Cognitive load weakens neural speech tracking without altering response timing Authors: Ruixiang Liu, Huan Zhang, Chang Liu, Xinmeng Xu, Xu Zhao, Xuewen Wang, Fulong Zhang, Xianzheng Sha, Limin Sun, Zhi Wang, Shuo Li, Shijie Chang Published: 2026 ABSTRACT Everyday speech perception often occurs under concurrent cognitive demands, but it remains unclear how speech-extrinsic load influences the neural processing of continuous speech. Here, electroencephalography (EEG) recordings from 29 healthy adults are analyzed with temporal response functions (TRFs), artificial neural network (ANN) decoding, and phase-lag index (PLI)-based connectivity measures during speech listening with or without concurrent mental arithmetic. Higher cognitive load reduces the amount of speech-envelope information recoverable from EEG, as reflected by lower delta-band TRF prediction accuracy and lower ANN-based decoding performance, whereas theta-band encoding accuracy did not differ significantly between conditions. In the delta-band TRF, cognitive load is associated with reduced response amplitude without an obvious shift in dominant peak latency. Higher load is also accompanied by selective changes in delta-band sensor-level phase coupling and a more clustered, more small-world-like topology in the thresholded delta-band PLI subnetwork. These findings suggest that speech-extrinsic cognitive load weakens neural representations of continuous speech and alters frontotemporal coupling patterns, even in the absence of speech degradation or competing talkers. Introduction Speech perception in everyday life rarely occurs in isolation, as people often need to understand speech while simultaneously holding information in mind, monitoring other task-relevant information, or planning an action1,2. Understanding how cognitive demands shape speech perception is therefore important not only for basic research on auditory attention and working memory, but also for related domains such as effortful listening3, age-related communication difficulties4, and audiological rehabilitation. Much of the existing literature mainly explores this issue through two typical scenarios: one focuses on speech perception under conditions of impaired sensory input, and the other concentrates on the processing mechanisms in concurrent speech stream competition scenarios. In the former type of research, the challenge lies in signal degradation, such as background noise, vocoding, or hearing-related acoustic impoverishment, which reduces the quality of bottom-up sensory evidence and increases listening effort3,5,6. In the latter type of research, the challenge lies in competition between streams, as in cocktail-party or two-talker paradigms, where listeners must selectively track one speech stream while suppressing another7,8. These two lines of work have shown that speech processing is highly sensitive to both acoustic adversity and attentional competition. In addition to these research paths, another group of works focuses on the impact of cognitive load on continuous speech perception, typically using a dual-task paradigm for manipulation. These studies are of great significance for revealing how high-level cognitive resources (such as working memory and attention control) participate in and regulate the speech processing process. However, studies manipulating external speech load are currently mainly at the behavioral level, focusing on the influence of cognitive load on speech recognition accuracy, word weight allocation, and time progression indicators9–12. Experimental work has also begun to address related issues. However, the dual tasks used in existing studies have typically remained closely tied to speech processing itself. That is, memory or control demands directly act on the auditory or linguistic material being processed. For example, Hjortkjær et al. 13 used an auditory n-back task performed on spoken number sequences, such that the speech stream itself served as the target of the memory operation, whereas Liu et al. 14 examined how working memory load modulates cortical tracking at different linguistic timescales during sentence processing. These paradigms are highly informative for understanding how mnemonic demands interact with speech-related processing. Although these paradigms differ in the source of challenge they impose on listening, they converge on a broader question: which aspect of speech processing is altered when perceptual and cognitive demands increase? In principle, poorer performance under adverse listening conditions may arise from degraded sensory encoding, disrupted temporal alignment to speech, altered higher-order linguistic integration, or a redistribution of resources across large-scale control networks. To examine how these factors affect continuous speech perception, a neural measure is needed that captures the degree to which cortical activity tracks ongoing speech. Continuous-speech neural tracking provides such a measure, because it quantifies the alignment between cortical activity and the temporal structure of continuous speech and can therefore be used to assess whether a given challenge primarily alters the strength, timing, or hierarchical organization of speech representations. Low-frequency cortical activity aligns with the temporal structure of speech, and this tracking is widely interpreted as reflecting the transformation of continuous acoustic input into progressively higher-order speech representations15–17. Importantly, delta- and theta-band activity are not functionally redundant. Theta-band tracking has been linked more closely to syllabic parsing and lower-level temporal sampling, whereas delta-band tracking is more strongly associated with slower prosodic, phrasal, and sentential organization18,19. In competitive-listening paradigms, prior work has emphasized selective neural tracking of attended speech during stream competition, and some studies further suggest that distraction can disrupt the temporal fidelity or phase consistency of speech-related neural responses, particularly in the theta range7,20,21. Recent evidence indicates that such distraction reflects a combination of reduced sensory gain and increased temporal jitter, with unattended sounds being most susceptible22. Moreover, selective attention operates along a hierarchy: near the auditory cortex, neural responses are amplitude‑modulated for both attended and ignored speech, whereas higher‑order regions become fully selective for the attended stream, a selectivity that sharpens over time7. This selective entrainment, which degrades as the number of competing talkers increases23, exemplifies ‘active sensing’, where top‑down predictions dynamically shape bottom‑up input24. Beyond local cortical tracking, speech perception is also shaped by interactions with broader control systems, including frontal influences on auditory entrainment25,26. Together, these findings indicate that cognitive demands may influence speech processing at multiple levels, from local neural representation to distributed network coordination. Despite this progress, an important gap remains. Existing studies that isolate extrinsic load from the speech material have been largely behavioral or behavior-dominant, whereas neural studies have often used paradigms in which the concurrent demand is closely coupled to the auditory or linguistic content itself. As a result, it remains unclear how an independent, speech-extrinsic cognitive load influences continuous speech processing across multiple neural levels simultaneously. In particular, few studies have examined whether such load reshapes both the local representation of speech and the broader functional coordination among brain regions during natural continuous speech perception. This distinction is theoretically important because a reduction in speech tracking under load could reflect not only weaker local encoding of speech but also a reallocation of resources at the network level. The present study was designed to address this gap. Participants listened to a single continuous speech stream under a low-load condition and under a high-load condition that required concurrent mental arithmetic. This design isolates a speech-extrinsic dual-task condition in which continuous speech perception and an independent cognitive task compete for limited resources in real time. To assess the neural consequences of this competition, we used TRF analysis to quantify the strength and latency of neural speech tracking, ANN-based decoding to provide a complementary measure of recoverable speech-envelope information from EEG, PLI-based analysis to test for condition-related changes in sensor-level phase coupling, and graph-theoretic metrics to evaluate whether cognitive load was associated with broader reorganization of network topology during speech listening. We therefore asked whether higher cognitive load primarily weakens speech representations, alters their temporal dynamics, or both, and whether these local effects are accompanied by changes in distributed network organization. Results Behavioral results Figure 1 presents the NASA-TLX scores for participants under the low cognitive load (LCL) and high cognitive load (HCL) conditions. Under the HCL condition, NASA-TLX scores were significantly higher than under the LCL condition (p < 0.001), with a mean score of 60.9 (SD = 8.4) in the HCL condition compared to 31.3 (SD = 5.7) in the LCL condition, indicating that the manipulation of subjective cognitive load was successful. A paired-sample t-test on the HCL–LCL difference confirmed this increase, t(28) = 17.39, p < 0.001, Cohen’s d = 3.23, mean difference = 29.59, 95% CI [26.12, 33.07]. Consistent with this effect, the overall distribution of NASA-TLX scores shifted upward in the HCL condition relative to the LCL condition, with a higher median and a wider range of scores. These results confirm that participants experienced significantly greater task demands in the HCL condition than in the LCL condition. Supplementary Figure 1 illustrates the distribution of the number of responses and accuracy rates across participants. Only one participant had an accuracy rate below 80%, and the EEG recordings were excluded from subsequent analyses.Fig. 1Box plots illustrate the cognitive load assessment distributions under LCL and HCL conditions: LCL represents cognitive load assessments in pure auditory experiments, while HCL denotes assessments in auditory experiments combined with mental arithmetic tasks.Results indicate that HCL conditions consistently yielded higher scores than LCL conditions, with statistically significant differences (paired tests, p < 0.001), suggesting that dual-tasking effectively increases cognitive load. Boxplots show the median, interquartile range, and whiskers extending to 1.5 times the interquartile range. Each point represents one participant; n = 29 biologically independent participants. The sign of n.s. means no significance, * means 0.01 ≤ p < 0.05, ** means 0.001 ≤ p < 0.01, and *** means 0.001 ≤ p < 0.001. Box plots illustrate the cognitive load assessment distributions under LCL and HCL conditions: LCL represents cognitive load assessments in pure auditory experiments, while HCL denotes assessments in auditory experiments combined with mental arithmetic tasks. Results indicate that HCL conditions consistently yielded higher scores than LCL conditions, with statistically significant differences (paired tests, p < 0.001), suggesting that dual-tasking effectively increases cognitive load. Boxplots show the median, interquartile range, and whiskers extending to 1.5 times the interquartile range. Each point represents one participant; n = 29 biologically independent participants. The sign of n.s. means no significance, * means 0.01 ≤ p < 0.05, ** means 0.001 ≤ p < 0.01, and *** means 0.001 ≤ p < 0.001. Cognitive load reduces the speech-envelope information recoverable from the EEG We first examined whether concurrent cognitive load reduced the amount of speech-envelope information recoverable from the EEG. TRF-based encoding analysis revealed significantly lower prediction accuracy in the delta band under high cognitive load than under low cognitive load (LCL = 0.049 ± 0.021; HCL = 0.033 ± 0.022; paired-sample t-test, t(28) = 3.73, p < 0.001, Cohen’s d = 0.69; Fig. 2). A similar condition difference was observed for the full band (LCL = 0.038 ± 0.015; HCL = 0.026 ± 0.015; paired-sample t-test, t(28) = 3.78, p < 0.001, Cohen’s d = 0.70). In contrast, no significant within-participant HCL–LCL differences were found in the theta band (t(28) = 1.59, p = 0.122, Cohen’s d = 0.30), alpha band (t(28) = 0.26, p = 0.793, Cohen’s d = 0.05), beta band (t(28) = 0.20, p = 0.846, Cohen’s d = 0.04), or gamma band (t(28) = 0.67, p = 0.510, Cohen’s d = 0.12).Fig. 2Neural encoding accuracy across frequency bands.Subplots A–F represent the neural decoding scores for the Fullband, Delta, Theta, Alpha, Beta, and Gamma frequency bands, respectively. The values represent the mean Pearson correlation coefficient between the TRF-predicted EEG and the actual EEG across channels. Results are shown separately for the full-band signal and for delta, theta, alpha, beta, and gamma bands. Each data point corresponds to the averaged accuracy for a single subject within each band. Boxplots show the median, interquartile range, and whiskers extending to 1.5 times the interquartile range. Each point represents one participant; n = 29 biologically independent participants. The sign of n.s. means no significance, * means 0.01 ≤ p < 0.05, ** means 0.001 ≤ p < 0.01, and *** means 0.001 ≤ p < 0.001. Neural encoding accuracy across frequency bands. Subplots A–F represent the neural decoding scores for the Fullband, Delta, Theta, Alpha, Beta, and Gamma frequency bands, respectively. The values represent the mean Pearson correlation coefficient between the TRF-predicted EEG and the actual EEG across channels. Results are shown separately for the full-band signal and for delta, theta, alpha, beta, and gamma bands. Each data point corresponds to the averaged accuracy for a single subject within each band. Boxplots show the median, interquartile range, and whiskers extending to 1.5 times the interquartile range. Each point represents one participant; n = 29 biologically independent participants. The sign of n.s. means no significance, * means 0.01 ≤ p < 0.05, ** means 0.001 ≤ p < 0.01, and *** means 0.001 ≤ p < 0.001. To further assess the recoverability of speech envelope information from EEG, we performed stimulus reconstruction using an ADT Network–based decoding model and quantified decoding accuracy as the Pearson correlation coefficient between the reconstructed and original speech envelopes on the test set. Accuracy was significantly lower in the HCL condition than in the LCL condition (LCL: 0.105 ± 0.034; HCL: 0.087 ± 0.029; paired-sample t-test on the LCL–HCL difference, t(28) = 3.64, p = 0.001, Cohen’s d = 0.68, mean difference = 0.0189, 95% CI [0.0083, 0.0295]; Fig. 3). Together, these results indicate that concurrent cognitive load reduces the neural recoverability of speech-envelope information, with the most prominent effect observed in the delta band.Fig. 3Pearson correlation coefficients for speech envelope reconstruction using the ADT Network across two experimental conditions, with each sample representing the averaged results from each subject.Boxplots show the median, interquartile range, and whiskers extending to 1.5 times the interquartile range. Each point represents one participant; n = 29 biologically independent participants. The sign of n.s. means no significance, * means 0.01 ≤ p < 0.05, ** means 0.001 ≤ p < 0.01, and *** means 0.001 ≤ p < 0.001. Pearson correlation coefficients for speech envelope reconstruction using the ADT Network across two experimental conditions, with each sample representing the averaged results from each subject. Boxplots show the median, interquartile range, and whiskers extending to 1.5 times the interquartile range. Each point represents one participant; n = 29 biologically independent participants. The sign of n.s. means no significance, * means 0.01 ≤ p < 0.05, ** means 0.001 ≤ p < 0.01, and *** means 0.001 ≤ p < 0.001. Cognitive load attenuates response magnitude without altering temporal alignment We next examined whether the load-related reduction in speech tracking was associated with altered temporal timing or reduced response magnitude. In the delta-band TRF, the latency of the dominant response peak did not differ significantly between conditions (Wilcoxon signed-rank test, W = 186, p = 0.510, rank-biserial r = 0.058), with peak responses occurring at comparable latencies in HCL and LCL (approximately 343 and 351 ms, respectively; Fig. 4A, C). In contrast, peak value differed significantly between conditions, with a smaller absolute response under HCL than under LCL (LCL = −0.00385; HCL = −0.00283; paired-sample t-test on the LCL–HCL difference, t(28) = − 2.43, p = 0.022, Cohen’s d = −0.45; Fig. 4B).Fig. 4Analysis of TRF weights for the envelope in the delta frequency band.A The box plot of peak latency for TRF. B The box plot of absolute peak intensity for TRF. C TRF weights in the Cz channel. D–F Source analysis of differences between the two experimental conditions (HCL–LCL) at peak time. G, H Scalp topography maps of TRF weights for HCL and LCL. Boxplots show the median, interquartile range, and whiskers extending to 1.5 times the interquartile range. Each point represents one participant; n = 29 biologically independent participants. The sign of n.s. means no significance, * means 0.01 ≤ p < 0.05, ** means 0.001 ≤ p < 0.01, and *** means 0.001 ≤ p < 0.001. Analysis of TRF weights for the envelope in the delta frequency band. A The box plot of peak latency for TRF. B The box plot of absolute peak intensity for TRF. C TRF weights in the Cz channel. D–F Source analysis of differences between the two experimental conditions (HCL–LCL) at peak time. G, H Scalp topography maps of TRF weights for HCL and LCL. Boxplots show the median, interquartile range, and whiskers extending to 1.5 times the interquartile range. Each point represents one participant; n = 29 biologically independent participants. The sign of n.s. means no significance, * means 0.01 ≤ p < 0.05, ** means 0.001 ≤ p < 0.01, and *** means 0.001 ≤ p < 0.001. This pattern was also evident in the condition-averaged TRF waveforms at Cz channel, which showed similar overall morphology and peak timing across conditions, but a weaker response magnitude in HCL within the main response window (Fig. 4C). Consistent with this observation, scalp topographies at the peak latency showed broadly similar spatial distributions in the two conditions, with lower response strength under HCL than under LCL (Fig. 4G, H). Source-level contrast maps further indicated that the within-participant HCL–LCL difference at the peak response window was concentrated predominantly over bilateral temporal regions, particularly along the lateral temporal cortex, with comparatively limited differences in other cortical areas (Fig. 4D–F). Together, these results suggest that increased cognitive load primarily reduces the magnitude of delta-band speech tracking while leaving its principal temporal profile largely unchanged, and that this attenuation is most prominently expressed over temporal cortical regions. High cognitive load selectively enhances delta-band sensor-level phase coupling We next examined whether the reduction in speech-envelope tracking was accompanied by altered delta-band phase coupling. Because the unthresholded PLI matrices were relatively dense (Fig. 5A, B), we focused on the statistically significant difference network. In the HCL–LCL difference matrix, black boxes mark edges showing significant within-participant HCL–LCL differences (Fig. 5C). These significant effects were confined to a limited subset of connections rather than being broadly distributed across the scalp, with most showing stronger coupling under the HCL condition than under the LCL condition.Fig. 5Delta-band PLI connectivity and the key differential subnetwork.A, B Condition-averaged PLI connectivity matrices for the LCL and HCL conditions. C Difference matrix (HCL–LCL); black boxes indicate edges with significant within-participant HCL–LCL differences. D–F Key subnetwork extracted from the significant difference network, summarizing the dominant connected structure of altered delta-band coupling. In panels A and B, color scale bars indicate mean PLI values. In panel C, the color scale bar indicates HCL–LCL PLI differences. Connectivity statistics were computed using within-participant HCL–LCL comparisons; n = 29 biologically independent participants, with 5000 permutations. Delta-band PLI connectivity and the key differential subnetwork. A, B Condition-averaged PLI connectivity matrices for the LCL and HCL conditions. C Difference matrix (HCL–LCL); black boxes indicate edges with significant within-participant HCL–LCL differences. D–F Key subnetwork extracted from the significant difference network, summarizing the dominant connected structure of altered delta-band coupling. In panels A and B, color scale bars indicate mean PLI values. In panel C, the color scale bar indicates HCL–LCL PLI differences. Connectivity statistics were computed using within-participant HCL–LCL comparisons; n = 29 biologically independent participants, with 5000 permutations. Spatially, the significant edges were concentrated mainly in a fronto-temporal network, with additional links extending to parietal and occipital electrodes (Fig. 5C). To summarize the dominant organization of these effects, Fig. 5D–F shows the key subnetwork extracted from the significant difference network. This subnetwork was centered on frontal and temporal nodes and included several long-range anterior–posterior connections, indicating that the load-related connectivity differences were sparse but topologically structured rather than diffuse. Overall, increased cognitive load was associated with selective changes in delta-band scalp-level phase coupling, concentrated in a sparse fronto-temporal-centered subnetwork. Graph metrics indicate a more clustered and more strongly small-world delta-band topology under HCL To further characterize the topology of the thresholded delta-band PLI network, we computed graph-theoretic metrics including density, clustering coefficient, characteristic path length, global efficiency, edge count, and the small-worldness index sigma (Fig. 6). Compared with LCL condition, the HCL condition exhibited significantly higher density (paired-sample t-test on the LCL–HCL difference, t(28) = − 2.79, p = 0.009, Cohen’s d = −0.52, 95% CI [−0.257, −0.040]), significantly higher clustering coefficient (t(28) = − 2.81, p = 0.009, Cohen’s d = −0.52, 95% CI [−0.267, −0.042]), significantly shorter characteristic path length (t(23) = 3.72, p = 0.001, Cohen’s d = 0.76, 95% CI [0.026, 0.090]), and significantly higher sigma (t(28) = − 2.51, p = 0.009, Cohen’s d = −0.47, 95% CI [−0.509, -0.010]). Global efficiency showed an increasing trend but did not reach significance (t(28) = − 1.91, p = 0.066, Cohen’s d = −0.35, 95% CI [ − 0.250, 0.009]), whereas edge count did not differ significantly between conditions (t(28) = − 1.38, p = 0.178, Cohen’s d = −0.26, 95% CI [ − 27.316, 5.316]).Fig. 6Graph-theoretic summary of the thresholded delta-band PLI subnetwork.Boxplots show condition-level distributions of density, clustering coefficient, characteristic path length, global efficiency, edge count, and small-worldness index sigma for the LCL and HCL conditions. Relative to LCL, HCL showed significantly higher density, clustering coefficient, and sigma, as well as significantly shorter characteristic path length. Global efficiency showed a nonsignificant increase, whereas edge count did not differ significantly between conditions. Boxplots show the median, interquartile range, and whiskers extending to 1.5 times the interquartile range. Each point represents one participant; n = 29 biologically independent participants for panels (A, B, D–F), and n = 24 biologically independent participants for panel (C). The sign of n.s. means no significance, * means 0.01 ≤ p < 0.05, ** means 0.001 ≤ p < 0.01, and *** means 0.001 ≤ p < 0.001. Graph-theoretic summary of the thresholded delta-band PLI subnetwork. Boxplots show condition-level distributions of density, clustering coefficient, characteristic path length, global efficiency, edge count, and small-worldness index sigma for the LCL and HCL conditions. Relative to LCL, HCL showed significantly higher density, clustering coefficient, and sigma, as well as significantly shorter characteristic path length. Global efficiency showed a nonsignificant increase, whereas edge count did not differ significantly between conditions. Boxplots show the median, interquartile range, and whiskers extending to 1.5 times the interquartile range. Each point represents one participant; n = 29 biologically independent participants for panels (A, B, D–F), and n = 24 biologically independent participants for panel (C). The sign of n.s. means no significance, * means 0.01 ≤ p < 0.05, ** means 0.001 ≤ p < 0.01, and *** means 0.001 ≤ p < 0.001. Overall, this metric profile indicates that high cognitive load was associated with a more locally clustered and topologically compact delta-band network, together with a stronger tendency toward small-world organization. The increase in the small-worldness index sigma, accompanied by higher clustering and shorter path length, suggests that the HCL network exhibited more pronounced small-world properties than the LCL network. Discussion In this study, we examined how a speech-extrinsic cognitive load modulates continuous speech processing when listeners hear a single ongoing speech stream while simultaneously performing mental arithmetic. Three main findings emerged. First, cognitive load reduced the amount of speech-envelope information recoverable from EEG, as reflected by lower delta-band TRF prediction accuracy and lower ANN-based decoding performance under HCL conditions. Second, in the delta-band TRF, the HCL condition was associated with reduced response amplitude but not a shift in the dominant peak latency, suggesting that the principal temporal profile of the response was preserved at the condition level even when response strength was attenuated. Third, the HCL condition was accompanied by selective changes in delta-band sensor-level phase coupling and by a more clustered, more strongly small-world topology of the thresholded delta-band PLI subnetwork. Together, these findings indicate that speech-extrinsic load affects both local speech representation and large-scale coupling during continuous speech perception. Our results extend prior work on speech perception under adverse or competitive conditions by showing that an independent concurrent cognitive task can weaken speech-related neural representations even when the speech signal itself is not degraded, and no competing talker is present27. In the present data, the load effect was most prominent in the delta range, whereas theta-band encoding accuracy did not differ significantly between conditions in the main analysis. This pattern is compatible with the idea that speech-extrinsic load may particularly affect slower processes related to prosodic or phrasal-scale organization28,29, while the current dataset provides less evidence for a robust theta-band reduction in envelope encoding. At the same time, because acoustic-envelope tracking can covary with higher-order linguistic structure30,31, the present results should not be interpreted as reflecting purely acoustic processing in isolation. Rather, they likely index the aggregate consequence of load on temporally aligned speech representations across multiple hierarchical levels. The dissociation between reduced TRF amplitude and preserved peak latency is central to our interpretation. One possibility is that cognitive load reduces the effective neural gain allocated to speech processing, yielding weaker response magnitude while leaving the dominant response timing relatively stable. This interpretation is consistent with resource-competition accounts, according to which concurrent mental arithmetic draws on limited executive resources that would otherwise support speech processing6,14. However, preserved condition-level peak latency argues against a large shift in the dominant response timing, but it does not completely exclude finer-grained temporal variability or trial-to-trial jitter that may also contribute to an apparent amplitude reduction in averaged TRF32,33. Since the peak width did not show an obvious change, the present pattern appears more consistent with reduced response gain than with markedly increased temporal dispersion, although finer-grained latency variability cannot be ruled out. At the network level, the increase in cognitive load was associated with selective changes in sensor-level phase coupling in the delta band; significant differences were concentrated in a sparse subnetwork centered on frontal and temporal electrodes and extending to parietal and occipital regions. Conservatively speaking, the data suggest that dual-task loading altered coupling patterns at the scalp level, while the intensity of speech tracking was reduced. The accompanying increase in clustering and sigma values, coupled with a reduction in characteristic path length, indicates that under HCL conditions, the network configuration shifts toward a more locally clustered and topologically compact state, known as small-world organization. One possible explanation is that this configuration reflects an adaptive redistribution of processing capacity under high load conditions34,35. Of course, another possible explanation is that the change in topological structure may partly reflect the superposition of couplings associated with the two parallel tasks (speech listening and mental arithmetic), rather than a pure reorganization of a single speech network. The absence of an edge-count difference after hard-thresholding makes a trivial edge-density account less likely, but it does not rule out contributions from arithmetic-related coupling superimposed on speech-related interactions. The ANN-based decoding results complement the TRF findings by showing that the load-related reduction in speech recoverability is not limited to a linear encoding framework. Taken together, the TRF, decoding, and connectivity analyses provide a broader description of how cognitive load alters continuous-speech processing across representational and network levels. Several limitations should be noted. First, the secondary task was limited to mental arithmetic, and different forms of cognitive demand may produce different patterns of interference. Second, the functional-connectivity results were obtained at the sensor level and were based on PLI36, which limits anatomical specificity and does not address directionality37. Third, the speech materials were synthesized using text-to-speech, which helped reduce speaker-specific variability but may not capture all properties of naturally produced speech. Finally, the sample consisted of young healthy adults, so generalization to other populations remains to be established. In summary, speech-extrinsic cognitive load reduced the amount of speech-envelope information recoverable from EEG during continuous speech listening, as shown by lower delta-band TRF prediction accuracy and lower ANN-based decoding performance. This effect was expressed mainly as reduced delta-band TRF amplitude without an obvious shift in dominant peak latency, suggesting attenuated response strength with relative preservation of the principal temporal profile at the condition level. Higher load was also associated with selective changes in delta-band sensor-level phase coupling and a more clustered, more small-world-like topology in the thresholded delta-band PLI subnetwork. Together, these findings indicate that cognitive load alters both local neural representations of speech and large-scale coupling patterns during continuous speech processing. Methods Participants The volunteers for this study were recruited from China Medical University, comprising 31 undergraduate and graduate students, including 16 females and 15 males, aged between 20 and 27 years (M = 22.9, SD = 2.1). All participants were native Mandarin speakers with normal hearing thresholds (≤20 dB), were right-handed, and had no history of auditory or cognitive disorders. All participants completed the entire experiment without knowledge of its purpose. Prior to the experiment, subjects were informed of the procedure and signed written informed consent forms. The study was approved by the Ethics Committee of China Medical University and the Shengjing Hospital of China Medical University (No. 2023PS1188K). All ethical regulations relevant to human research participants were followed. Experiments Figure 7 shows the details of our experiment. The experiment consisted of two within-participant conditions: a passive listening condition, in which participants listened to continuous speech, and a dual-task listening condition, in which participants listened to continuous speech while simultaneously performing complex mental arithmetic. In each part, subjects heard six audio clips through headphones, averaging approximately 5 min in length, with an average loudness of 60 dB per ear. Participants rested for 2–4 min between each pair of audio clips to alleviate fatigue. A longer break was provided between the two experimental sections. During this period, participants will be asked to complete the NASA-TLX scale to assess their cognitive load while performing the task.Fig. 7Schematic of the experimental paradigm for the auditory and dual-task conditions.A Experimental procedure showing repeated audio–rest blocks under two conditions, followed by NASA-TLX workload assessment and an inter-condition rest interval; B EEG acquisition setup for low-load passive listening and high-load dual-task listening during audio stimulation. The experiment consisted of two within-subject conditions designed to manipulate cognitive load: a passive listening (single-task) condition representing low cognitive load (LCL), and a dual-task listening condition representing high cognitive load (HCL). In both conditions, participants listened to six synthesized speech clips (derived from Animal Farm, ~5 min each, ~60 dB SPL) presented via headphones. Created in BioRender. Ruixiang, L. (2026) https://BioRender.com/vmlv8op. Schematic of the experimental paradigm for the auditory and dual-task conditions. A Experimental procedure showing repeated audio–rest blocks under two conditions, followed by NASA-TLX workload assessment and an inter-condition rest interval; B EEG acquisition setup for low-load passive listening and high-load dual-task listening during audio stimulation. The experiment consisted of two within-subject conditions designed to manipulate cognitive load: a passive listening (single-task) condition representing low cognitive load (LCL), and a dual-task listening condition representing high cognitive load (HCL). In both conditions, participants listened to six synthesized speech clips (derived from Animal Farm, ~5 min each, ~60 dB SPL) presented via headphones. Created in BioRender. Ruixiang, L. (2026) https://BioRender.com/vmlv8op. The speech stimuli were synthesized from passages of Animal Farm using an online text-to-speech platform (https://www.text-to-speech.cn/) rather than human voice recordings, in order to minimize variability arising from speaker-specific emotion, prosody, and other paralinguistic factors. Default synthesis settings were used for pitch, and the speaking rate was approximately 5 Chinese characters per second. The inter-sentence pause was set to 50 ms to maintain the continuity of the speech stream. No additional manual adjustment of acoustic parameters was applied. Data acquisition was conducted in a sound-attenuated, electromagnetically shielded chamber. Throughout the experiment, an operator continuously monitored the subject to ensure task compliance. Prior to the session, participants were explicitly instructed that vigilance checks, consisting of three content-comprehension questions, would be administered immediately following each audio segment. In the low cognitive load condition, these questions served to verify attentive listening. In the high cognitive load condition, in addition to the comprehension checks, participants were supervised to perform the mental arithmetic task, with specific instructions to maximize calculation accuracy while maintaining attention to the speech stream. Under the high cognitive load condition, the mental arithmetic questions were designed such that the next question appeared only after a correct answer was given. Consequently, the number of questions answered varied across participants. Accuracy was above 80% for all participants except for one, who reported after the experiment that the mental arithmetic task was too difficult and completed only three questions. We excluded this participant’s data from subsequent analyses. The distribution of response counts and accuracy rates is reported in Supplementary Fig. 1. EEG recording and preprocessing The ANT Neuro system was used to record subjects’ electroencephalogram (EEG) activity, employing 61 standard electrode positions with a sampling rate of 500 Hz. Additionally, electrooculogram (EOG) and mastoid electrodes (M1 and M2) were recorded. Preprocessing was performed using the EEGLAB toolbox38. Data underwent three-step initial filtering using finite impulse response (FIR) filters: first, a notch filter (cutoff = [49, 51] Hz; order = 3380), followed by a high-pass filter (low cutoff = 0.5; order = 16,896) and a low-pass filter (high cutoff = 64 Hz; order = 100). These filters are zero-phase, non-causal filters. Following the EEGLAB toolbox documentation, the bandpass filtering process is achieved through one high-pass filter and one low-pass filter. Next, the resting-state EEG recordings were removed from the data. The remaining data was downsampled to 128 Hz, followed by independent component analysis (ICA) to remove eye and muscle artifacts. Three experienced operators identified artifact components among the 61 components. Two operators made independent judgments; components were discarded or retained only if both operators agreed they represented artifacts or brain signals. Disputed components were resolved by a third operator’s final decision. One subject was excluded because the noise level following ICA analysis of the recording was too high, and most of the independent components were classified as “others.” In the 29 recordings included in the analysis, an average of 13.8 ± 3.4 independent components were removed from the LCL condition, compared to 17.6 ± 4.2 from the HCL condition. After ICA-based artifact correction, EEG amplitudes in all processed recordings were confined to within ±30 µV, consistent with a substantial reduction of large-amplitude artifacts. The data were then saved with all subject-identifying information removed. After preprocessing the data, we then used MNE-Python39 to extract and save the EEG data. The EEG signals were filtered into separate bands using the firwin function from SciPy via MNE’s built-in filtering routines. The bands included delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–14 Hz), beta (14–32 Hz), gamma (32–64 Hz), and a wideband (0.5–32 Hz). For the audio signals, speech envelopes were extracted following procedures commonly used in previous studies. Specifically, each audio signal was filtered using a 28-channel gammatone filterbank with center frequencies logarithmically distributed between 50 Hz and 5 kHz. The envelope of each subband was obtained via the Hilbert transform. These envelopes were then compressed using a power-law exponent of 0.6, averaged across subbands, and downsampled to 128 Hz to match the EEG data. Consistent with the six EEG frequency bands analyzed, the speech envelope was also decomposed into six corresponding frequency bands. TRF modeling The Temporal Response Function (TRF)40 is a method that analyzes the dynamic relationship between neural signals in the brain and continuous stimuli (such as speech envelopes) through time-lag regression analysis. Its core principle involves constructing a linear convolution model to quantify the influence of stimulus features on neural responses at different time delays. By fitting the delay relationship between input and output signals, TRF helps us understand the temporal relationship between internal neural oscillations and external stimuli. TRF models typically estimate the influence of input signals on output signals through linear regression. The TRF model can generally be expressed in the following linear regression form:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X(t)={\int }_{{t}_{min }}^{{t}_{max }}h\left(\tau \right)S\left(t-\tau \right)d\tau +\epsilon (t)$$\end{document}X(t)=∫tmintmaxhτSt−τdτ+ϵ(t)where, X(t)X(t) represents the neural response of the EEG signal at time t; S(t)S(t) denotes the stimulus signal (e.g., speech envelope); h(t)h(t) indicates external input received by the brain; dτdτ signifies the weighting factor of the stimulus signal’s influence on the EEG signal at different time lags (τ )(τ) which characterizes the temporal correlation of the EEG’s response to the stimulus; ε (t)ϵ(t) represents noise information; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{\min } \sim {t}_{\max }$$\end{document}tmin~tmax defines the time window that determines the brain’s response range to the corresponding stimulus. In this study, the Eelbrain toolbox41 was employed to compute TRFs for quantifying cortical encoding of speech envelopes. Eelbrain utilizes a Boosting algorithm to fit TRF models, mitigating overfitting phenomena observed in correlated features. The Boosting algorithm is an optimization method based on Coordinate Descent. When constructing multivariate temporal response functions (mTRFs), it optimizes by incrementally adjusting the weights of individual filters. Unlike Ridge Regression, which uniformly adjusts all filter weights through global optimization, Boosting employs sparse learning strategies. After each weight update, the model is evaluated against validation data, enabling more localized filter weight adjustments. Training halts when error reduction ceases, preventing overfitting and ensuring irrelevant weights remain zero, ultimately yielding a sparse solution. The core formula of the Boosting algorithm is as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h}_{j}(t+1)={h}_{j}(t)+{{\eta }}{\sum }_{i=1}^{N}{S}_{i}.{r}_{i}(t)$$\end{document}hj(t+1)=hj(t)+η∑i=1NSi.ri(t) Through the Boosting algorithm, h(t)h(t) is progressively optimized to reduce prediction error \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{i}(t)$$\end{document}ri(t). Specifically in this study, the Boosting algorithm updates only one filter weight h(t)h(t) per iteration, enabling the objective loss function to converge gradually. Consequently, the Boosting algorithm provides a more sparse, interpretable, and efficient optimization strategy for TRF response modeling. Its mechanism of progressively adjusting weights combined with cross-validation allows it to adapt to complex neural signal processing tasks. For each subject, the TRF model was trained on 12 trials across the six frequency bands, yielding filter weights for each trial at different frequencies. During model training, to prevent overfitting to the onset effect of speech sounds, a 4-min segment was extracted from the middle of each trial for 4-fold cross-validation without overlapping. Each segment was thus used for training three times and testing once across all folds. The training segment was used to adjust the model, the validation segment to optimize it, and the test segment to evaluate it. The base function in Eelbrain was adjusted to 1 × 10−³, and scale_data was set to True. The model’s time lag was fitted from −200 to 600 ms, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{\min }$$\end{document}tmin corresponds to −200 ms and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{\max }$$\end{document}tmax to 600 ms. Source estimation of TRF weights difference Cortical source localization of the TRF weights was performed using MNE-Python. The forward model was constructed based on the standard fsaverage brain template42 with a surface-based source space (oct-6 spacing). A three-layer boundary element method (BEM) model (conductivity: 0.3, 0.006, and 0.3 S/m for scalp, skull, and brain, respectively) was used to compute the forward solution. For the inverse problem, the peak latency of the grand-averaged difference waveform (e.g., LCL minus HCL) within the 0–600 ms time window was identified. An ad-hoc noise covariance matrix was generated, and the cortical current density at the peak latency was estimated using the Minimum Norm Estimate (MNE) method (SNR = 3.0), incorporating depth weighting (0.8) and loose orientation constraints (0.2). ANN-based model for decoding speech envelope In this study, an artificial neural network named Auditory Decoding Transformer (ADT) network proposed by Liu et al. is employed for reconstructing envelopes from EEG signals43. The ADT network achieves interpretable and robust speech envelope decoding through spatio-temporal convolutions and an inverse-causal masking mechanism. The input data consisted of wideband EEG (0.5–32 Hz) and the corresponding speech envelope. To prevent data leakage, each continuous recording was first split into non-overlapping temporal segments before data augmentation. The first 40% and last 40% of each recording were used for training, whereas the middle 20% was equally divided into validation (10%) and test (10%) sets. Data augmentation was applied only to the training set after the split. Consequently, no samples from the validation or test sets, nor their augmented variants, were included in model training. Performance was evaluated on the held-out test set. Phase lag index connectivity analysis In functional connectivity analysis, coupling methods from signal processing are employed to measure correlations between different electrode pairs or between two brain regions, yielding the corresponding brain functional connectivity matrix. The size of this matrix is typically determined by the number of electrode channels and the volume of experimental data. Each value within the matrix represents the correlation or causal relationship between the electrode pair or between the two brain regions. To mitigate the influence of common-source artifacts on analysis outcomes, this study employs the volume-conductance-independent PLI metric for functional connectivity calculations36. In this study, the dimensions of the preprocessed EEG data are \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E\times T\times S$$\end{document}E×T×S, where EE represents the number of electrodes, TT denotes the number of points per segment, and SS indicates the total number of segments. Each segment spans 4 s with no overlap, and the sampling rate is 128 Hz, resulting in TT= 512. PLI measures the degree of phase coherence between electrodes XX and YY across specific frequency bands. First, bandpass filtering is applied to the EEG data from electrodes XX and YY to select the target frequency bands. The filtered signals from electrodes XX and YY undergo a Hilbert transformation to extract the phase values \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varphi }_{x}$$\end{document}φx and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varphi }_{y}$$\end{document}φy at each time point. The Hilbert transformation formula is as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H\left(x\left(t\right)\right)=\frac{1}{\pi }{\int }_{\!\!\!\!-\infty }^{\infty }\frac{x(\tau )}{t-\tau }{d}_{\tau }$$\end{document}Hxt=1π∫−∞∞x(τ)t−τdτwhere, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x\left(t\right)$$\end{document}xt is the original signal, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H\left(x\left(t\right)\right)$$\end{document}Hxt is the Hilbert transform of the original signal. Based on the obtained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varphi }_{x}$$\end{document}φx and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varphi }_{y}$$\end{document}φy values, calculate the phase difference \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \varphi ={\varphi }_{x}{-\varphi }_{y}$$\end{document}Δφ=φx−φy between electrode XX and electrode YY. Then compute the PLI for each segment using the following formula:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${PLI}={|E}[{sgn}(\Delta {\varphi }_{(t)})]|$$\end{document}PLI=∣E[sgn(Δφ(t))]∣\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${sgn}\left(x\right)=\left\{\begin{array}{c}1,x > 0\\ \hfill 0,x=0\,\\ -1,x < 0\,\,\,\end{array}\right.$$\end{document}sgnx=1,x>00,x=0−1,x<0where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {\varphi }_{(t)}$$\end{document}Δφ(t) is the instantaneous phase difference between the signals of electrode XX and electrode YY at time tt, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathrm{sgn}}}(x)$$\end{document}sgn(x) is the sign function. In this study, functional connectivity matrices were computed individually for each subject. Graph theory analysis and statistical tests were then performed at the subject level. Visualization utilized subject-level averages. Region of interest In this study, we selected electrodes of interest as nodes in the brain network connectivity map. Based on existing research44,45, the frontal and temporal lobes play a significant role in the brain’s processing of natural continuous speech. Therefore, we selected 16 channels (FP1, FP2, F7, F8, F3, F4, T7, T8, P7, P8, C3, C4, P3, P4, O1, O2; Supplementary Fig. 2) as regions of interest. These 16 nodes were then grouped into five regions based on their corresponding locations: the frontal lobe (FP1, FP2, F7, F8, F3, F4), the central region (C3, C4), the parietal lobe (P3, P4), the temporal lobe (T7, T8, P7, P8), and the occipital lobe (O1, O2). Key difference subnetwork extraction To summarize the connectivity pattern most strongly affected by cognitive load, we derived a key difference subnetwork from the significant connectivity results. Edges showing a significant within-participant HCL–LCL difference in permutation testing (p < 0.05) were first retained to form a significant difference network, with edge weights defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta w{{\mathrm{PLI}}}=w{{{\mathrm{PLI}}}}_{{{\mathrm{HCL}}}}-w{{{\mathrm{PLI}}}}_{{{\mathrm{LCL}}}}$$\end{document}ΔwPLI=wPLIHCL−wPLILCL. The largest connected component (LCC) of this network was then extracted. Within the LCC, nodes were ranked by weighted betweenness centrality, and the top 15 nodes were retained together with the significant edges among them to form a candidate backbone subnetwork. If this subnetwork contained fewer than 10 edges, additional edges from the original LCC were added back in descending order of absolute effect size until the minimum edge count was reached. The final subnetwork was used as a compact descriptive representation of the most salient condition-related connectivity changes and was visualized on the standard fsaverage cortical surface. Network topology analysis Graph-theoretic analysis was performed on the hard-thresholded binary delta-band PLI network. The threshold is the PLI value to retain the top 30% connections under the two conditions. For each participant and condition, the thresholded connectivity matrix was represented as a symmetric binary adjacency matrix A=A= \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[{a}_{{ij}}]$$\end{document}[aij], where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${a}_{{ij}}=$$\end{document}aij= 11 indicated a surviving connection between nodes ii and jj, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${a}_{{ij}}$$\end{document}aij =0=0 otherwise. The following metrics were computed: edge count,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E=\frac{1}{2}{\sum }_{i=1}^{N}{\sum }_{j=1}^{N}{a}_{{ij}};\,$$\end{document}E=12∑i=1N∑j=1Naij;density,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D=\frac{2E}{N\left(N-1\right)};$$\end{document}D=2ENN−1;mean clustering coefficient,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C=\frac{1}{N}{\sum }_{i=1}^{N}{C}_{i},\,{C}_{i}=\left\{\begin{array}{c}\frac{2{e}_{i}}{{k}_{i}\left({k}_{i}-1\right)},\,{k}_{i}\ge 2\\ 0,\,{k}_{i} < 2\end{array}\right.;$$\end{document}C=1N∑i=1NCi,Ci=2eikiki−1,ki≥20,ki<2;characteristic path length\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L=\frac{1}{M}{\sum }_{i\ne j,\,{d}_{{ij}} < \infty }{d}_{{ij}};$$\end{document}L=1M∑i≠j,dij<∞dij;global efficiency,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E}_{{glob}}=\frac{1}{N\left(N-1\right)}{\sum }_{i\ne j}\frac{1}{{d}_{{ij}}};$$\end{document}Eglob=1NN−1∑i≠j1dij;and small-worldness index sigma,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma =\frac{C/{C}_{{rand}}}{L/{L}_{{rand}}}.$$\end{document}σ=C/CrandL/Lrand.Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{i}$$\end{document}ki is the degree of node ii, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${e}_{i}$$\end{document}ei is the number of edges among its neighbors, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}_{{ij}}$$\end{document}dij is the shortest path length between nodes ii and jj. Cr and) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{{{{\rm{rand}}}}}$$\end{document}Lrand were obtained from matched random reference networks. These metrics quantified network connection extent, local clustering, global integration, and small-world organization. Statistics and reproducibility All statistical analyses were conducted at the participant level. The final sample included 29 participants after exclusion of two participants, one because of poor behavioral performance in the mental arithmetic task and one because of excessive EEG noise after ICA-based preprocessing. Statistical analyses were conducted using paired within-participant comparisons between the LCL and HCL conditions. Effective sample sizes are reported separately for each analysis because the number of valid paired observations differed across behavioral, encoding, decoding, TRF-peak, and graph-theoretic analyses after preprocessing and metric-specific exclusions. Subjective workload scores, TRF-based encoding accuracy, ANN-based decoding accuracy, and graph-theoretic metrics were compared between conditions using two-sided paired-sample t-tests. Effect sizes for paired t-tests are reported as Cohen’s d, and 95% confidence intervals are reported for the mean LCL–HCL difference when available. For delta-band TRF peak measures, the normality of paired differences was assessed before inferential testing. Peak value was analyzed using a paired-sample t-test, whereas peak latency. Effect sizes for Wilcoxon signed-rank tests are reported as rank-biserial correlation. Edge-level PLI differences were assessed using permutation testing (n = 5000), and graph-theoretic metrics were calculated separately for each valid participant and condition. Statistical significance was evaluated at p < 0.05, and exact p values are reported where possible. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplemental Material Description of Additional Supplementary Files Supplementary Data 1 Reporting Summary Transparent Peer Review file Supplemental Material Description of Additional Supplementary Files Supplementary Data 1 Reporting Summary Transparent Peer Review file