Introduction

Human cognition and behavior is coordinated by a complex network of interconnected brain regions (Sporns and Betzel, 2016; Bassett and Sporns, 2017). A large body of literature has described these networks leveraging functional connectivity (FC) across regions during rest (Biswal et al., 1995; Fox et al., 2005; Greicius et al., 2009) as well as cognitive processing (Friston et al., 1993; Cole et al., 2014, 2021). Resting state data has identified reproducible intrinsic networks of functionally connected regions, such as the default mode, dorsal attention, and sensorimotor network, that resemble FC patterns observed during tasks (Biswal et al., 1995; Fox and Raichle, 2007; Smith et al., 2009). While task-based FC studies are able to reproduce intrinsic network, they also provide evidence for network reconfiguration during changing cognitive demand (Cole et al., 2014; Shine et al., 2016; Hearne et al., 2017).

One hallmark of human behavior is our ability to produce complex and rapid speech utterances.

However, there are only a select few studies examining functional connectivity during the production of speech and language (Fuertinger et al., 2015; Riecker et al., 2005; Geranmayeh et al., 2014; Ewald et al., 2012). This is mainly due to the severe motor artifacts introduced by the jaw, tongue, and laryngeal movements during speech that are inherent to noninvasive neuroimaging and electrophysiological recordings (Grabski et al., 2012; Abbasi et al., 2021). As a result, the network principles that support the rapid sensorimotor transformations of speech are still poorly understood. However, human invasive electrocorticographic (ECoG) recordings circumvent these motor artifacts. Intracranial signals are obtained during neurosurgical procedures directly from the depth and surface of the brain providing a high spatial resolution and millisecond temporal precision. The vast majority of previous intracranial studies have focused on local high-gamma neural activity as a proxy for population firing (Mesgarani et al., 2014; Pasley et al., 2012; Edwards et al., 2010; Flinker et al., 2015; Chang et al., 2010; Bouchard et al., 2013; Mugler et al., 2014; Cheung et al., 2016). These studies have profoundly advanced our understanding of how local neural activity encodes articulatory gestures and phonetic features (Bouchard et al., 2013; Mesgarani et al., 2014; Chang et al., 2010; Mugler et al., 2014; Cheung et al., 2016), and how speech production sequences involve distinct temporal activation patterns across language-relevant cortical regions (Edwards et al., 2010; Flinker et al., 2015). However, the dynamics of inter-regional connectivity during speech production tasks remain largely unexplored.

Measuring functional connectivity in ECoG typically involves assessing synchronous cofluctuations in broadband amplitude, or within specific frequency bands, most often high-gamma (i.e. 70-150 Hz frequency band), using pairwise correlation or coherence (Axmacher et al., 2010; Kucyi et al., 2018). Analyses of long-duration resting-state recordings have revealed stable correlation templates that closely mirror canonical fMRI networks (Kramer et al., 2011; Kucyi et al., 2018). During overt speech, high-gamma coherence between inferior frontal gyrus and superior temporal cortex increases selectively during active vocalization relative to passive listening (Kingyon et al., 2015), underscoring the value of correlation-based FC for probing rapid sensorimotor integration. Yet most FC studies in ECoG still rely on long analysis windows or condition-averaged data. This practice can mask the transient network dynamics that support rapid behaviors such as speech. Undirected FC has also been rarely applied to single-trial decoding of cognitive stages, so low-amplitude but behaviorally relevant connectivity patterns are often lost in averaging. As a result, we still do not understand which inter-regional connectivity patterns are specific to different successive stages of speech perception and production, or whether these interactions reveal network structure invisible to analyses that focus only on local activity.

Here, we directly address these gaps by examining single-trial functional connectivity dynamics in ECoG recordings from a cohort of 42 participants performing speech-production tasks. We employ a robust machine learning approach—support vector machine (SVM) classification—to decode cognitive states from trial-specific FC patterns, and we translate the resulting weights into interpretable connectivity signatures using forward model transformations (Haufe et al., 2014). We aim to address several main questions: (i) Are the cortical networks supporting speech perception and production decodable based on connectivity patterns alone? And (ii) are the associated discriminative connections specific to regions with high neural activity, or do they include network structures obscured by analyses focusing solely on activity?

Results

Single-Trial Functional Connectivity

To investigate the dynamics of single trial functional connectivity during speech perception and production, we collected electrocorticography (ECoG) data, which provide high spatiotemporal resolution. We recorded neural activity from 42 participants while performing three word production tasks (see Experimental Tasks and Cognitive States, Figure 1A and B) and defined five cognitive states of interest (Figure 1C): Auditory perception, picture perception, reading perception, speech production, and baseline (production and baseline are shared across tasks).

Figure 1.

Analyzing functional connectivity (FC) at the single-trial level can reveal nuanced dynamics, including weak yet consistent interactions, that may be obscured during the common practice of trial-averaging (He et al., 2008; Solomon et al., 2018; Huang et al., 2023). We segmented all singletrial neural activity into state-specific intervals and computed FC for each (see Single-Trial Functional Connectivity). In order to provide a view of single-trial connectivity, we create “connectivity vectors” comprised of the upper-triangular elements of the FC matrices (Figure 2A, left). The entries of these vectors represent all connections between pairs of cortical electrodes. Grouping the connectivity vectors by cognitive state provides a comprehensive view of the connectivity patterns. We illustrate this for one participant, revealing substantial trial-to-trial variability (Figure 2A, right). Some states (e.g. picture and reading perception) show obviously similar frontal–frontal and occipital–frontal connections. However, we were interested in identifying patterns of connectivity that are distinct across states.

Figure 2.

In order to test whether these patterns contained sufficient information to differentiate cognitive states we employed a decoding approach. We trained a multi-class Support Vector Machine (SVM) classifier on the connectivity vectors in a representative participant, achieving a mean cross-validated accuracy of 74% (chance level is 20%, Figure 2B). The performance of the model provides evidence for robust state-specific connectivity patterns in this participant. In order to identify these state-specific connections we examined the model weights. We assessed significance of weights after following common practices of transformation (Haufe et al., 2014), to ensure unbiased interpretability (see Decoding Framework). This identified specific networks driving classification performance, for example intra-auditory-cortex connections during auditory perception (Figure 2C). We call these discriminative connections, providing a distinction for connectivity patterns that significantly differentiate cognitive states.

State-Specific Connectivity Patterns Across Participants

To test whether the state-specific connectivity patterns observed in the representative participant generalize to the entire cohort, we replicated our decoding pipeline within each of the N=42 participants. Across all participants, the classifiers performed robustly (Figure 3A): perceptual states were decoded with 71.5–74.4% accuracy, speech and baseline with 56.7–59.1%. Participant-wise model accuracies over all states ranged from 52% to 86% (mean = 64.4%; Figure 3B), all well above the 20% chance level for five classes. We then transformed the SVM weights to identify each state’s discriminative connections (Figure 3D) revealing that each state is represented by a small subset of connections (mean across states <≈ 0.002% of all possible connections) largely unique to that state (uniqueness ratio = 0.817; Figure 3C, see Statistical Analyses). Spatially, auditory perception relied on STG–frontal links, visual states on occipital–frontal connections, and speech production on enhanced speech-motor cortex coupling as well as motor–auditory anti-correlations (Figure 3D and Figure 5B). In contrast, a traditional analysis of strongest connections (i.e. correlation magnitude thresholding; see Single-Trial Functional Connectivity) yielded denser, overlapping connections (uniqueness = 0.27; Figure 3C and E). These results show that single-trial FC decoding isolates a highly specific connectivity signature for each cognitive state, highlighting functionally meaningful interactions that are obscured when connectivity is addressed solely by correlation magnitude.

Figure 3.

Relationship between discriminative connectivity and local neural activity

Speech production is notoriously difficult to study with non-invasive measures given the inherent motor artifacts. Conversely, intracranial studies have shown great promise in characterizing speech production dynamics, however an overwhelming number of studies have focused on local neural activity alone. We were interested in characterizing the unique contribution of our approach compared with the common practice of analyzing local high-gamma neural activity. During speech production, high-gamma active electrodes (see Statistical Analyses) on the cortex were predominantly distributed over the superior temporal and peri-central cortices (Figure 4A and E). This highly selective distribution has been a hallmark of local activity and is in stark contrast to the widespread distribution of electrodes across cortex that show a strong functional correlation (Figure 4B,D,E). Our decoding approach provides evidence for a more selective set of discriminative functional connections with broad coverage of the language network (Figure 4C,D,E). Our approach reveals a larger ratio of electrodes identified across non-sensory (and non-motor) cortices compared with local activity (permutation test, see Statistical Analyses), notably the inferior temporal (p = 0.0319), superior frontal (p = 0.0114), parietal (p = 0.0005), occipital (p = 0.0002), and fusiform cortices (p = 0.0279). This indicates that local activity alone may be insufficient for identifying important regions in the language network and a connectivity-based approach is necessary as a complement. While our approach elucidates discriminative connections, we were interested in understanding the relationship between the degree of local activity and discriminative connectivity. To quantify this, we extracted the median correlation for all discriminative connections and the mean neural activity of their respective nodes (for every cognitive state; see Neural Activity versus Connectivity Analysis). The resulting connectivity values span both positive and negative correlations (Figure 5A), yet organize into clear, state-specific patterns when projected onto the cortex (Figure 5B). Node activity shows a similar broad distribution (Figure 5C) with a clear spatial gradient from low to high activity regions (Figure 5D). A natural question is whether nodes with higher neural activity contribute more to functional connectivity. To test this, we correlated the mean connectivity of each node in discriminative connections with its mean activity for every cognitive state and found significant correlations across states (Spearman Correlations, Auditory rho = 0.4472, p < 0.0001; Picture rho = 0.2925, p = 0.0003; Reading rho = 0.4306, p < 0.0001; Speech rho = 0.1195, p = 0.0254.). When visualizing the relationship (Figure 5E, color-coded by cognitive state), most nodes align along the diagonal, indicating balanced activity and connectivity. More active regions tend to exhibit stronger correlations. However, many electrodes diverge from this relationship: some are highly connected despite low neural activity, whereas others are strongly active but weakly connected. We quantified this imbalance with a connectivity–activity relative-distance index (RDI) and mapped it on the cortex (Figure 5F). The RDI highlights precentral and frontal gyri during auditory perception and speech-motor cortex during picture and reading perception, where connectivity exceeds activity. Conversely, during speech production discriminative nodes include both positive RDI (i.e. low activity and high connectivity) and negative RDI sites (i.e. high activity and low connectivity), particularly within the superior temporal, inferior frontal and precentral gyri. Together, these findings show that discriminative connectivity results in a unique, functionally meaningful network signature that is invisible to activity-only or correlation-magnitude-only views and underscores that connectivity adds a complementary dimension to local activity for differentiating cognitive states.

Figure 4.

Figure 5.

Discussion

Our study leverages single trial connectivity decoding across a large cohort of patients (N=42) to investigate functional connectivity. Our approach is able to achieve high decoding across five speech-related cognitive states while providing interpretable weights that highlight the discriminative connections. Compared to conventional methods (i.e. local high-gamma activity), our approach reveals previously obscured cortical sites that discriminate cognitive states. These results are among a select few functional connectivity studies focusing on speech related cortical networks, providing a rare window into the functional networks supporting this function.

Despite valuable insights, most human FC studies in ECoG share design choices that blur fast connectivity dynamics. Many studies estimate connectivity from long windows of >1s (Kucyi et al., 2018; Nir et al., 2008). Investigating connectivity during the transient cognitive states involved in speech perception and production will require a shorter time window, enabling us to capture these state specific characteristics of connectivity. Further, studies often average the signal or connectivity over many repetitions, which eliminates trial-to-trial variability (Casimo et al., 2019; Foster et al., 2015). Averaging signals or correlation matrices can result in robust estimations that highlights the strongest connections. However, this step also blurs the distinction between state specific signals or connections that maintain a steady low value and those that vary drastically. To be able to capture the transient, state-specific connections that are stable across trials regardless of their strength, we compute FC in 500-ms windows and keep every trial separate. This allows us to preserve within-state variability while avoiding the loss of temporal detail. As a result, this framework unveils the fast dynamics of speech perception and production with a resolution that respects both the transient and distributed nature of these networks.

To investigate the discriminative connections underlying speech perception and production, we computed functional connectivity between every pair of electrodes. By leveraging the interpretability of the transformed weights of a linear SVM model, we detect sparse state-specific connections based solely on their ability to differentiate cognitive states (Figure 3D, Figure 5B). As a result, we identify both strong and weak discriminative connections, including ones between regions that fall outside the canonical language areas. In contrast, the standard connectivity magnitude thresholding method results in a widespread and state-agnostic set of connections, masking these distinctions (Figure 3E). A common practice in ECoG-based FC studies that avoids this problem is to restrict analyses to a handful of predefined nodes using seed-based or region-of-interest strategies (Keller et al., 2013). Most often, these seeds are selected based on electrodes with large high-gamma responses or those deemed critical by electrical stimulation mapping (Kingyon et al., 2015; Rolston and Chang, 2018; Fox et al., 2020; Hsieh et al., 2024). In the context of language, seed-based strategies typically target inferior-frontal and auditory cortices, privileging well-established nodes but marginalising others, like middle frontal, precentral, postcentral and temporal pole. While methodologically convenient, this choice significantly narrows the scope of network inference and implicitly reinforces the dominant cortical regions. Moreover, it overlooks potentially informative and functionally relevant regions with lower signal (activity or connectivity), such as those identified by our RDI close to 1 or -1 in Figure 5F. Our results suggest that a more inclusive (e.g. all electrodes) approach coupled with a decoding framework reveals a set of meaningful connections and avoids anatomical blind spots.

Our findings reveal connections across peri-sylvian cortices as well as the temporal pole during shifts from perception to production — a dynamic that has been difficult to capture with previous approaches. While fMRI studies have been instrumental in mapping large-scale networks (Biswal et al., 1995; Smith et al., 2009; Fox et al., 2005; Greicius et al., 2003; Power et al., 2011; Yeo et al., 2011) and in characterizing the language network (Saur et al., 2008; Braga et al., 2020; Tomasi and Volkow, 2012; Branco et al., 2020; Lipkin et al., 2022), most task-based studies focus on listening, reading, or covert speech. This has left overt speech underexplored in fMRI, due to motion artifacts (Birn et al., 1999; Gracco et al., 2005). Moreover, the temporal smoothing of the BOLD response limits fMRI’s ability to resolve the rapid state transitions between perception and production of speech. In contrast, ECoG, with its millisecond resolution, allows us to track these fast dynamics. Our approach captures connectivity patterns directly relevant to speech processes, revealing perception-to-production shifts with millisecond resolution. Specifically, we find positive correlations during perception across frontal and relevant sensory regions, however speech production is associated with peri-rolandic and temporal (STG and ATL) networks. Interestingly, the discriminative connections for speech production included many negative correlations (Figure 5B) between STG and ventral frontal cortices (IFG and vPrG) which likely reflects speech induced suppression during vocalization (Khalilian-Gourtani et al., 2024; Kurteff et al., 2024). Although ECoG coverage is clinically constrained, it complements the whole-brain approach of fMRI by offering a high-resolution view of speech processing dynamics at behaviorally relevant timescales.

Our results further show that discriminative connections emerge in regions with both high and low neural activity. In the ECoG literature, neural activity is most often measured as the amplitude of the high-gamma (70–150 Hz) frequency range, with task-relevant regions mapped by significant increases above baseline. Numerous studies have shown that high-gamma active regions during speech tasks localize strongly to IFG, precentral gyrus, STG, and MTG (Chang et al., 2010; Edwards et al., 2010; Flinker et al., 2015; Moses et al., 2016). This approach maps active electrodes but often overlooks distributed connectivity as well regions that are not significantly active. Our connectivity findings provide evidence for connections involving low-activity electrodes, particularly in parts of precentral and inferior frontal cortices during perceptual states (Figure 5F). The fact that we find these connections in addition to discriminative connections between electrodes with high neural activity, suggests that the functional networks supporting speech extend beyond the most metabolically active regions. The identification of both active-to-active and active-to-silent discriminative connections demonstrates that connectivity analyses provide orthogonal information to local high-gamma measures, potentially revealing preparatory or coordinating regions that studies relying on activation alone may overlook.

Together, these findings highlight how single-trial functional connectivity decoding provides a powerful complement to traditional analyses of neural activity. By uncovering state-specific and sparse connections across regions, we reveal a distributed speech network that cannot be captured by focusing on local activation alone or seed-based connectivity. The ability to decode cognitive states from transient connectivity patterns underscores the dynamic and flexible nature of the underlying cognitive processes and allows for mapping the full cortical networks involved in speech perception and production.

Methods and Materials

Participants

We used intracranial data from 42 neurosurgical patients undergoing surgical monitoring for Epilepsy (20 females, mean age: 30.38 ± 11.75 yo, 18 right grid, 24 left grid coverage). Patients were selected based on sufficient electrode coverage over temporal, parietal, frontal, and occipital cortices relevant to language processing and their ability to perform the experimental tasks. The electrode arrays included grids and strips placed directly on the cortical surfaces. The implantation and location of the electrodes were guided solely by clinical requirements. All participants provided informed consent following approval from the NYU Langone institutional review board. Surface reconstructions, electrode localization, and Montreal Neurological Institute coordinates were extracted by aligning postoperative brain imaging (MR or CT) to preoperative brain MRI using previously published methods (Yang et al., 2012).

Experimental Tasks and Cognitive States

Participants completed three tasks that elicited production of the same target words from auditory or visual stimuli (Figure 1A). The tasks were auditory word repetition (repeat an auditorily presented word), visual word reading (read a visually presented written word), and picture naming (name a word from a visually presented line drawing). Each task comprised 100 trials drawn from an identical set of 50 words that appeared twice across different modalities. For each task and trial, the stimulus was randomly presented. Each trial began with a silent baseline of at least 500ms, followed by stimulus presentation, and then speech production (Figure 1C). Participants are instructed to produce the word as soon as they are ready without a go signal (Shum et al., 2020; Yu et al., 2025). We defined five cognitive states using 500 ms windows time-locked to event markers: auditory perception (auditory stimulus onset), picture perception (picture stimulus onset), reading perception (written word stimulus onset), speech production (speech onset; shared across tasks), and baseline (pre-stimulus presentation period, also shared across tasks).

ECoG Data Acquisition and Preprocessing

ECoG neural signals were recorded at 2048 Hz with the Neuroworks Quantum Amplifier (Natus Biomedical, Appleton, WI) and were downsampled to 512 Hz during export. Electrodes exhibiting seizure activity or excessive noise identified clinically were excluded from further analysis. Data were re-referenced using a common average referencing scheme to mitigate common-mode noise. We band-passed signals into eight logarithmically spaced sub-bands from 70–150 Hz and averaged them to isolate high-gamma activity. We adopted this multi-band extraction to avoid dominance of lower frequencies (Hamilton et al., 2021; Moses et al., 2016). We computed analytic amplitude as the absolute value of the Hilbert transform of the filtered signal in each band. We then normalized the data (i.e., z-score) and averaged across bands.

Single-Trial Functional Connectivity

For computing functional connectivity, we segmented the ECoG high-gamma signals into 500 ms epochs for each trial and cognitive state, aligned to task event markers described above. For every epoch, we computed pairwise Pearson correlations of high-gamma analytic amplitude between all electrodes, producing a symmetric functional connectivity matrix per trial and state. To avoid redundancy, we vectorized the upper-triangular off-diagonal of each matrix to obtain a trial-specific connectivity vector for each state. Each vector contains the correlation for all possible electrode pairs (features), with the sign preserved.

In the analysis to identify the strongest connections (correlation magnitude-thresholding, Figure 3E), we first summarized each state’s connectivity matrix by computing the median value of each connection across trials. We then evaluated how much each connection’s median deviated from the overall distribution of medians. Connections that showed significantly high or low values (beyond what we would expect by chance, after accounting for multiple comparisons, α< 0.05) were retained. All others were set to zero. After this thresholding step, we reconstructed the participant- and state-specific connectivity matrices using only the significant connections. Finally, we combined these matrices into a block-diagonal group matrix, preserving individual participant structure while enabling pooled network analysis.

Decoding Framework

We classified trial-specific connectivity vectors into cognitive states using a linear Support Vector Machine (SVM) in MATLAB (MathWorks). To address class imbalance, we applied the Synthetic Minority Oversampling Technique (SMOTE, Chawla et al., 2002) within each participant and training fold to avoid leakage. We evaluated performance with stratified five-fold cross-validation per participant, averaged accuracy across folds, and derived confusion matrices. We estimated chance levels as theoretical chance (20% for five states, after accounting for SMOTE) and verified this estimate with a permutation approach (i.e. shuffling state labels 10000× within participant). To identify discriminative connections in Figure 2 and Figure 3D, we transformed SVM weights from the decision boundary hyperplanes using the Haufe forward model (Haufe et al., 2014). This transformation enables interpretation of classifier weights by accounting for feature dependencies and mapping them back to neural sources. We normalized the resulting weights using their L2 norm, and then evaluated how far each transformed weight deviated from the median of the distribution, using the median absolute deviation (MAD) as a robust scale estimator. We set to zero all weights whose deviation from the median did not exceed a threshold derived from a Laplacian tail bound (α< 0.05). We selected this threshold to reflect the heavy-tailed distribution of weights. We defined discriminative connections as those whose weights exceeded this threshold, indicating they contributed most strongly to the separation of cognitive states in the classifier. To quantify the specificity of the identified connections for each state, we computed a uniqueness ratio. This ratio represents the proportion of connections uniquely associated with a given cognitive state relative to all connections identified for that state.

Neural Activity versus Connectivity Analysis

For the analysis in Figure 5, we focused on electrodes that participated in the discriminative connections. For each such electrode and state, we quantified local activity as the mean high-gamma analytical amplitude across trials and time points. For each discriminative connection, we quantified connectivity strength as the median Pearson’s correlation across trials of that state. We then computed a connectivity–activity relative distance index (RDI) at the electrode level:

Where A denotes average local activity and C denotes average connectivity. RDI ranges from -1 (high activity, low connectivity) to +1 (low activity, high connectivity), with values near zero indicating balance. We projected RDI onto each participant’s cortical surface to visualize its spatial distribution.

Statistical Analyses

For assessing the significant weights and connections (i.e. Decoding Framework and Single-Trial Functional Connectivity, respectively), We performed two-sided hypothesis tests at the single participant level. To control the family-wise error rate (FWER) across multiple comparisons, we applied a Bonferroni correction, maintaining an FWER of α = 0.05. For active electrode selection, we compared the high-gamma neural activity of each trial averaged over time, with the distribution of high-gamma neural activity during baseline, and used a one-sided threshold with α = 0.05 of the baseline to label an electrode as active on that trial.

To compute the significance of the difference between uniqueness ratios (Figure 3C), we performed permutation test, randomly swapping the connections identified across the two methods (for each connection, values are 1 for significant/discriminative connection and 0 otherwise, then swapped between methods with probability 0.5), with 10,000 permutations. We controlled for multiple comparisons using the false-discovery rate for a p < 0.05 two-sided significance (Benjamini and Hochberg, 1995).

To identify brain regions where the two methods yielded different results (Figure 4D), we compared the fraction of electrodes identified by the discriminative FC method against the fraction identified by the neural signal-only method. For each state s and region r, we defined the fraction(ratio) as the number of identified electrodes in r divided by the total electrodes recorded in r. This was statistically compared to a null hypothesis distribution by randomly permuting region labels across electrodes (preserving each region’s total electrode count) while holding fixed which electrodes were identified by that method, recomputing the ratio difference in each region for 10,000 permutations. Two-sided p-values used the permutation tail; For each region and state, we controlled for multiple comparisons using the false-discovery rate (Benjamini and Hochberg (1995), α = 0.05).

Supplementary Information

Supplementary Figure 1.

Supplementary Figure 2.

Supplementary Figure 3.

Data availability

All code will be available under the Flinker Lab GitHub, and data will be available upon request.

Acknowledgements

This work was supported by National Institutes of Health grants R01NS109367, R01NS115929, and R01DC018805 (A.F.) and National Science Foundationshum IIS-2309057.

Additional information

Funding

HHS | NIH | National Institute of Neurological Disorders and Stroke (NINDS) (R01NS109367)

• Adeen Flinker

HHS | NIH | National Institute of Neurological Disorders and Stroke (NINDS) (R01NS115929)

• Adeen Flinker

HHS | NIH | National Institute on Deafness and Other Communication Disorders (NIDCD) (R01DC018805)

• Adeen Flinker

National Science Foundation (NSF) (IIS-2309057)

• Adeen Flinker