subsequent paper. Additionally, the 25 color words were evenly distributed between left-to-right and right-to-left switches within each story, minus the one remainder. Onset timings for the color words were calculated using Gentle, an audio forced aligner (https://github.com/lowerquality/gentle). The Gentle forced aligner software finds the start and end points of spoken words, allowing us to pinpoint exactly when the color words start, which was needed for determining switch times.

Experimental conditions

After cropping the audio, the seven stories were assigned to specific experimental conditions: two stories presented alone (“mono-talker”), two sets of paired stories (“dual-talker”), and one for initial practice. Half of the stories (one mono-talker and one dual-talker pair) had modified Cheech characteristics to study their effects on brainstem encoding (results forthcoming in future reports); however, this modification did not impact cortical responses, so we used all six nonpractice stories for our N1 analyses. Since our focus is on predicting SiN performance, this paper primarily examines the dual-talker conditions (excluding Fig. 4 and associated analyses). In the dual-talker paradigm, each story was presented twice—once as the “target story” and once as the “masker story” within its pair. This experimental manipulation allowed us to assess the effects of directed spatial attention, as participants were instructed to either “attend” to or “ignore” the same audio depending on its role in the trial. The order of stories and conditions was counterbalanced across participants using a Latin square design to minimize potential sequence-dependent learning effects. Additionally, no pair of stories was presented consecutively.

Story pairs in the dual-talker condition required additional considerations. As described above, Cheech, one voice in each story pair was pitch and vocal tract modulated to simulate a female speaker. Consistent with findings on the combined effects of voice gender and spatial release from masking (Oh et al., 2021, 2022), two distinct talker F0s facilitated better listener differentiation between spatially separated voices and improved task performance during piloting. The mono-talker stories were presented in their original, unmodified male voice. To minimize the influence of stimulus variability on the N1, the analysis in this paper focused solely on the male target and male masker conditions. Due to the naturalistic embedding of color words within the stories, some overlap occurred in the dual-talker pairings, i.e., color words in the target and masker speech occurred in close temporal proximity. Story pairs were selected to minimize color word overlaps, resulting in six overlapping words that were excluded from the behavioral analysis (four from one story pair and two from the other pair). A few target story switch times were adjusted to ensure that no color words in the masker story occurred during the switch (i.e., masker color word onset was not between −0.5 and 0.1 ms from a switch time).

Stimulus presentation equipment

Participants were seated in a soundproof, electromagnetically shielded booth ∼176 cm from two computer monitors. The subjects and cables were positioned at a sufficient distance from the monitors. Testing confirmed that this arrangement effectively prevented any electromagnetic interference from the monitors. The Dell UltraSharp monitors had a 60 Hz refresh rate with screen settings maintained at standard factory defaults. Stimulus presentation, including both audio and visual components, was controlled using a custom MATLAB and Psychtoolbox code. Auditory stimuli were routed from a Hammerfall audio card to an RME Fireface UFX II audio interface, then through an RME ADI-2DAC FS headphone amplifier, and finally delivered via ER-2 insert earphones (Etymotic Research/Interacoustics). The ER-2 insert wires were electromagnetically shielded up to the transducer box to minimize EEG artifacts (Campbell et al., 2012). The Cheech stimuli were presented at 75 dB SPL.

In addition to the two-channel audio tracks, a pair of channel outputs carrying trigger events was routed from the Fireface audio interface to two Brain Products StimTraks. These StimTraks monitored triggers for the target and masker stories, respectively, converting the audio trigger signals into pulses sent to a TriggerBox (Brain Products). This configuration ensured that audio playback was synchronized with the chirp triggers embedded during the Cheech process, minimizing latency jitter, and maintaining sample-level synchronicity. Triggers managed by Psychtoolbox—such as participant responses, block sequence triggers, and spatial location switches—were also routed to the TriggerBox through the computer parallel port. The TriggerBox then transmitted both parallel port and StimTrak triggers to the BioSemi signal receiver box (refer below, EEG recordings). To calibrate our stimuli, we used a Larson-Davis PRM902 preamplifier connected to a Larson-Davis 2221 preamplifier power supply, paired with a B&K Ear Simulator Type 4157 (IEC 60318-4 Coupler). Audio stimuli were calibrated using either C-weighting (for Cheech stimuli) or Z-weighting (for cognitive and SiN tests), with a default preamplifier gain of 20 dB applied. Signal intensity measurements were then analyzed using a custom MATLAB script to ensure accuracy and consistency across all stimuli.

Attention task procedure

The order of events within each block is as follows: a 1 min rest period (for resting state EEG analysis), 7.5 min of audio presentation, comprehension questions, subjective workload questions, a flash reaction time task, and a break period. During the rest period, participants were instructed to remain still with their eyes open. They were instructed to attend to the target story as the auditory stream alternates between left and right spatial locations. They must immediately shift both their selective attention and eye gaze to the fixation point (“+”) on the computer screen as it switches the cue. Participants were also instructed to press the spacebar upon hearing the embedded color words in the story as quickly and accurately as possible. In dual-talker condition blocks, they were instructed to ignore color words from the masker story and only respond to those from the target story. Following the audio presentation, participants completed various behavioral tasks, including comprehension questions, subjective task workload assessments, and reaction time tests which are not discussed in this paper. Finally, participants were given a break before starting the next block. The entire experiment presentation was controlled by a custom MATLAB and Psychtoolbox code.

Measure of SiN perception

SiN perception during listening was measured by the correct identification of color words embedded in the target story, referred to as “hits.” A hit was defined as pressing a designated keyboard button within 2 s after the onset of a target word. Any color words in the dual-talker condition that occurred within ±2 s of each other in the target and masker were excluded from analyses. Specifically, the relevant measure was the proportion of correct hits out of the total possible, nonoverlapping target words (n = 25 in the mono-talker condition, n = 21 or n = 23 in the dual-talker condition, depending on the story pair).

EEG recordings

EEG recordings were obtained in an electrically shielded booth using a BioSemi ActiveTwo system (BioSemi, www.biosemi.com) with a sample rate of 8,192 Hz from 64 electrodes following the International 10–20 standard, plus an additional 4 electrode pairs around ears (earlobe, mastoid, posterior to the earlobe and inferior to the mastoids, and superior–anterior to the tragus near the zygomatic arch) for 72 total electrodes. Prior to recording, offsets from all electrodes were measured to be below 20 µV relative to the common mode sense electrode using the ActiveView2 software. Data were recorded using the Lab Streaming Layer (LSL) software (https://github.com/sccn/labstreaminglayer) in XDF format on a desktop computer running Windows 10. Events were recorded alongside EEG data as a separate trigger channel in the XDF file.

EEG preprocessing

The EEG data were preprocessed in MATLAB 2021b using EEGLAB (Delorme and Makeig, 2004) and a custom MATLAB code. Data were first imported from XDF files into EEGLAB using the XDF importer plugin (version 1.19, https://github.com/xdf-modules/xdf-EEGLAB) after which events were extracted from the trigger channel using a custom MATLAB code. Given the high number of densely packed event codes (>130,000/subject at an average rate of one event every 20 ms), events from the three different event streams (parallel port and two StimTraks) often overlapped in the trigger channel, causing occasional event code errors. We therefore checked all recorded events against expected events and corrected for any missing, extra, or jittered events using a custom MATLAB code. We then applied a second set of corrections to adjust for differences in the acoustic signal and the event times. These corrections accounted for the 1.973 ms audio delay due to the HRTF process, a 1 ms audio delay due to the audio travel time from the transducer via the insert earphone tubes, a 1 ms trigger delay in one of the StimTrak channels caused by the TriggerBox hardware, and a 0.541 ms trigger delay to the chirp triggers accounting for a difference between the expected chirp timing from the Cheech process and the chirp onsets.

Data were then pruned to retain only sections of the data from the rest periods and the attention task periods with 10 s buffers on both sides of the selected periods to prevent filtering artifacts. Each remaining section was then DC corrected and filtered using a noncausal second-order 0.1 Hz high-pass Butterworth filter using ERPLAB's filter function to remove slow drifts and then referenced to electrode Cz. Data were then checked and cleaned for 60 Hz line noise, which was occasionally found in some channels despite the electrically shielded booth. Only channels with detected line noise were cleaned (mean, 4.5 channels ± 7.7 SD; range for a given subject, 0–39 channels) to minimize the effects of the cleaning. This was achieved by first checking for line noise using a frequency tagging approach. Individual channel spectra were calculated across all frequencies. For each frequency, x, the average spectral power of four neighboring frequencies except those immediately adjacent (x − 2, x − 3, x + 2, x + 3) were subtracted. If the resulting neighbor corrected power at 60 Hz or any of its first four harmonics (120, 180, 240, 300 Hz) was greater than eight standard deviations calculated from the corrected power from all other frequencies, that channel was cleaned of line noise using the Cleanline plugin (https://www.nitrc.org/projects/cleanline). This repeated up to a maximum of five times or until no further line noise was detected.

We used independent component analysis (ICA) to further clean the data of eyeblink, eye movement, and heartbeat artifacts following the methodology outlined by Luck (2022). This involved making a copy of the data processed up to this point (ICA copy), downsampling, and cleaning the ICA copy, followed by ICA and marking and rejecting artifact components and then transferring the cleaned ICA weight matrix from the ICA copy to the original data. This approach allows for a massive reduction in ICA computation time while still retaining all the information in the data relevant for decomposing eye and heart artifacts. Specifically, after making a copy for ICA, we downsampled to 256 Hz to speed ICA processing and then manually inspected and rejected bad channels (mean, 3.67 ± 2.16 SD channels; range = 0–9 channels). Following channel rejection, a noncausal 1–50 Hz eighth-order Butterworth bandpass filter using ERPLAB's filter function was applied. We then removed sections of data with significant movement and muscle artifacts that can hinder ICA's ability to isolate eye and heart components, using ERPLAB's continuous artifact rejection function (mean, 126.54 ± 86.85 s SD or 3.52 ± 2.36%; range, 4.98–298.79 s or 0.14–7.96%). The rejection process operated with a 1 s sliding window with a 0.25 s step size and an initial rejection threshold of ±200 µV. In order to retain eye activity for ICA, the rejection threshold was applied against all channels except those mostly likely to contain eye-related activations (FP1, FPz, FP2, AF7, AF3, AFz, AF4, AF8). The rejection threshold was inspected and adjusted for each subject to maximize rejecting movement and muscle artifacts while retaining eyeblinks (mean, 220.83 ± 29.18 SD µV; range = 200–300 µV). We then used the Infomax ICA algorithm after first using principal component analysis to reduce the number of channels from an average of 67–32 for each subject to further speed ICA. ICA components constituting eyeblinks, eye movements, and heart artifacts (when present) were manually selected for rejection (mean, 2.79 ± 0.66 SD components; range, 2–4 components). The resulting ICA weight matrix was then transferred to the original dataset, after first rejecting the same bad channels in the original dataset as were identified in the ICA dataset.

After the data were cleaned with ICA, we downsampled it to 512 Hz and applied a noncausal 0.5–40 Hz eighth-order Butterworth bandpass filter using ERPLAB's filter function. Data were then re-referenced to the average of the left and right earlobes. In the case that one or both earlobe channels were rejected as bad, the average of the left and right mastoid electrodes was used as reference instead. While re-referencing, activity at channel Cz (the previous reference) was recalculated and placed back into the data. We then used spherical interpolation on any channels rejected.

EEG data were epoched −50 to +500 ms from the first chirp of each word in each condition, with the prestimulus period used as the baseline. Each condition had roughly the same number of events (MonoMale, 2,417; DualTargetMale, 2,392; DualMaskerMale, 2,392), except for one subject in which data recording ended early due to computer error (MonoMale, 1,249; DualTargetMale, 2,000; DualMaskerMale, 2,392). Epochs with artifacts were cleaned with a two-step process: first, a voltage threshold of ±300 µV was applied to remove any epochs with extreme artifacts. Second, any remaining epochs with voltage activity outside of 6 SD for a single channel or 2 SD across all channels were rejected (MonoMale, mean, 5.01 ± 0.83% SD rejected epochs; range, 3.07–6.85%; DualTargetMale, mean, 4.84 ± 0.86% SD rejected epochs; range, 3.75–7.02%; DualMaskerMale, mean, 5.12 ± 1.1% SD rejected epochs; range, 2.9–8.12%).

N1 extraction

N1 component activity was captured as the mean amplitude at electrode Fz between 110 and 160 ms postchirp onset. The electrode site Fz was chosen based on previous research using the Cheech paradigm (Backer et al., 2019) and literature on the scalp regions where auditory evoked responses generally reach peak amplitudes (Kappenman and Luck, 2011). Given the N1 component lacked a clear peak (as is typical during continuous speech presentation) and was usually followed by a sustained negativity (Figs. 3, 4), a mean amplitude over the 50 ms interval was used instead of more traditional peak fitting approaches, with the time course chosen based on visual inspection of mono and dual target conditions (Fig. 4). Difference scores were calculated by subtracting mean dual target N1 amplitudes from mean dual masker N1 amplitudes.