Testing the Role of Temporal Attention in Speech: Pretarget Alpha Predicts Memory Encoding Rather Than Effects of Linguistic Focus

Abstract

The rapid, continuous flow of speech often places strong demands on attention. Listeners are thought to meet these demands by using linguistic focus cues to predict when important information will occur and to allocate temporal attention accordingly. Yet, direct neural evidence for focus-driven modulations of preparatory attention is lacking. We address this gap by measuring EEG alpha power (8–13 Hz), an index of temporal attentional preparation (N = 40; 26 females, 12 males, 2 nonbinary). Using an auditory question–answer paradigm, we tested whether pretarget alpha power tracks the temporal prioritization of focused words in the discourse. We found that pretarget alpha was comparable before focused and defocused words, inconsistent with the view that attention is temporally preallocated to discourse-relevant information. Instead, trial-by-trial fluctuations in pretarget alpha modulated a centroparietal ERP subsequent memory effect for focused words, linking preparatory attention to the depth of processing and memory encoding during comprehension. These findings constrain theories of attention in language, indicating that alpha-band activity primarily indexes an encoding-ready state that predicts subsequent memory rather than focus-driven temporal attentional preallocation.

Significance Statement

As spoken language unfolds over time, listeners are hypothesized to prepare for informative content by preallocating attention in time, yet direct neural support for this view has been scarce. Using pretarget EEG alpha power as an index of preparatory attention, we examined whether listeners use linguistic focus cues to preallocate attention to the timing of focus-marked (novel) information in discourse. We observed no focus-related increase; instead, moment-to-moment fluctuations in preparatory alpha predicted the depth of processing and the encoding of novel information into memory, modulating subsequent neural and behavioral responses. Thus, preparatory alpha reflects an encoding-ready state that supports successful memory encoding rather than the dynamic preallocation of attention to linguistically focused information.

Introduction

Spoken language rapidly unfolds over time, conveying old and new information organized into the “information structure” of sentences. Efficient comprehension depends on identifying and integrating new information; listeners can use the discourse context to predict when that information will occur. For example, the question “Who brought the cake?” signals that the answer will likely occur early in the sentence (“Emma brought the cake”), whereas the question “What did Emma bring?” suggests the answer will be later (“…the cake”). Similarly, new information can be focused prosodically (“EMMA brought the cake”) or syntactically [“It was Emma who…”; see Vallduví (2016) for debates on terminology]. Listeners are thought to preallocate attention to key points in time when focused information is expected to occur (Cutler and Fodor, 1979; Wang et al., 2014), but evidence for the involvement of preparatory attention during comprehension remains modest.

Psycholinguistics studies show that cues to the timing of upcoming information sharpen processing, leading to faster phoneme detection (Cutler, 1976; Cutler and Fodor, 1979; Beier and Ferreira, 2022), enhanced ERP responses to semantic (N400) and syntactic (P600) violations (Wang et al., 2009, 2011, 2012), and boosted memory for focused words (Cutler and Fodor, 1979; Fraundorf et al., 2010, 2012). These effects are often attributed to anticipatory attention (Cutler and Fodor, 1979; Akker and Cutler, 2003; Wang et al., 2014). Evidence for the engagement of attention includes greater activation of domain-general attentional networks (Kristensen et al., 2013) and larger neural responses (P2 ERPs) to focused words (Chen et al., 2014; Yang et al., 2019). However, these studies investigated attention “in response to” focused words rather than attentional “preallocation.” Alternative accounts suggest that information structure modulates semantic preactivation of focused words (Ferreira and Lowder, 2016), which could explain prior ERP and memory effects (Hubbard et al., 2019; Hodapp and Rabovsky, 2021; Spalek et al., 2025) without invoking attentional mechanisms.

In the present study, we measure EEG alpha (8–13 Hz) to examine whether the comprehension benefits of discourse focus involve the preallocation of attention, paralleling alpha effects reported in visual and auditory temporal-orienting tasks (Breska and Deouell, 2017; Nobre and Van Ede, 2018). Alpha-band activity may function as a dynamic gate on speech processing, indexing attentional engagement and resource allocation: increases in alpha (synchronization) implement functional inhibition of task-irrelevant inputs, whereas decreases (desynchronization) release task-relevant regions for processing (Jensen and Mazaheri, 2010). Alpha tracks attention to speech during dichotic listening (Wöstmann et al., 2016), the suppression of speech distractors (Wöstmann et al., 2017), and attentional lapses during naturalistic listening (Boudewyn and Carter, 2018). Trial-wise alpha modulates depth of processing, predicting ERP indices of linguistic operations (e.g., resolving reference, detecting semantic conflict; Boudewyn et al., 2015, 2017) and behavioral performance (Dimitrijevic et al., 2017; Boudewyn and Carter, 2018).

Using a discourse question–answer paradigm (Cutler and Fodor, 1979; Beier and Ferreira, 2022), we cued focus to either an early or late target word in the upcoming sentence (Fig. 1). Based on prior work (Boudewyn and Carter, 2018), we expected reduced alpha at central electrodes prior to focused words, reflecting attention preallocation. Failure to observe prestimulus alpha modulations would argue against robust temporal preallocation of attention, as reflected in alpha power. Instead, facilitatory effects of focus may arise from nonattentional mechanisms, bottom–up attentional capture at the target word, or alternative attentional preparatory mechanisms not reflected in alpha power (e.g., phase-locking; Schwartze and Kotz, 2013; Beier et al., 2021).

Figure 1.

Example stimulus, showing the experimental design. The Early target word is focused when preceded by the Early focus question and defocused when preceded by the Late focus question and vice versa for the Late target word.

We also examined whether discourse focus and preparatory attention modulated the depth of processing and memory encoding of the target words, both behaviorally and neurally. To examine potential effects of focus on semantic processing, we measured the N400 to target words—which reflects processing demands even in the absence of violations (Van Petten and Kutas, 1990; Bentin et al., 1993). Finally, we addressed whether there were subsequent memory effects (SMEs; Klimesch et al., 1996; Mecklinger and Kamp, 2023), whereby trial-level ERPs or pretarget alpha predicted subsequent memory accuracy.

Materials and Methods

Transparency and code accessibility

The methods, analyses, and hypotheses of this study were preregistered (AsPredicted #133353) prior to the start of data collection. Code used for the statistical analyses, aggregated data, and a full list of stimuli are available on the Open Science Framework (https://osf.io/va6xj/?view_only=c6ebdea56ab34e08bb0962f69988e22d).

Participants

Forty-eight participants took part in the experiment after providing informed consent. Eight participants were removed from analysis following our preregistered exclusion criteria: failure to complete the experiment (N = 1); technical issues during data collection (N = 5); and >25% of trials rejected due to artifacts in any of the primary experimental conditions (N = 2). This resulted in our preregistered sample size of 40 participants (26 females, 12 males, 2 nonbinary; median age, 20 years; SD = 3.39). This sample size was selected based on recommendations in the field (Boudewyn et al., 2018). All participants were right-handed (Grey et al., 2017) native English speakers (defined as having learned English before age 5) who reported no hearing, speech, or language impairments and no neurological disorders. Participants were recruited through a University of California, Davis participant pool and were compensated with either class credit or cash ($15 per hour).

Stimuli and experimental design

Stimuli consisted of 80 question–answer pairs, which included the 36 stimuli used in Beier and Ferreira (2022) and another 44 stimuli modeled after those used by Cutler and Fodor (1979). Each stimulus consisted of an Early focus question (e.g., “Which man was wearing the hat?”) and a Late focus question (e.g., “What hat was the man wearing?”), cuing focus to either an Early or Late target word in a related upcoming sentence (e.g., “The man on the corner was wearing the dark hat”; target words underlined). Because the focus condition was determined by the preceding question, the target sentence was acoustically identical across both focus conditions.

As in Cutler and Fodor's (1979) phoneme monitoring experiment, all target words started with one of three possible phonemes (/b/, /d/, or /k/). Although the present study did not involve a phoneme monitoring task, we retained this design feature to permit future analyses related to phonological processing, which are beyond the scope of this report.

All stimuli were recorded by a male native English speaker from California and digitized at 48,000 Hz (later downsampled to 44,100 Hz) using a Shure KSM44 microphone. Stimuli were amplitude normalized at 65 dB using a custom Praat script (Boersma and Weenink, 2022). The onsets of each word in the sentences, including the Early and Late target words, were determined through the Montreal Forced Aligner (McAuliffe et al., 2017).

Filler items

Stimuli also included 100 filler items, which served four purposes. (1) Forty-eight filler items (26% of all stimuli) were followed by comprehension questions to ensure participants listened for meaning (see below, Procedure and apparatus). We used filler items rather than experimental items for these questions to avoid potential memory confounds: repetitions of the target words during the comprehension checks could have influenced later recall. (2) To prevent participants from noticing that the memory test focused only on the Early/Late target words in the experimental items (always subject or object modifiers), the memory probes for the filler items targeted other types of words (e.g., verbs, head nouns, prepositions). (3) Filler items provided some variability in sentence structure to prevent participants from adopting a shallow-processing strategy due to repeated exposure to similar constructions (Bock, 1986; Pickering and Ferreira, 2008), which can influence ERP responses (Ledoux et al., 2007; Tooley et al., 2009). (4) Finally, 60 filler items contained a manipulation of semantic constraint. Because our preregistered hypotheses included a potential effect of discourse focus on the N400, these filler items provided a check that our paradigm could reliably detect the well-established effect of semantic prediction on the N400.

The 60 filler items containing a manipulation of semantic constraint were adapted from a prior study on the effect of semantic constraint on the N400 (Brothers et al., 2019). There were two versions of each sentence: a High Constraint version leading up to a highly predictable target word (e.g., “The bride refused to invite her father to the wedding for some reason”; target word underlined) and a Low Constraint version where the same target word was less predictable yet plausible (“The lazy teenager refused to go to the wedding for some reason”). The original study (Brothers et al., 2019) ensured the effectiveness of this manipulation through a cloze norming task, where participants read each sentence frame (e.g., “The bride refused to invite her father to the…”) and were asked to write in the first word that comes to mind; cloze probability is the percentage of participants completing the sentence with the target word. This confirmed that the target word was more predictable in the High Constraint (95% cloze probability) than in the Low Constraint condition (<1%). To prevent participants from noticing a difference between these filler items and the experimental items and to maintain a similar trial structure, we preceded each sentence by a question (e.g., “Who refused to do something?”). The questions were designed not to focus the semantic target word (e.g., “wedding”) to avoid confounding the effects of discourse focus and semantic constraint. Additionally, the questions were constructed to avoid increasing the predictability of the subsequent sentence.

The remaining 40 filler items were adapted from Beier and Ferreira (2022). Importantly, the proportion of Early and Late focus questions across all filler and practice items was 50:50. During the experiment, a comprehension question followed all of these 40 filler items and 8 of the semantic filler items, for a total of 48 comprehension questions.

An additional 18 practice items, with characteristics similar to the experimental items and fillers, were used in an initial practice block. The filler and practice items were recorded using the same procedure and the same native English speaker as for the experimental items.

Procedure and apparatus

After providing informed consent, participants completed a demographic survey. Following placement of the EEG electrodes (see below, EEG recording), participants sat in an electrically shielded, sound-attenuated booth in front of a VIEWPixx/EEG LED monitor (model VPX−2006A; VPixx Technologies; 120 Hz refresh rate). Auditory stimuli were presented through earphones via air-tubes (ER-1, Etymotic Research) in order to minimize electrical interference with the EEG. After receiving written and verbal instructions about the task, participants completed a practice block.

Stimuli were presented using the Presentation software (Version 23.0, Neurobehavioral Systems). Participants started each trial by pressing the spacebar on a computer keyboard. After 1,000 ms, they heard one of the focus questions (Early or Late). The ISI between the end of the question and the onset of the sentence was jittered between 1,180 and 2,200 ms to prevent participants from forming a precise temporal prediction except that engendered by the Early/Late focus manipulation. Each trial ended 1,000 ms after the end of the sentence, after which participants could proceed to the next trial by pressing the spacebar. Participants were instructed to fixate on a central cross on the video monitor throughout each trial to reduce eye movement artifacts in the EEG. They were encouraged to use the self-paced breaks between trials to blink and rest their eyes.

On 48 filler trials (see above, Filler items), the end of the sentence was followed by a question mark presented at the center of the screen. These trials served as the comprehension questions: participants were instructed to repeat aloud the answer to the question they just heard. For example, after hearing the question “Which rabbit was chewing the carrot?” followed by the sentence “The white rabbit was chewing the delicious carrot,” participants should respond with: “the white rabbit.” Participants had 5 s to answer the question, and their responses were recorded. This task both encouraged participants to use the Early/Late focus questions as cues for when to pay particular attention during the sentence and ensured that they remained attentive throughout the experiment.

Stimuli were presented in 10 blocks of 18 trials each; each block lasted ∼5 min. Blocks ended with a memory test similar to that used by Cutler and Fodor (1979) and Beier and Ferreira (2022). We chose to test memory after each block rather than at the end of the entire experiment (as in those previous studies) due to the increased number of stimuli and duration of the study. During the memory test, each sentence was presented visually on the video monitor with one target word blocked out. Participants were instructed to select the correct word completing the sentence among four alternatives shown on the monitor. For the experimental items, half of the judgments were about the Early target words and half were about the Late target words; furthermore, a random half were for words that had been focused and half for those defocused by the preceding questions. For the filler items, memory was tested for other parts of the sentence (e.g., verbs, head nouns, prepositions) to prevent participants from developing a strategy based on the typical position of the probed words.

The experimental conditions (Early/Late focus question, Early/Late target tested in the memory test) were counterbalanced across four lists. The lists were also counterbalanced with respect to High Constraint and Low Constraint conditions for the 60 semantic constraint filler items. Stimulus order was pseudorandomized within each block so that no more than three stimuli in the same focus condition and no more than two stimuli of the same type (experimental or filler) were presented consecutively. In addition, each block always started with two filler items and ended with a filler item, so that the experimental items were never at the very beginning or end of the block. Participants were encouraged to take longer breaks between blocks to reduce fatigue and maintain focus.

EEG recording

EEG was recorded using a 64-channel actiCAP active electrode system (Brain Products) at the following recording sites of the international 10–10 system (Jurcak et al., 2007): FP1, FP2, AF7, AF3, AFz, AF4, AF8, F7, F5, F3, F1, Fz, F2, F4, F6, F8, FT7, FC5, FC3, FC1, FCz, FC2, FC4, FC6, FT8, T7, C5, C3, C1, Cz, C2, C4, C6, T8, TP9, TP7, CP5, CP3, CP1, CPz, CP2, CP4, CP6, TP8, TP10, P7, P5, P3, P1, Pz, P2, P4, P6, P8, PO7, PO3, POz, PO4, PO8, PO9, O1, Oz, Iz, O2, and PO10. Additionally, bipolar EOG channels were used to capture blinks and horizontal eye movements through passive tin electrodes. (Due to a technical issue, EOG data was not recorded for 13 out of 40 participants. No significant difference was found between participants with and without EOG data in any of the behavioral or EEG measures of interest.) Data were amplified with a Neuroscan SynAmps2 input board and amplifier and digitized at 1,000 Hz using the CURRY 8 acquisition software (Compumedics). Electrode Fpz served as the ground and electrode FCz as the reference during recording. The fitted elastic electrode caps were centered for each participant so that electrode Cz was at the half-way point between nasion and inion and between preauricular points. Electrode impedances were kept below 15 kΩ through the application of high viscosity electrolyte gel at each electrode site.

EEG preprocessing

Data were preprocessed in MATLAB (MathWorks) using the EEGLAB Toolbox (Delorme and Makeig, 2004) with the ERPLAB plugin (Lopez-Calderon and Luck, 2014). Preprocessing scripts were adapted from the standardized preprocessing pipeline recommended by Boudewyn et al. (2023) and available on the Open Science Framework (https://osf.io/wdrj3/), after removing noisy or broken channels through visual inspection (to be later interpolated).

Independent component analysis (ICA) was used on the continuous data to remove the contribution of blink artifacts to the EEG. Data were filtered between 0.05 and 30 Hz using the EEGLAB function pop_basicfilter. The ERPLAB function pop_erplabInterpolateElectrodes was then used for spherical interpolation of noisy or broken channels. Lastly, EEG data were rereferenced to the algebraic average of electrodes TP9 and TP10 (placed over the left and right mastoids, respectively).

For ERP analyses, EEG data were epoched from −200 to 1,000 ms relative to the onset of the target words, and baseline correction was applied using the 200 ms prior to word onset. For time–frequency analyses, data were epoched from −3,000 to 3,000 ms relative to the onset of the target words and other events of interest (e.g., question and sentence onsets). For these epochs, we did not apply baseline correction to the EEG voltage at this point, given that we were interested in oscillatory activity before target word onsets (see below, Time–frequency analyses, for details on the baseline correction later applied to spectral power). The longer epochs were used to measure oscillatory activity over the periods before and after the target words and to avoid edge artifacts, which occur when wavelet convolution extends beyond the available data at the edges of each epoch, potentially distorting power estimates.

Finally, artifact rejection was performed separately for the shorter and longer epochs using the peak-to-peak algorithm through the ERPLAB function pop_artmwppth on data from all ICA-corrected electrodes (excluding the EOG bipolar channels), starting with an amplitude threshold of 100 µV which was iteratively adjusted for each participant through visual inspection (resulting in a range from 60 to 160 µV).

Behavioral analyses

Accuracy on the memory test was analyzed through a logistic mixed-effects model with fixed effects of Target Position (Early, Late), Focus Position (Early, Late), and their interaction. The model included by-subject and by-item intercepts and slopes for Target Position and Focus Position.

We also performed an exploratory analysis on the time it took participants to make their selection on the memory test. This measure included the time it took participants to read the sentence (with the target word blocked out) and to view the four possible presented options; note that they were not instructed to respond as fast as possible. Response times faster than 2 s and slower than 30 s were excluded from the analysis. Given that the distribution was right-skewed, response times were log-transformed to achieve a normal distribution (Baayen and Milin, 2010). Response times were then analyzed through a linear mixed-effects model with the same fixed and random effects as the model on accuracy. As the full model did not converge, we removed the random slope of Focus Position. For all models, in case of convergence issues or singular fit, the random effects structure was simplified by iteratively removing the random effect explaining the least amount of variance until the model converged (Barr et al., 2013).

ERP analyses

Following our preregistered hypotheses, we initially extracted mean amplitude over a window of 300–500 ms following target word onset, matching the typical latency of the N400 (Šoškić et al., 2022). To explore potential SMEs, we also extracted mean amplitudes over a later time window of 600–1,000 ms. Mean amplitude was calculated on a trial-by-trial basis for each participant and analyzed through mixed-effects models using the lme4 and lmerTest packages in R (Bates et al., 2014; Kuznetsova et al., 2017). We used mixed-effects models rather than ANOVAs to account for by-subject and by-item variability and because this analytical method has been shown to be more powerful (Baayen et al., 2008; Stevens and Brysbaert, 2016) and more robust to noise in ERP data (Heise et al., 2022). To account for differences in the scalp distribution of the effects, we included a fixed effect of electrode cluster (central, FC1, FC2, C1, CP1, CPz, C2, Cz, CP2; left frontal, F7, F3, FC5, FT7, FC3, Fp1, F5, AF7; left posterior, PO7, PO9, P5, P3, O1, P7, PO3; left temporal, TP7, CP3, C3, C5, T7, CP5; midfrontal, Fz, AF4, AFz, F2, AF3, F1; midposterior, P1, Pz, Oz, POz, P2, Iz; right frontal, F6, F4, FT8, FC4, AF8, F8, FC6, Fp2; right posterior, PO4, P4, PO8, P6, O2, PO10, P8; right temporal, CP4, T8, C6, C4, CP6, TP8). The effect of cluster was treatment-coded, such that model estimates reflected the comparison of each cluster against the central cluster.

The model on the semantic constraint filler items included a fixed effect of semantic constraint (High, Low; treatment-coded) and its interaction with cluster, as well as by-subject and by-item random intercepts and slopes for semantic constraint.

The model on the experimental items included a fixed effect of focus (Focused, Defocused; treatment-coded) and its interaction with cluster, as well as by-subject and by-item random intercepts and slopes for focus. This analysis was performed separately for the Early and Late target words to alleviate model complexity given the large number of interaction terms involving electrode cluster and because we did not have a priori hypotheses for whether the effects would be similar or different across the two targets.

Time–frequency analyses

Event-related spectral power was estimated using the EEGLAB function newtimef, time-locked to the onset of the Early and Late target words. Power was estimated for a range of linearly spaced frequencies from 3 to 30 Hz using Morlet wavelets starting at 3 cycles and increasing by 0.5 cycles for every 1 Hz increase in frequency. Given that any differences induced by the focus condition could in principle begin during the focus question, we used the 1,000 ms prior to the question onset as the baseline window, averaging across focus conditions. Baseline correction during the trial was applied by first dividing each time–frequency power estimate by the baseline power spectrum of the matching frequency and then converting to decibel using the base-10 log. Alpha power, calculated as the mean power from 8 to 13 Hz, was compared between the Focused and Defocused conditions in the 1,500 ms prior to and following each target word, using cluster-based permutation statistics (Maris and Oostenveld, 2007). For this, we calculated the sum of t values for each cluster of significant adjacent time points and compared this t-sum statistic to a randomization distribution of the same t-sum statistics in shuffled data between conditions (10,000 permutations). Our initial analysis was conducted on electrode Cz, where we expected to find lower alpha power corresponding to greater preparatory attention based on Boudewyn et al. (2015, 2017) and Boudewyn and Carter (2018). Follow-up analyses were performed on a range of central and occipital midline electrodes (CPz, Pz, POz, Oz). We constrained this analysis to midline electrodes given that we did not have any a priori hypotheses about lateralized effects.

Trial-level analyses

We performed follow-up analyses on trial-level pretarget alpha at all electrodes, estimated using the BOSC method (Whitten et al., 2011) through the fBOSC toolbox in MATLAB (Seymour et al., 2022). Similarly to Boudewyn and Carter (2018), we employed this approach to address an important limitation: given the variable duration and number of words across stimuli, there is likely to be jitter in the exact onset of any differences in alpha power across conditions. Thus, averaging alpha power across trials may obscure an effect that has inconsistent timing across stimuli. The BOSC method overcomes this issue by calculating the percentage of timepoints over a window (1,500 ms prior to target word onset) that contain periodic activity within the alpha band above and beyond background 1/f aperiodic activity. In the fBOSC toolbox, 1/f aperiodic activity is calculated using the FOOOF algorithm (Donoghue et al., 2020). We then identified episodes of periodic activity exceeding a power threshold (above the 95th percentile of the theoretical probability distribution) and duration threshold (three cycles per frequency), for each frequency, averaging over the alpha band (8–13 Hz). The resulting p-episode metric therefore represents the percentage of timepoints over the selected window that contains alpha activity above these thresholds, regardless of its exact timing across each stimulus.

We first tested whether trial-level pretarget alpha varied as a function of focus condition. To deal with the non-normal, zero-and-one-inflated distribution of the p-episode metric, we used a Bayesian ordered-beta regression through the ordbetareg package in R (Kubinec, 2023). Similarly to the ERP analyses, the model was conducted separately for the Early and Late target and included fixed effects of focus (Focused, Defocused; treatment-coded), cluster, and their interaction. It also included by-subject and by-item random intercepts and slopes for focus. As in the ERP analyses, we included p-episode data from all electrodes in the model, controlling for differences in scalp distribution by including cluster as a fixed effect.

We also tested whether ERP mean amplitude varied as a function of pretarget alpha on a trial-by-trial basis, aggregating across the Early and Late targets. The model included fixed effects of focus (Focused, Defocused) and pretarget alpha, as well as by-subject and by-item random intercepts and slopes for focus. For this model, the effect of focus was sum-coded, in order to estimate the main effect of pretarget alpha across focus conditions.

Lastly, we tested whether memory test accuracy was predicted by pretarget alpha or the ERP mean amplitude over the 600–1,000 ms post-target window, on a trial-by-trial basis. Data were aggregated across the Early and Late target words. We used a logistic mixed-effects model with fixed effects of focus (Focused, Defocused), pretarget alpha, and ERP mean amplitude. The effect of focus was sum-coded, while pretarget alpha and ERP mean amplitude were z-score transformed to help with model convergence. The model also included by-subject and by-item random intercepts.

Results

Focused words are remembered better

On average, participants performed with 77% accuracy (SD, 42%) on the memory test (similarly to prior studies; Cutler and Fodor, 1979). A logistic mixed-effects model found that accuracy was higher for the Early than the Late target word (β = −0.81; SE = 0.21; p < 0.0001), and higher for the Early than the Late focus condition (β = −0.86; SE = 0.15; p < 0.0001). Critically, there was a significant interaction between Target Position and Focus Position, indicating that memory for each target word was higher when it had been focused by the preceding question (β = 1.61; SE = 0.19; p < 0.0001; Fig. 2), replicating Cutler and Fodor (1979).

Figure 2.

Memory test accuracy (left) and response time (right) for Early versus Late target words as a function of Early (red) versus Late (blue) focus. Higher accuracy and faster response times were found for the Early target when focused by the Early focus question. For the Late target, higher accuracy and faster response times were obtained when focused by the Late focus question. Data points represent mean accuracies and response times for each participant. Error bars represent the standard error of the mean.

The exploratory analysis on response times in the memory test revealed that participants were faster at selecting an answer for the Early than the Late target word (β = 0.1; SE = 0.02; p < 0.0001) and faster for the Early than the Late focus condition (β = 0.09; SE = 0.02; p < 0.0001). Like for accuracy, there was a significant interaction, indicating that participants were faster for words that had been focused by the preceding question (β = −0.17; SE = 0.3; p < 0.0001).

Thus, participants were both more accurate and faster at selecting a response to the memory test for words that had been focused during the trials. We interpret the faster selection as reflecting a greater amount of confidence in their response.

Similar pretarget alpha across focus conditions

Event-related spectral power in the alpha band as a function of Early versus Late target word and focus is shown in Figure 3. In contrast to our hypothesis that attention would be preallocated as a function of focus, as indexed by alpha power at electrode site Cz (reflecting attention allocation to auditory inputs; Boudewyn et al., 2015, 2017; Boudewyn and Carter, 2018), no such effect was observed: cluster-based permutation analysis revealed no reduction in pretarget alpha power at electrode Cz for either target as a function of focus. There was also no difference observed in alpha power following the Early target. In contrast, following the Late target, we found a significant cluster at Cz from 960 to 1,303 ms, resulting from lower alpha power in the Focused compared with the Defocused condition.

Figure 3.

Power averaged over the alpha band (8–13 Hz), locked to the onset of the Early target word (A) and Late target word (B), for electrodes Cz, POz, and Oz. Shaded areas represent standard error of the mean. Horizontal black lines represent significant clusters. Significant differences in alpha as a function of focus condition only emerged following the Late target, rather than in preparation for either target word.

Visualization of changes in spectral power relative to the 1,000 ms baseline period prior to the question onset (Fig. 4) shows that there was a marked increase in alpha power over parieto-occipital electrodes with the onset of the question (the first auditory stimulus of the trial), which then returned to baseline at the end of the sentence (the last auditory stimulus of the trial, on average ∼1,140 ms after the Late target word onset). This fits with prior studies showing an increase in parieto-occipital alpha in preparation for an auditory stimulus (Fu et al., 2001), potentially reflecting a gating-by-inhibition mechanism whereby irrelevant visual information is suppressed (through increased alpha power) in favor of auditory inputs (Jensen and Mazaheri, 2010).

Figure 4.

A, Event-related spectral power at electrode POz throughout the trial, showing an increase in alpha-band power (8–13 Hz) with the onset of the question and a return to baseline following the Late target word. B, Scalp distribution of alpha power throughout the trial, indicating that alpha was greater at parieto-occipital recording sites.

We performed follow-up analyses at midline parietal and occipital electrodes matching the observed scalp distribution of alpha power (CPz, Pz, POz, Oz), which still revealed no significant clusters prior to target onset. However, there was lower post-target alpha following the Focused Late target word, resulting in a significant cluster between 446 and 1,500 ms at electrode POz and from 397 to 1,500 ms at Oz (Fig. 3B). This effect may reflect a faster return to baseline alpha power (releasing the occipital inhibition) after hearing the Focused Late target, in preparation for the behavioral task of answering the question out loud (which was indicated visually through a question mark on the screen). We return to this interpretation in Discussion.

Given that any potential pretarget effects of focus on alpha may have been obscured by the jitter naturally present in our stimuli due to sentence-level differences in duration and number of words, we also performed a follow-up analysis of pretarget alpha p-episode to address this issue, following Boudewyn and Carter (2018). The p-episode metric, estimated using the BOSC method (Whitten et al., 2011; Seymour et al., 2022), represents the percentage of timepoints that contain above-threshold alpha within a time window of interest (here, 1,500 ms prior to target onset; see Materials and Methods for more details). Thus, this measure captures changes in alpha activity regardless of the exact timing of its dynamics. We analyzed trial-level p-episode at all electrodes using a Bayesian ordered-beta regression to account for its non-normal distribution.

We found several main effects of electrode cluster, such that p-episode was greater at posterior electrode sites and lower at frontal sites, relative to central sites, matching the observed scalp distribution of alpha power across conditions (see Table 1 for model output; for Bayesian models, we considered an effect to be “significant” when the 95% credible interval (CI) for the effect did not overlap with 0; Nicenboim and Vasishth, 2016). However, it did not reveal an effect of focus for either the Early target (95% CI = [−0.08, 0.06]) or the Late target (95% CI = [−0.01, 0.14]). To test the strength of the evidence for the null hypothesis, we calculated Bayes factors for all effects using the bayestestR package in R (Makowski et al., 2019). The Bayes factor for the effect of focus was 0.007 for the Early target and 0.026 for the Late target, indicating strong evidence in favor of the null hypothesis (Quintana and Williams, 2018). Thus, based on both this analysis and the cluster-based permutations on alpha power, we conclude that focus condition did not have a significant effect on pretarget alpha in our study. Nonetheless, we tested whether pretarget alpha affected the ERP and behavioral responses on a trial-by-trial basis (described in the next sections), under the assumption that—contrary to Boudewyn and Carter (2018)—“higher” pretarget alpha reflected “greater” auditory attention (through greater inhibition of visual information).

Table 1.

Bayesian ordered-beta mixed–effects model output. Alpha p-episode at −1,500 to 0 ms for the Early and Late targets, as a function of focus and electrode cluster

Predictor	Early target			Late target
	Estimate	Est. error	95% CI	Estimate	Est. error	95% CI
Intercept	0.22	0.06	[0.11 0.33]*	0.17	0.06	[0.05 0.30]*
Focus (Focused)	−0.01	0.03	[−0.08 0.06]	0.06	0.04	[−0.01 0.14]
Cluster (left frontal)	−0.22	0.01	[−0.25 −0.20]*	−0.25	0.01	[−0.28 −0.23]*
Cluster (left posterior)	0.26	0.01	[0.24 0.29]*	0.31	0.01	[0.28 0.33]*
Cluster (left temporal)	0.05	0.01	[0.03 0.08]*	0.05	0.01	[0.02 0.08]*
Cluster (midfrontal)	−0.29	0.01	[−0.31 −0.26]*	−0.30	0.01	[−0.33 −0.27]*
Cluster (midposterior)	0.26	0.01	[0.23 0.29]*	0.33	0.01	[0.30 0.36]*
Cluster (right frontal)	−0.28	0.01	[−0.30 −0.25]*	−0.28	0.01	[−0.30 −0.25]*
Cluster (right posterior)	0.26	0.01	[0.23 0.29]*	0.31	0.01	[0.28 0.33]*
Cluster (right temporal)	0.01	0.01	[−0.01 0.04]	0.02	0.01	[−0.01 0.05]
Focus × left frontal	0.05	0.02	[0.01 0.08]*	−0.00	0.02	[−0.04 0.03]
Focus × left posterior	0.03	0.02	[−0.01 0.06]	−0.02	0.02	[−0.06 0.01]
Focus × left temporal	0.01	0.02	[−0.02 0.05]	−0.02	0.02	[−0.06 0.01]
Focus × midfrontal	0.04	0.02	[0.00 0.08]*	−0.00	0.02	[−0.04 0.04]
Focus × mid posterior	0.03	0.02	[−0.01 0.07]	−0.01	0.02	[−0.05 0.03]
Focus × right frontal	0.05	0.02	[0.02 0.09]*	−0.02	0.02	[−0.06 0.01]
Focus × right posterior	0.03	0.02	[−0.01 0.06]	0.00	0.02	[−0.03 0.04]
Focus × right temporal	0.04	0.02	[0.00 0.08]*	−0.02	0.02	[−0.06 0.02]

Note.*95% CI do not overlap with 0. Fixed effects were treatment-coded, using the Defocused condition and the Central cluster as reference levels.

In summary, we did not find evidence that discourse focus modulated pretarget alpha power, either through cluster-based permutation testing or p-episode analysis, and Bayesian statistics provided strong support for the null hypothesis. While a post-target reduction in alpha power was observed following Focused Late targets over posterior electrodes, this effect may reflect preparatory processes for the upcoming behavioral response rather than anticipatory attention to focused words. Thus, contrary to our predictions, focus did not significantly recruit preparatory attention, as reflected by alpha activity.

Semantic constraint elicits an N400 effect

As expected, central electrodes showed the N400 effect, that is a stronger negative deflection for Low Constraint as compared with High Constraint filler items in the time period from 300 to 500 ms after word onset (β = −2.26; SE = 0.4; p < 0.0001; Fig. 5). Significant interactions with electrode cluster confirmed that the difference between High and Low Constraint was significantly greater at central sites compared with other electrode clusters (see Table 2 for full model output). This analysis establishes that our design had sufficient power to detect the expected effect of semantic constraint on the N400.

Figure 5.

ERPs (top) at electrode Cz, time-locked to the onset of the target word for the semantic constraint filler items, as a function semantic constraint (High vs Low). The negative-going deflection in the ERPs (red) is the N400 elicited by the Low semantic constraint items. Positive voltage is plotted upward for all ERP plots. Topographic voltage map (bottom) showing the scalp distribution of the difference between the ERPs (Low minus High Constraint) averaged over the 300–500 ms post-target period (i.e., the N400 time window). The N400 effect is visible as the negative (blue) voltage over the central–parietal scalp.

Table 2.

Mixed-effects model output. Mean ERP amplitude at 300–500 ms for the filler items, as a function of semantic constraint and electrode cluster

Predictor	β	SE	p
(Intercept)	0.07	0.29	0.805
Semantic Constraint (Low Constraint)	−2.26	0.40	<0.001***
Cluster (left frontal)	−0.86	0.12	<0.001***
Cluster (left posterior)	0.48	0.13	<0.001***
Cluster (left temporal)	−0.14	0.13	0.298
Cluster (midfrontal)	−0.59	0.13	<0.001***
Cluster (midposterior)	0.47	0.13	<0.001***
Cluster (right frontal)	−0.20	0.12	0.114
Cluster (right posterior)	0.56	0.13	<0.001***
Cluster (right temporal)	0.14	0.13	0.302
Low Constraint × left frontal	1.25	0.18	<0.001***
Low Constraint × left posterior	0.92	0.18	<0.001***
Low Constraint × left temporal	1.00	0.19	<0.001***
Low Constraint × midfrontal	0.99	0.19	<0.001***
Low Constraint × midposterior	0.43	0.19	0.024*
Low Constraint × right frontal	1.05	0.18	<0.001***
Low Constraint × right posterior	0.78	0.18	<0.001***
Low Constraint × right temporal	1.05	0.19	<0.001***

Note. Significance codes: ***p < 0.001; **p < 0.01; *p < 0.05; p < 0.10. Fixed effects were treatment-coded, using the High Constraint condition and the Central cluster as reference levels. SE, standard error.

Focused words elicit a centroparietal late positivity

Figure 6 shows the ERPs as a function of focus. There was no difference in the amplitude of the N400 (300–500 ms latency range) as a function of focus for either Early or Late targets (both p > 0.05; see Table 3 for full model outputs). We did, however, observe a late positivity for Focused as compared with Defocused target words for both Early and Late target words. This effect of focus at 600–1,000 ms was significant for the Late target (β = 1.08; SE = 0.4; p < 0.001) and marginally significant for the Early target (β = 0.69; SE = 0.39; p = 0.08). The effect was maximal at central and posterior scalp sites, as reflected through interactions of focus and electrode cluster (see Table 4 for full model output).

Figure 6.

ERPs (top) at electrode Pz, time-locked to the onsets of the Early and Late target words as a function of focus condition. There was no difference in the N400 (300–500 ms), but there was a later (600–1,000 ms) positive-going ERP effect for the Focus condition. The topographic maps (bottom) show the scalp distribution of the ERP differences of Focused minus Defocused conditions over the later 600–1,000 ms post-target time period for Early and Late targets. The effect of focus is maximal over parietal scalp sites (orange-red colors).

Table 3.

Mixed-effects model output. Mean ERP amplitude at 300–500 ms for the Early and Late targets, as a function of focus and electrode cluster

Predictor	Early target			Late target
	β	SE	p	β	SE	p
(Intercept)	−3.23	0.35	<0.001***	−3.15	0.37	<0.001***
Focus (focused)	−0.36	0.32	0.258	−0.31	0.32	0.336
Cluster (left frontal)	1.03	0.29	<0.001***	2.32	0.33	<0.001***
Cluster (left posterior)	2.18	0.30	<0.001***	1.09	0.34	0.002**
Cluster (left temporal)	1.45	0.31	<0.001***	1.37	0.35	<0.001***
Cluster (midfrontal)	0.98	0.31	0.003**	1.77	0.35	<0.001***
Cluster (midposterior)	1.42	0.31	<0.001***	0.30	0.35	0.396
Cluster (right frontal)	1.20	0.29	<0.001***	2.02	0.33	<0.001***
Cluster (right posterior)	2.00	0.30	<0.001***	0.86	0.34	0.014*
Cluster (right temporal)	1.22	0.31	<0.001***	1.02	0.35	0.006**
Focus × left frontal	−0.16	0.15	0.291	−0.96	0.16	<0.001***
Focus × left posterior	0.44	0.16	0.006**	0.90	0.16	<0.001***
Focus × left temporal	0.25	0.16	0.123	0.07	0.17	0.688
Focus × mid frontal	−0.26	0.16	0.118	−0.71	0.17	<0.001***
Focus × midposterior	0.14	0.16	0.379	0.83	0.17	<0.001***
Focus × right frontal	−0.01	0.15	0.965	−0.28	0.16	0.070
Focus × right posterior	0.09	0.16	0.563	0.83	0.16	<0.001***
Focus × right temporal	0.03	0.16	0.853	0.29	0.17	0.082

Note. Significance codes: ***p < 0.001; **p < 0.01; *p < 0.05; p < 0.10. Fixed effects were treatment-coded, using the Defocused condition and the Central cluster as reference levels. SE, standard error.

Table 4.

Mixed-effects model output. Mean ERP amplitude at 600–1,000 ms for the Early and Late targets, as a function of focus and electrode cluster

Predictor	Early target			Late target
	β	SE	p	β	SE	p
(Intercept)	−2.65	0.34	<0.001***	−1.35	0.33	<0.001***
Focus (Focused)	0.70	0.39	0.079	1.08	0.40	0.009**
Cluster (left frontal)	0.78	0.12	<0.001***	0.35	0.13	0.006**
Cluster (left posterior)	2.52	0.13	<0.001***	1.46	0.13	<0.001***
Cluster (left temporal)	1.64	0.14	<0.001***	0.84	0.14	<0.001***
Cluster (midfrontal)	0.62	0.16	<0.001***	0.43	0.16	0.007**
Cluster (midposterior)	1.63	0.14	<0.001***	1.07	0.14	<0.001***
Cluster (right frontal)	1.04	0.12	<0.001***	0.02	0.13	0.878
Cluster (right posterior)	2.15	0.13	<0.001***	1.25	0.13	<0.001***
Cluster (right temporal)	1.37	0.14	<0.001***	0.45	0.14	0.001**
Focus × left frontal	−1.10	0.18	<0.001***	−2.13	0.18	<0.001***
Focus × left posterior	−0.32	0.19	0.083	0.26	0.19	0.176
Focus × left temporal	−0.55	0.20	0.005**	−0.62	0.20	0.002**
Focus × midfrontal	−0.77	0.22	<0.001***	−1.40	0.22	<0.001***
Focus × midposterior	−0.02	0.20	0.907	0.49	0.20	0.013*
Focus × right frontal	−0.80	0.18	<0.001***	−0.76	0.18	<0.001***
Focus × right posterior	−0.18	0.19	0.348	0.32	0.19	0.094
Focus × right temporal	−0.31	0.20	0.112	0.17	0.20	0.384

Note. Significance codes: ***p < 0.001; **p < 0.01; *p < 0.05; p < 0.10. Fixed effects were treatment-coded, using the Defocused condition and the Central cluster as reference levels. SE, standard error.

In sum, we did not see an effect of focus on the N400; rather, we found that focused words elicited a centroparietally distributed late positivity (600–1,000 ms) relative to defocused words. Late positive ERPs, particularly those with a centroparietal distribution following the N400, have been associated with stimulus evaluation, semantic integration, and depth of encoding (Van Petten and Luka, 2012; DeLong et al., 2014; Quante et al., 2018). Such ERPs have also been linked to strategic processing and enhanced memory for task-relevant information (Mecklinger and Kamp, 2023). Given that focused words elicited a robust late positivity in our study, we sought to determine whether this effect reflected deeper encoding processes—indexed by improved memory performance—and whether it was modulated by preparatory brain states, as indexed by pretarget alpha power.

Trial-level pretarget alpha and late positivity jointly modulate subsequent memory

To understand how trial-level variability in attention affected neural and behavioral responses, we first addressed whether pretarget alpha (p-episode during the 1,500 ms prior to each target) predicted the amplitude of the late positivity (averaged across 600–1,000 ms post-target). Data were aggregated across the Early and Late target words. Greater pretarget alpha corresponded to a more positive late positivity (β = 1.35; SE = 0.06; p < 0.0001; Fig. 7A), while the effect of focus was no longer significant (p > 0.05). However, pretarget alpha interacted with focus, such that the effect of pretarget alpha on mean amplitude was larger for Focused words (β = −0.87; SE = 0.06; p < 0.0001), suggesting that trial-level variability in preparatory attention mediated the effect of focus on the late positivity.

Figure 7.

A, Mean amplitude of the late positivity ERP (averaged over 600–1,000 ms post-target) as a function of pretarget alpha (p-episode) and focus condition, across all electrodes. Regression lines represent the relationship between mean amplitude and pretarget alpha, calculated from trial-level data; shaded ribbons represent confidence intervals. Individual data points and side density plots represent mean data for each participant, averaged across trials. B, Memory test accuracy as a function of the mean amplitude of the late positivity ERP (averaged over 600–1,000 ms post-target), pretarget alpha (p-episode), and focus condition. The y-axis is truncated at 0.5 for easier visualization. For visualization purposes only, pretarget alpha was turned to a categorical variable (high, medium, low; see legend), by splitting trials into three quantile-based groups—though analyses were conducted using a continuous measure of p-episode. Regression lines represent the relationship between mean amplitude and accuracy, calculated from trial-level data; shaded ribbons represent confidence intervals. Individual data points and side density plots represent mean data for each participant, averaged across trials. Greater accuracy was found for Focused words and as a function of mean amplitude of the late positivity. The SME of the late positivity on accuracy was stronger in the Defocused condition and for trials with lower pretarget alpha (less attentional preparation).

We then tested whether subsequent memory test accuracy was modulated by the late positivity and pretarget alpha during stimulus presentation, aggregating over the Early and Late targets. As expected, accuracy was higher for words that were Focused (β = −0.35; SE = 0.007; p < 0.0001), similarly to the prior behavioral analyses (where this effect was shown as a significant interaction between focus position and target position). Accuracy was also higher for words that elicited a more positive late positivity (mean amplitude over 600–1,000 ms; β = 0.11; SE = 0.006; p < 0.0001; Fig. 7B). The latency and scalp distribution of this ERP component match prior reports of centroparietal positive SMEs (Paller et al., 1987; Friedman and Johnson Jr., 2000; Schott et al., 2002; Mecklinger and Kamp, 2023). The effect of mean amplitude on accuracy also interacted with focus, such that there was a larger effect for words that were Defocused (β = 0.06; SE = 0.006; p < 0.0001). This might be partly due to a ceiling effect in the Focused condition, but it also matches prior findings of larger SMEs under shallow-processing conditions (Schott et al., 2002).

Furthermore, we found that pretarget alpha significantly interacted with the effect of the late positivity on accuracy (β = −0.06; SE = 0.006; p < 0.0001), such that there was a larger effect of the late positivity on accuracy for trials with lower pretarget alpha (Fig. 7B). This further supports the idea that SMEs are larger when the encoding conditions are suboptimal.

Lastly, pretarget alpha did not directly predict accuracy, nor was there a significant three-way interaction between focus, mean amplitude, and pretarget alpha (all p > 0.05). We checked for multicollinearity using the performance package in R (Lüdecke et al., 2021). We found that all effects had a variance inflation factor below 5, indicating low multicollinearity, despite the prior analysis indicating a relationship between pretarget alpha and the late positivity.

Overall, these analyses revealed that trial-level variability in attentional preparation affected both neural and behavioral measures of processing. Greater attentional preparation modulated the late positivity elicited by target words, mediating the effect of focus. In turn, the amplitude of the late positivity predicted subsequent accuracy on the memory test, and this effect was largest for words receiving shallow processing (when defocused by the preceding question) and less attentional preparation (lower pretarget alpha).

Discussion

We tested whether discourse focus cues lead listeners to dynamically allocate attention to the anticipated timing of critical information in spoken sentences. Prior findings of enhanced phoneme detection, semantic and syntactic processing, and memory for focused words are often attributed to attentional preparation (Cutler and Fodor, 1979; Kristensen et al., 2013; Wang et al., 2014; Yang et al., 2019; Beier and Ferreira, 2022). Yet direct evidence for preallocation of attention to focused words is lacking. We address this gap by measuring EEG alpha activity, which is widely used to index the time-course of temporal attention in nonlinguistic tasks (Breska and Deouell, 2017; Nobre and Van Ede, 2018) and auditory attention to speech (Boudewyn et al., 2015, 2017; Wöstmann et al., 2016, 2017; Dimitrijevic et al., 2017; Boudewyn and Carter, 2018).

Contrary to our predictions based on the temporal preallocation account, we observed no reliable difference in pretarget alpha power between target words that were focused or defocused by the preceding discourse question. Bayesian analyses provided strong evidence for the null hypothesis, indicating that this absence is unlikely due to insufficient statistical power. Furthermore, alpha was sensitive to other task demands: posterior alpha increased during speech presentation relative to baseline, consistent with gating-by-inhibition of the visual cortex to favor auditory processing (Jensen and Mazaheri, 2010). Following focused late target words, posterior alpha returned to baseline more quickly, suggesting a release of occipital inhibition to prepare for the imminent visual response cue, paralleling temporal-orienting findings outside language (Breska and Deouell, 2017).

While there was no effect on pretarget alpha, discourse focus did modulate the processing of focused words, which elicited a late centroparietal positivity whose amplitude predicted later memory. This aligns with previously reported positive-going parietal SMEs (Paller et al., 1987; Friedman and Johnson Jr., 2000; Mecklinger and Kamp, 2023) linked to familiarity-based memory (Yonelinas et al., 2010; Meng et al., 2014). Indeed, our multiple-choice memory test encouraged familiarity-based recognition rather than free recall. Parietal SMEs may reflect the formation of item-specific memory traces via integration with prior knowledge (Mecklinger and Kamp, 2023). Consistent with this interpretation, in our study focused words were processed within a meaningful sentence context that encouraged semantic integration. Interestingly, the SME (remembered > forgotten) was larger for defocused words. This pattern is compatible with accounts of greater SMEs under variable or suboptimal encoding (Schott et al., 2002): when words are focused, processing is uniformly deep, compressing variance and attenuating SMEs; when not in focus, memory success is dependent on trial-by-trial engagement, yielding larger SMEs.

Critically, we found that trial-by-trial variation in prestimulus alpha predicted both the amplitude of the centroparietal late positivity and its coupling to memory. Greater pretarget alpha (greater preparatory auditory attention via inhibition of visual input) was associated with a larger late positivity, and this relationship was greater for focused words. While not predictive of behavioral performance directly, pretarget alpha modulated the SME of the late positivity: the SME effect was greater for trials with lower pretarget alpha. This again matches prior findings of stronger SMEs under shallow processing (Schott et al., 2002), where attentional variability plays a larger role in determining memory outcomes. Together, these results suggest that both discourse focus and attentional state shape encoding success and that the strength of these effects differs depending on depth of processing and the degree of preparatory attention.

Role of alpha in preparatory attention to speech

Previous work found that increased alpha during the discourse context predicted reduced sensitivity to referential ambiguity (Boudewyn et al., 2015). In the present study, greater pretarget alpha was linked to enhanced memory encoding. This apparent contradiction may indicate different functional roles of alpha across tasks. In the earlier study, increased alpha at central sites during the discourse context likely reflected reduced attention to the unfolding linguistic input, impairing the formation of referential representations. In contrast, in the current study, increased alpha preceded the target word itself; its posterior scalp distribution suggests that it primarily arose through a gating mechanism, whereby visual input was inhibited to prioritize auditory processing (Fu et al., 2001). This illustrates that alpha power is not a unitary marker of “attention” or “engagement” but rather reflects flexible, context-sensitive neural filtering mechanisms. Depending on task demands, increased alpha can either hinder processing (disengaging from critical context) or facilitate it (suppressing irrelevant input), matching prior reports of both alpha synchronization and desynchronization during language comprehension (Dimitrijevic et al., 2017). Furthermore, alpha may index other cognitive processes involved in language comprehension, such as working memory (Prystauka and Lewis, 2019). Future work should clarify these different roles of alpha as a function of task and stimulus characteristics.

Overall, our results do not support the view that listeners preallocate attention to the timing of new information via domain-general attentional mechanisms, as reflected by alpha activity. Prior fMRI and ERP findings may instead reflect bottom–up attention to salient input (e.g., pitch accents; Kristensen et al., 2013) or other anticipatory attentional mechanisms not captured by alpha power (e.g., phase-locking; Schwartze and Kotz, 2013). Furthermore, effects of focus could arise from nonattentional processes, such as semantic prediction (Ferreira and Lowder, 2016) and discourse integration (Spalek and Oganian, 2019).

At the same time, discourse focus shaped how preparatory attentional states influenced encoding. Fluctuations in pretarget alpha predicted post-target memory encoding as reflected in ERPs, and this effect was strongest for focused words. Thus, focus may determine the functional role of these trial-level fluctuations in attention, such that moments of greater preparatory attention are more likely to facilitate the encoding of focused information into memory. These memory-related effects link online discourse processing to subsequent encoding outcomes at both the neural and behavioral levels.

Methodological considerations and future directions

In the present study, target timing varied across trials due to variable sentence durations, improving ecological validity but departing from typical temporal prediction tasks with fixed intervals (Breska and Deouell, 2017). We addressed this through the BOSC method to quantify pretarget alpha regardless of its exact timing, and we still found robust trial-level effects of alpha on post-target responses. Future work could employ a more controlled—albeit less natural—design with consistent target onset across trials.

A second consideration concerns task demands. Due to the large number of stimuli, memory tests were interleaved throughout the experiment, probing both focused and defocused words. This may have encouraged broadly distributed attention across sentences, reducing sensitivity to focus-related alpha modulations. Importantly, however, the focus manipulation modulated ERPs and subsequent memory, in line with prior findings. Thus, although the memory task may have increased global attentiveness, it did not eliminate discourse-driven distinctions between focused and defocused targets. Therefore, the absence of focus effects on pretarget alpha suggests that the observed ERP and memory effects are more parsimoniously attributed to mechanisms other than temporal attentional preallocation as measured by pretarget alpha activity.

The use of clear, read speech may also have reduced the need for effortful attentional preallocation, which may be more relevant with reduced speech intelligibility. Prior work has shown a decrease in alpha power in response to acoustically degraded speech (Obleser and Weisz, 2012); future studies could address whether this attentional effect is modulated by discourse focus.

Finally, we did not see an effect of focus on the N400, but our stimuli did not manipulate semantic constraint. While prior work points to a larger N400 effect for focused semantic violations (Wang et al., 2011; but see Verdonschot et al., 2024), future studies could test whether focus likewise affects the processing of words that were not predicted but that are plausible within the context. This is particularly relevant given that speakers are more likely to focus novel information (which is therefore less predictable, yet plausible). Furthermore, focus and newness have been shown to lead to distinct ERPs (Chen et al., 2014); dissociating the two may lead to new insights into attentional preparation. Overall, a central question for future research is to determine how exactly semantic processing interacts with listeners’ expectations for the timing of upcoming focused words based on discourse, prosodic, and syntactic cues.

Conclusions

Leading theories claim that listeners strategically allocate attention to the anticipated timing of focused input (Cutler and Fodor, 1979). However, we found no evidence for focus-driven preparatory attention: pretarget EEG alpha power—a putative index of temporal attention—was not modulated by focus. This raises the possibility that the cognitive and neural effects of information structure arise from nonattentional processes (e.g., shifts in semantic prediction) or bottom–up attentional capture once relevant input is encountered.

At the same time, pretarget alpha predicted the size of a centroparietal late positivity and its coupling to subsequent memory, suggesting that alpha reflects an encoding-ready state rather than precise temporal orienting to speech. Importantly, we observed a larger SME for trials with lower pretarget alpha, consistent with accounts in which encoding success for less-prioritized input depends more on trial-by-trial engagement.

Together, these findings refine models of attention in language: preparatory alpha modulates the effect of focus on memory encoding, without aligning attention to the precise timing of the focused word. Future work should test the boundary conditions for timing-based preparation (e.g., prosodic/syntactic timing cues, degraded speech) and examine whether oscillatory dynamics beyond changes in power—such as phase alignment or cross-frequency coupling—drive temporal orienting of attention during comprehension.