Author response:
The following is the authors’ response to the original reviews
Public Reviews:
Reviewer #1 (Public Review):
Summary:
This paper reports an intracranial SEEG study of speech coordination, where participants synchronize their speech output with a virtual partner that is designed to vary its synchronization behavior. This allows the authors to identify electrodes throughout the left hemisphere of the brain that have activity (both power and phase) that correlates with the degree of synchronization behavior. They find that high-frequency activity in the secondary auditory cortex (superior temporal gyrus) is correlated to synchronization, in contrast to primary auditory regions. Furthermore, activity in the inferior frontal gyrus shows a significant phase-amplitude coupling relationship that is interpreted as compensation for deviation from synchronized behavior with the virtual partner.
Strengths:
(1) The development of a virtual partner model trained for each individual participant, which can dynamically vary its synchronization to the participant's behavior in real-time, is novel and exciting.
(2) Understanding real-time temporal coordination for behaviors like speech is a critical and understudied area.
(3) The use of SEEG provides the spatial and temporal resolution necessary to address the complex dynamics associated with the behavior.
(4) The paper provides some results that suggest a role for regions like IFG and STG in the dynamic temporal coordination of behavior both within an individual speaker and across speakers performing a coordination task.
We thank the Reviewer for their positive comments on our manuscript.
Weaknesses:
(1) The main weakness of the paper is that the results are presented in a largely descriptive and vague manner. For instance, while the interpretation of predictive coding and error correction is interesting, it is not clear how the experimental design or analyses specifically support such a model, or how they differentiate that model from the alternatives. It's possible that some greater specificity could be achieved by a more detailed examination of this rich dataset, for example by characterizing the specific phase relationships (e.g., positive vs negative lags) in areas that show correlations with synchronization behavior. However, as written, it is difficult to understand what these results tell us about how coordination behavior arises.
We understand the reviewer’s comment. It is true that this work, being the first in the field using real-time adapting synchronous speech and intracerebral neural data, is a descriptive work, that hopefully will pave the way for further studies. We have now added more statistical analyses (see point 2) to go beyond a descriptive approach and we have also rewritten the discussion to clarify how this work can possibly contribute to disentangle different models of language interaction. Most importantly we have also run new analyses taking into account the specific phase relationship, as suggested.
We already had an analysis using instantaneous phase difference in the phase-amplitude coupling approach, that bridges phase of behaviour to neural responses (amplitude in the high-frequency range). However, this analysis, as the reviewer noted, does not distinguish between positive and negative lags, but rather uses the continuous fluctuations of coordinative behaviour. Following the reviewer’s suggestion, we have now run a new analysis estimating the average delay (between virtual partner speech and patient speech) in each trial, using a cross-correlation approach. This gives a distribution of delays across trials that can then be “binned” as positive or negative. We have thus rerun the phase-amplitude coupling analyses on positive and negative trials separately, to assess whether the phase amplitude relationship depends upon the anticipatory (negative lags) or compensatory (positive lags) behaviour. Our new analysis (now in the supplementary, see figure below) does not reveal significant differences between positive and negative lags. This lack of difference, although not easy to interpret, is nonetheless interesting because it seems to show that the IFG does not have a stronger coupling for anticipatory trials. Rather the IFG seems to be strongly involved in adjusting behaviour, minimizing the error, independently of whether this is early or late.
We have updated the “Coupling behavioural and neurophysiological data” section in Materials and methods as follows:
“In the third approach, we assessed whether the phase-amplitude relationship (or coupling) depends upon the anticipatory (negative delays) or compensatory (positive delays) behaviour between the VO and the patients’ speech. We computed the average delay in each trial using a cross-correlation approach on speech signals (between patient and VP) with the MATLAB function xcorr. A median split (patient-specific ; average median split = 0ms, average sd = 24ms) was applied to conserve a sufficient amount of data, classifying trials below the median as “anticipatory behaviour” and trials above the median as “compensatory behaviour”. Then we conducted the phase-amplitude coupling analyses on positive and negative trials separately.”
We also added a paragraph on this finding in the Discussion:
“Our results highlight the involvement of the inferior frontal gyrus (IFG) bilaterally, in particular the BA44 region, in speech coordination. First, trials with a weak verbal coordination (VCI) are accompanied by more prominent high frequency activity (HFa, Fig.4; Fig.S4). Second, when considering the within-trial time-resolved dynamics, the phase-amplitude coupling (PAC) reveals a tight relation between the low frequency behavioural dynamics (phase) and the modulation of high-frequency neural activity (amplitude, Fig.5B ; Fig.S5). This relation is strongest when considering the phase adjustments rather than the phase of speech of the VP per se : larger deviations in verbal coordination are accompanied by increase in HFa. Additionally, we also tested for potential effects of different asynchronies (i.e., temporal delay) between the participant's speech and that of the virtual partner but found no significant differences (Fig.S6). While lack of delay-effect does not permit to conclude about the sensitivity of BA44 to absolute timing of the partner’s speech, its neural dynamics are linked to the ongoing process of resolving phase deviations and maintaining synchrony.”
(2) In the results section, there's a general lack of quantification. While some of the statistics reported in the figures are helpful, there are also claims that are stated without any statistical test. For example, in the paragraph starting on line 342, it is claimed that there is an inverse relationship between rho-value and frequency band, "possibly due to the reversed desynchronization/synchronization process in low and high frequency bands". Based on Figure 3, the first part of this statement appears to be true qualitatively, but is not quantified, and is therefore impossible to assess in relation to the second part of the claim. Similarly, the next paragraph on line 348 describes optimal clustering, but statistics of the clustering algorithm and silhouette metric are not provided. More importantly, it's not entirely clear what is being clustered - is the point to identify activity patterns that are similar within/across brain regions? Or to interpret the meaning of the specific patterns? If the latter, this is not explained or explored in the paper.
The reviewer is right. We have now added statistical analyses showing that:
(1) the ratio between synchronization and desynchronization evolves across frequencies (as often reported in the literature).
(2) the sign of rho values also evolves across frequencies.
(3) the clustering does indeed differ when taking into account behaviour. We have also clarified the use of clustering and the reasoning behind it.
We have updated the Materials and methods section as follows:
“The statistical difference between spatial clustering in global effect and brain-behaviour correlation was estimated with linear model using the R function lm (stat package), post-hoc comparisons were corrected for multiple comparisons using the Tukey test (lsmeans R package ; Lenth, 2016). The statistical difference between clustering in global effect and behaviour correlation across the number of clusters was estimated using permutation tests (N=1000) by computing the silhouette score difference between the two conditions.” We have updated the Results section as follows:
(1) “This modulation between synchronization and desynchronization across frequencies was significant (F(5) = 6.42, p < .001 ; estimated with linear model using the R function lm).”
(2) “The first observation is a gradual transition in the direction of correlations as we move up frequency bands, from positive correlations at low frequencies to negative ones at high frequencies (F(5) = 2.68, p = .02). This effect, present in both hemispheres, mimics the reversed desynchronization/synchronization process in low and high frequency bands reported above.”
(3) “Importantly, compared to the global activity (task vs rest, Fig 3A), the neural spatial profile of the behaviour-related activity (Fig 3B) is more clustered, in the left hemisphere. Indeed, silhouette scores are systematically higher for behaviour-related activity compared to global activity, indicating greater clustering consistency across frequency bands (t(106) = 7.79, p < .001, see Figure S3). Moreover, silhouette scores are maximal, in particular for HFa, for five clusters (p < .001), located in the IFG BA44, the IPL BA 40 and the STG BA 41/42 and BA22 (see Figure S3).”
(3) Given the design of the stimuli, it would be useful to know more about how coordination relates to specific speech units. The authors focus on the syllabic level, which is understandable. But as far as the results relate to speech planning (an explicit point in the paper), the claims could be strengthened by determining whether the coordination signal (whether error correction or otherwise) is specifically timed to e.g., the consonant vs the vowel. If the mechanism is a phase reset, does it tend to occur on one part of the syllable?
Thank you for this thoughtful feedback. We agree that the relationship between speech coordination and specific speech units, such as consonants versus vowels, is an intriguing question. However, in our study, both interlocutors (the participant and the virtual partner) are adapting their speech production in real-time. This interactive coordination makes it difficult to isolate neural signatures corresponding to precise segments like consonants or vowels, as the adjustments occur in a continuous and dynamic context.
The VP's ability to adapt depends on its sensitivity to spectral cues, such as the transition from one phonetic element to another. This is likely influenced by the type of articulation, with certain transitions being more salient (e.g., between a stop consonant like "p" and a vowel like "a") and others being less distinct (e.g., between nasal consonants like "m" and a vowel). Thus, the VP’s spectral adaptation tends to occur at these transitions, which are more prominent in some cases than in others.
For the participants, previous studies have shown a greater sensitivity during the production of stressed vowels (Oschkinat & Hoole, 2022; Li & Lancia, 2024), which may reflect a heightened attentional or motor adjustment to stressed syllables.
Here, we did not specifically address the question of coordination at the level of individual linguistic units. Moreover, even if we attempted to focus on this level, it would be challenging to relate neural dynamics directly to specific speech segments. The question of how synchronization at the level of individual linguistic units might relate to neural data is complex. The lack of clear, unit-specific predictions makes it difficult to parse out distinct neural signatures tied to individual segments, particularly when both interlocutors are continuously adjusting their speech in relation to one another.
Therefore, while we recognize the potential importance of examining synchronization at the level of individual phonetic elements, the design of our task and the nature of the coordination in this interactive context (realtime bidirection adaptation) led us to focus more broadly on the overall dynamics of speech synchronization at the syllabic level, rather than on specific linguistic units.
We now state at the end of the Discussion section:
“It is worth noting that the influence of specific speech units, such as consonants versus vowels, on speech coordination remains to be explored. In non-interactive contexts, participants show greater sensitivity during the production of stressed vowels, possibly reflecting heightened attentional or motor adjustments (Oschkinat & Hoole, 2022; Li & Lancia, 2024). In this study, the VP’s adaptation relies on sensitivity to spectral cues, particularly phonetic transitions, with some (e.g., formant transitions) being more salient than others. However, how these effects manifest in an interactive setting remains an open question, as both interlocutors continuously adjust their speech in real time. Future studies could investigate whether coordination signals, such as phase resets, preferentially align with specific parts of the syllable.” References cited:
– Oschkinat, M., & Hoole, P. (2022). Reactive feedback control and adaptation to perturbed speech timing in stressed and unstressed syllables. Journal of Phonetics, 91, 101133.
– Li, J., & Lancia, L. (2024). A multimodal approach to study the nature of coordinative patterns underlying speech rhythm. In Proc. Interspeech, 397-401.
(4) In the discussion the results are related to a previously-described speech-induced suppression effect. However, it's not clear what the current results have to do with SIS, since the speaker's own voice is present and predictable from the forward model on every trial. Statements such as "Moreover, when the two speech signals come close enough in time, the patient possibly perceives them as its own voice" are highly speculative and apparently not supported by the data.
We thank the reviewer for raising thoughtful concerns about our interpretation of the observed neural suppression as related to speaker-induced suppression (SIS). We agree that our study lacks a passive listening condition, which limits direct comparisons to the original SIS effect, traditionally defined as the suppression of neural responses to self-produced speech compared to externally-generated speech (Meekings & Scott, 2021).
In response, we have reconsidered our terminology and interpretation. In the revised Discussion section, we refer to our findings as a "SIS-related phenomenon specific to the synchronous speech context". Unlike classic SIS paradigms, our interactive task involves simultaneous monitoring of self- and externally-generated speech, introducing additional attentional and coordinative demands.
The revised Discussion also incorporates findings by Ozker et al. (2022, 2024), which link SIS and speech monitoring, suggesting that suppressing responses to self-generated speech facilitates error detection. We propose that the decrease in high-frequency activity (HFa) as verbal coordination increases reflects reduced error signals due to closer alignment between perceived and produced speech. Conversely, HFa increases with reduced coordination may signify greater prediction error.
Additionally, we relate our findings to the "rubber voice" effect (Zheng et al., 2011; Lind et al., 2014; Franken et al., 2021), where temporally and phonetically congruent external speech can be perceived as self-generated. We speculate that this may occur in synchronous speech tasks when the participant's and VP's speech signals closely align. However, this interpretation remains speculative, as no subjective reports were collected to confirm this perception. Future studies could include participant questionnaires to validate this effect and relate subjective experience to neural measures of synchronization.
Overall, our findings extend the study of SIS to dynamic, interactive contexts and contribute to understanding internal forward models of speech production in more naturalistic scenarios.
We have now added these points to the discussion as follows:
“The observed negative correlation between verbal coordination and high-frequency activity (HFa) in STG BA22 suggests a suppression of neural responses as the degree of behavioural synchrony increases. This result is reminiscent of findings on speaker-induced suppression (SIS), where neural activity in auditory cortex decreases during self-generated speech compared to externally-generated speech (Meekings & Scott, 2021; Niziolek et al., 2013). However, our paradigm differs from traditional SIS studies in two critical ways: (1) the speaker's own voice is always present and predictable from the forward model, and (2) no passive listening condition was included. Therefore, our findings cannot be directly equated with the original SIS effect.
Instead, we propose that the suppression observed here reflects a SIS-related phenomenon specific to the synchronous speech context. Synchronous speech requires simultaneous monitoring of self- and externallygenerated speech, a task that is both attentionally demanding and coordinative. This aligns with evidence from Ozker et al. (2024, 2022), showing that the same neural populations in STG exhibit SIS and heightened responses to feedback perturbations. These findings suggest that SIS and speech monitoring are related processes, where suppressing responses to self-generated speech facilitates error detection. In our study, suppression of HFa as coordination increases may reflect reduced prediction errors due to closer alignment between perceived and produced speech signals. Conversely, increased HFa during poor coordination may signify greater mismatch, consistent with prediction error theories (Houde & Nagarajan, 2011; Friston et al., 2020). Furthermore, when self- and externally-generated speech signals are temporally and phonetically congruent, participants may perceive external speech as their own. This echoes the "rubber voice" effect, where external speech resembling self-produced feedback is perceived as self-generated (Zheng et al., 2011; Lind et al., 2014; Franken et al., 2021). While this interpretation remains speculative, future studies could incorporate subjective reports to investigate this phenomenon in more detail.” References cited:
– Franken, M. K., Hartsuiker, R. J., Johansson, P., Hall, L., & Lind, A. (2021). Speaking With an Alien Voice: Flexible Sense of Agency During Vocal Production. Journal of Experimental Psychology-Human perception and performance, 47(4), 479-494. https://doi.org/10.1037/xhp0000799
– Houde, J. F., & Nagarajan, S. S. (2011). Speech production as state feedback control. Frontiers in human neuroscience, 5, 82.
– Lind, A., Hall, L., Breidegard, B., Balkenius, C., & Johansson, P. (2014). Speakers' acceptance of real-time speech exchange indicates that we use auditory feedback to specify the meaning of what we say. Psychological Science, 25(6), 1198-1205. https://doi.org/10.1177/0956797614529797
– Meekings, S., & Scott, S. K. (2021). Error in the Superior Temporal Gyrus? A Systematic Review and Activation Likelihood Estimation Meta-Analysis of Speech Production Studies. Journal of Cognitive Neuroscience, 33(3), 422-444. https://doi.org/10.1162/jocn_a_01661
– Niziolek C. A., Nagarajan S. S., Houde J. F (2013) What does motor efference copy represent? Evidence from speech production Journal of Neuroscience 33:16110–16116Ozker M., Doyle W., Devinsky O., Flinker A (2022) A cortical network processes auditory error signals during human speech production to maintain fluency PLoS Biology 20.
– Ozker, M., Yu, L., Dugan, P., Doyle, W., Friedman, D., Devinsky, O., & Flinker, A. (2024). Speech-induced suppression and vocal feedback sensitivity in human cortex. eLife, 13, RP94198. https://doi.org/10.7554/eLife.94198
– Zheng, Z. Z., MacDonald, E. N., Munhall, K. G., & Johnsrude, I. S. (2011). Perceiving a Stranger's Voice as Being One's Own: A 'Rubber Voice' Illusion? PLOS ONE, 6(4), e18655.
(5) There are some seemingly arbitrary decisions made in the design and analysis that, while likely justified, need to be explained. For example, how were the cutoffs for moderate coupling vs phase-shifted coupling (k ~0.09) determined? This is noted as "rather weak" (line 212), but it's not clear where this comes from. Similarly, the ROI-based analyses are only done on regions "recorded in at least 7 patients" - how was this number chosen? How many electrodes total does this correspond to? Is there heterogeneity within each ROI?
The reviewer is correct, we apologize for this missing information. We now specify that the coupling values were empirically determined on the basis of a pilot experiment in order to induce more or less synchronization, but keeping the phase-shifted coupling at a rather implicit level.
Concerning the definition of coupling as weak, one should consider that, in the Kuramoto model, the strength of coupling (k) is relative to the spread of the natural frequencies (Δω) in the system. In our study, the natural frequencies of syllables range approximately from 2 Hz to 10Hz, resulting in a frequency spread of Δω = 8 Hz. For coupling to strongly synchronize oscillators across such a wide range, k must be comparable to or exceed Δω. Thus, since k = 0.1 is far much smaller than Δω, it is therefore classified as weak coupling.
We have now modified the Materials and methods section as follows:
“More precisely, for a third of the trials the VP had a neutral behaviour (close to zero coupling: k = +/- 0.01). For a third it had a moderate coupling, meaning that the VP synchronised more to the participant speech (k = -0.09). And for the last third of the trials the VP had a moderate coupling but with a phase shift of pi/2, meaning that it moderately aimed to speak in between the participant syllables (k = + 0.09). The coupling values were empirically determined on the basis of a pilot experiment in order to induce more or less synchronization but keeping the phase-shifted coupling at a rather implicit level. In other terms, while participants knew that the VP would adapt, they did not necessarily know in which direction the coupling went.”
Regarding the criterion of including regions recorded in at least 7 patients, our goal was to balance data completeness with statistical power. Given our total sample of 16 patients, this threshold ensures that each included region is represented in at least ~44% of the cohort, reducing the likelihood of spurious findings due to extremely small sample sizes. This choice also aligns with common neurophysiological analysis practices, where a minimum number of subjects (at least 2 in extreme cases) is required to achieve meaningful interindividual comparisons while avoiding excessive data exclusion. Additionally, this threshold maintains a reasonable tradeoff between maximizing patient inclusion and ensuring that statistical tests remain robust.
We have now added more information in the Results section “Spectral profiles in the language network are nuanced by behaviour” on this point as follows:
“To balance data completeness and statistical power, we included only brain regions recorded in at least 7 patients (~44% of the cohort) for the left hemisphere and at least 5 patients for the right hemisphere (~31% of the cohort), ensuring sufficient representation while minimizing biases due to sparse data.”
Reviewer #2 (Public Review):
Summary:
This paper investigates the neural underpinnings of an interactive speech task requiring verbal coordination with another speaker. To achieve this, the authors recorded intracranial brain activity from the left hemisphere in a group of drug-resistant epilepsy patients while they synchronised their speech with a 'virtual partner'. Crucially, the authors were able to manipulate the degree of success of this synchronisation by programming the virtual partner to either actively synchronise or desynchronise their speech with the participant, or else to not vary its speech in response to the participant (making the synchronisation task purely one-way). Using such a paradigm, the authors identified different brain regions that were either more sensitive to the speech of the virtual partner (primary auditory cortex), or more sensitive to the degree of verbal coordination (i.e. synchronisation success) with the virtual partner (secondary auditory cortex and IFG). Such sensitivity was measured by (1) calculating the correlation between the index of verbal coordination and mean power within a range of frequency bands across trials, and (2) calculating the phase-amplitude coupling between the behavioural and brain signals within single trials (using the power of high-frequency neural activity only). Overall, the findings help to elucidate some of the left hemisphere brain areas involved in interactive speaking behaviours, particularly highlighting the highfrequency activity of the IFG as a potential candidate supporting verbal coordination.
Strengths:
This study provides the field with a convincing demonstration of how to investigate speaking behaviours in more complex situations that share many features with real-world speaking contexts e.g. simultaneous engagement of speech perception and production processes, the presence of an interlocutor, and the need for inter-speaker coordination. The findings thus go beyond previous work that has typically studied solo speech production in isolation, and represent a significant advance in our understanding of speech as a social and communicative behaviour. It is further an impressive feat to develop a paradigm in which the degree of cooperativity of the synchronisation partner can be so tightly controlled; in this way, this study combines the benefits of using prerecorded stimuli (namely, the high degree of experimental control) with the benefits of using a live synchronisation partner (allowing the task to be truly two-way interactive, an important criticism of other work using pre-recorded stimuli). A further key strength of the study lies in its employment of stereotactic EEG to measure brain responses with both high temporal and spatial resolution, an ideal method for studying the unfolding relationship between neural processing and this dynamic coordination behaviour.
We sincerely appreciate the Reviewer's thoughtful and positive feedback on our manuscript.
Weaknesses:
One major limitation of the current study is the lack of coverage of the right hemisphere by the implanted electrodes. Of course, electrode location is solely clinically motivated, and so the authors did not have control over this. However, this means that the current study neglects the potentially important role of the right hemisphere in this task. The right hemisphere has previously been proposed to support feedback control for speech (likely a core process engaged by synchronous speech), as opposed to the left hemisphere which has been argued to underlie feedforward control (Tourville & Guenther, 2011). Indeed, a previous fMRI study of synchronous speech reported the engagement of a network of right hemisphere regions, including STG, IPL, IFG, and the temporal pole (Jasmin et al., 2016). Further, the release from speech-induced suppression during a synchronous speech reported by Jasmin et al. was found in the right temporal pole, which may explain the discrepancy with the current finding of reduced leftward high-frequency activity with increasing verbal coordination (suggesting instead increased speech-induced suppression for successful synchronisation). The findings should therefore be interpreted with the caveat that they are limited to the left hemisphere, and are thus likely missing an important aspect of the neural processing underpinning verbal coordination behaviour.
We have now included, in the supplementary materials, data from the right hemisphere, although the coverage is a bit sparse (Figures S2, S4, S5, see our responses in the ‘Recommendation for the authors’ section, below). We have also revised the Discussion section to add the putative role of right temporal regions (see below as well).
A further limitation of this study is that its findings are purely correlational in nature; that is, the results tell us how neural activity correlates with behaviour, but not whether it is instrumental in that behaviour. Elucidating the latter would require some form of intervention such as electrode stimulation, to disrupt activity in a brain area and measure the resulting effect on behaviour. Any claims therefore as to the specific role of brain areas in verbal coordination (e.g. the role of the IFG in supporting online coordinative adjustments to achieve synchronisation) are therefore speculative.
We appreciate the reviewer’s observation regarding the correlational nature of our findings and agree that this is a common limitation of neuroimaging studies. While elucidating causal relationships would indeed require intervention techniques such as electrical stimulation, our study leverages the unique advantages of intracerebral recordings, offering the best available spatial and temporal resolution alongside a high signal-tonoise ratio. These attributes ensure that our data accurately reflect neural activity and its temporal dynamics, providing a robust foundation for understanding the relationship between neural processes and behaviour. Therefore, while causal claims are beyond the scope of this study, the precision of our methodology allows us to make well-supported observations about the neural correlates of synchronous speech tasks.
Recommendations for the authors:
Reviewing Editor Comment:
After joint consultation, we are seeing the potential for the report to be strengthened and the evidence here to be deemed ultimately at least 'solid': to us (editors and reviewers) it seems that this would require both (1) clarifying/acknowledging the limitations of not having right hemisphere data, and (2) running some of the additional analyses the reviewers suggest, which should allow for richer examination of the data e.g. phase relationships in areas that correlate with synchronisation.
We have now added data on the right hemisphere (RH) that we did not previously report due to a rather sparse sampling of the RH. These results are now reported in the Results section as well as in the Supplementary section, where we put all right hemisphere figures for all analyses (Figure S2, S4, S5). We have also run additional analyses digging into the phase relationship in areas that correlate with synchronisation (Figure S6). These additional analyses allowed us to improve the Discussion section as well.
Reviewer #1 (Recommendations For The Authors):
In some sections, the writing is a bit unclear, with both typos and vague statements that could be fixed with careful proofreading.
We thank the reviewer for pointing out areas where the writing could be improved. We carefully proofread the manuscript to address typos and clarify any vague statements. Specific sections identified as unclear have been rephrased for better precision and readability.
In Figure 1, the colors repeat, making it impossible to tell patients apart.
We have now updated Figure 1 colormap to avoid redundancy and added the right hemisphere.
Line 132: "16 unilateral implantations (9 left, 7 bilateral implantations)". Should this say 7 right hemisphere? If so, the following sentence stating that there was "insufficient cover [sic] of the right hemisphere" is unclear, since the number of patients between LH and RH is similar.
The confusion was due to the fact that the lateralization refers to the presence/absence of electrodes in the Heschl’s gyrus (left : H’ ; right : H) exclusively.
We have thus changed this section as follows:
“16 patients (7 women, mean age 29.8 y, range 17 - 50 y) with pharmacoresistant epilepsy took part in the study. They were included if their implantation map covered at least partially the Heschl's gyrus and had sufficiently intact diction to support relatively sustained language production.” The relevant part (previously line 132) now states:
“Sixteen patients with a total of 236 electrodes (145 in the left hemisphere) and 2395 contacts (1459 in the left hemisphere, see Figure 1). While this gives a rather sparse coverage of the right hemisphere, we decided, due to the rarity of this type of data, to report results for both hemispheres, with figures for the left hemisphere in the main text and figures for the right hemisphere in the supplementary section.”
Reviewer #2 (Recommendations For The Authors):
(1) To address the concern regarding the absence of data from the right hemisphere, I would advise the authors to directly acknowledge this limitation in their Discussion section, citing relevant work suggesting that the right hemisphere has an important role to play in this task (e.g. Jasmin et al., 2016). You should also make this clear in your abstract e.g. you could rewrite the sentence in line 40 to be: "Then, we recorded the intracranial brain activity of the left hemisphere in 16 patients with drug-resistant epilepsy...".
We are grateful to the reviewer for this comment that incited us to look into the right hemisphere data. We have now included results in the right hemisphere, although the coverage is a bit sparse. We have also revised the Discussion section to add the putative role of right temporal regions. Interestingly, our results show, as suggested by the reviewer, a clear involvement of the RH in this task.
First, the full brain analyses show a very similar implication of the RH as compared to the LH (see Figure below). We have now added in the Results section:
“As expected, the whole language network is strongly involved, including both dorsal and ventral pathways (Fig 3A). More precisely, in the left temporal lobe the superior, middle and inferior temporal gyri, in the left parietal lobe the inferior parietal lobule (IPL) and in the left frontal lobe the inferior frontal gyrus (IFG) and the middle frontal gyrus (MFG). Similar results are observed in the right hemisphere, neural responses being present across all six frequency bands with medium to large modulation in activity compared to baseline (Figure S2A) in the same regions. Desynchronizations are present in the theta, alpha and beta bands while the low gamma and HFa bands show power increases.”
As to compared to the left hemisphere, assessing brain-behaviour correlations in the right hemisphere does not provide the same statistical power, because some anatomical regions have very few electrodes. Nonetheless, we observe a strong correlation in the right IFG, similar to the one we previously reported in the left hemisphere, and we now report in the Results section:
“The decrease in HFa along the dorsal pathway is replicated in the right hemisphere (Figure S4). However, while both the right STG BA41/42 and STG BA22 present a power increase (compared to baseline) — with a stronger increase for the STG BA41/42 — neither shows a significant correlation with verbal coordination (t(45)=-1.65, p=.1 ; t(8)=-0.67, p=.5 ; Student’s T test, FDR correction). By contrast, results in the right IFG BA44 are similar to the one observed in the left hemisphere with a significant power increase associated with a negative brainbehaviour correlation (t(17) = -3.11, p = .01 ; Student’s T test, FDR correction).”
Interestingly, the phase-amplitude coupling analysis yields very similar results in both hemispheres (exception made for BA22). We have thus updated the Results section as follows:
“Notably, when comparing – within the regions of interest previously described – the PAC with the virtual partner speech and the PAC with the phase difference, the coupling relationship changes when moving along the dorsal pathway: a stronger coupling in the auditory regions with the speech input, no difference between speech and coordination dynamics in the IPL and a stronger coupling for the coordinative dynamics compared to speech signal in the IFG (Figure 5B ). When looking at the right hemisphere, we observe the same changes in the coupling relationship when moving along the dorsal pathway, except that no difference between speech and coordination dynamics is present in the right secondary auditory regions (STG BA22; Figure S5).”
We also included in the Discussion section the right hemisphere results also mentioning previous work of Guenther and the one of Jasmin. On the section “Left secondary auditory regions are more sensitive to coordinative behaviour” one can read:
“Furthermore, the absence of correlation in the right STG BA22 (Figure S4) seems in first stance to challenge influential speech production models (e.g. Guenther & Hickok, 2016) that propose that the right hemisphere is involved in feedback control. However, one needs to consider the the task at stake heavily relied upon temporal mismatches and adjustments. In this context, the left-lateralized sensitivity to verbal coordination reminds of the works of Floegel and colleagues (2020, 2023) suggesting that both hemispheres are involved depending on the type of error: the right auditory association cortex monitoring preferentially spectral speech features and the left auditory association cortex monitoring preferentially temporal speech features. Nonetheless, the right temporal pole seems to be sensitive to speech coordinative behaviour, confirming previous findings using fMRI (Jasmin et al., 2016) and thus showing that the right hemisphere has an important role to play in this type of tasks (e.g. Jasmin et al., 2016).”
References cited:
– Floegel, M., Fuchs, S., & Kell, C. A. (2020). Differential contributions of the two cerebral hemispheres to temporal and spectral speech feedback control. Nature Communications, 11(1), 2839.
– Floegel, M., Kasper, J., Perrier, P., & Kell, C. A. (2023). How the conception of control influences our understanding of actions. Nature Reviews Neuroscience, 24(5), 313-329.
– Guenther, F. H., & Hickok, G. (2016). Neural models of motor speech control. In Neurobiology of language (pp. 725-740). Academic Press.
(2) When discussing previous work on alignment during synchronous speech, you may wish to include a recently published paper by Bradshaw et al (2024); this manipulated the acoustics of the accompanist's voice during a synchronous speech task to show interactions between speech motor adaptation and phonetic convergence/alignment.
We thank the reviewer for pointing to this recent and interesting paper. We added the article as reference as follows
“Furthermore, synchronous speech favors the emergence of alignment phenomena, for instance of the fundamental frequency or the syllable onset (Assaneo et al., 2019 ; Bradshaw & McGettigan, 2021 ; Bradshaw et al., 2023; Bradshaw et al., 2024).”
(3) Line 80: "Synchronous speech resembles to a certain extent to delayed auditory feedback tasks"- I think you mean "altered auditory feedback tasks" here.
In the case of synchronous speech it is more about timing than altered speech signals, that is why the comparison is done with delayed and not altered auditory feedback. Nonetheless, we understand the Reviewer’s point and we have now changed the sentence as follows:
“Synchronous speech resembles to a certain extent to delayed/altered auditory feedback tasks”
(4) When discussing superior temporal responses during such altered feedback tasks, you may also want to cite a review paper by Meekings and Scott (2021).
We thank the reviewer for this suggestion, indeed this was a big oversight!
The paper is now quoted in the introduction as follows:
“Previous studies have revealed increased responses in the superior temporal regions compared to normal feedback conditions (Hirano et al., 1997 ; Hashimoto & Sakai, 2003 ; Takaso et al., 2010 ; Ozerk et al., 2022 ; Floegel et al., 2020 ; see Meekings & Scott, 2021 for a review of error-monitoring and feedback control in the STG during speech production).”
Furthermore, we updated the discussion part concerning the speaker-induced suppression phenomenon (see below our response to the point 10).
(5) Line 125: "The parameters and sound adjustment were set using an external low-latency sound card (RME Babyface Pro Fs)". Can you please report the total feedback loop latency in your set-up? Or at the least cite the following paper which reports low latencies with this audio device.
Kim, K. S., Wang, H., & Max, L. (2020). It's About Time: Minimizing Hardware and Software Latencies in Speech Research With Real-Time Auditory Feedback. Journal of Speech, Language, and Hearing Research, 63(8), 25222534. https://doi.org/10.1044/2020_JSLHR-19-00419
We now report the total feedback loop latency (~5ms) and also cite the relevant paper (Kim et al., 2020).
(6) Line 127 "A calibration was made to find a comfortable volume and an optimal balance for both the sound of the participant's own voice, which was fed back through the headphones, and the sound of the stimuli." What do you mean here by an 'optimal balance'? Was the participant's own voice always louder than the VP stimuli? Can you report roughly what you consider to be a comfortable volume in dB?
This point was indeed unlcear. We have now changed as follows:
“A calibration was made to find a comfortable volume and an optimal balance for both the sound of the participant's own voice, which was fed back through the headphones, and the sound of the stimuli. The aim of this procedure was that the patient would subjectively perceive their voice and the VP-voice in equal measure. VP voice was delivered at approximately 70dB.”
(7) Relatedly, did you use any noise masking to mask the air-conducted feedback from their own voice (which would have been slightly out of phase with the feedback through the headphones, depending on your latency)?
Considering the low-latency condition allowed with the sound card (RME Babyface Pro Fs), we did not use noise masking to mask the air-conducted feedback from the self-voice of the patients.
(8) Line 141: "four short sentences were pre-recorded by a woman and a man." Did all participants synchronise with both the man and woman or was the VP gender matched to that of the participant/patient?
We thank the reviewer for this important missing detail. We know changed the text as follows:
“Four stimuli corresponding to four short sentences were pre-recorded by both a female and a male speaker. This allowed to adapt to the natural gender differences in fundamental frequency (i.e. so that the VP gender matched that of the patients). All stimuli were normalised in amplitude.”
(9) Can you clarify what instructions participants were given regarding the VP? That is, were they told that this was a recording or a real live speaker? Were they naïve to the manipulation of the VP's coupling to the participant?
We have now added this information to the task description as follows:
“Participants, comfortably seated in a medical chair, were instructed that they would perform a real-time interactive synchronous speech task with an artificial agent (Virtual Partner, henceforth VP, see next section) that can modulate and adapt to the participant’s speech in real time.”
“The third step was the actual experiment. This was identical to the training but consisted of 24 trials (14s long, speech rate ~3Hz, yielding ~1000 syllables). Importantly, the VP varied its coupling behaviour to the participant. More precisely, for a third of the sequences the VP had a neutral behaviour (close to zero coupling : k = +/- 0.01). For a third it had a moderate coupling, meaning that the VP synchronised more to the participant speech (k = - 0.09). And for the last third of the sequences the VP had a moderate coupling but with a phase shift of pi/2, meaning that it moderately aimed to speak in between the participant syllables (k = + 0.09). The coupling values were empirically determined on the basis of a pilot experiment in order to induce more or less synchronization, but keeping the phase-shifted coupling at a rather implicit level. In other terms, while participants knew that the VP would adapt, they did not necessarily know in which direction the coupling went.”
(10) The paragraph from line 438 entitled "Secondary auditory regions are more sensitive to coordinative behaviour" includes an interesting discussion of the relation of the current findings to the phenomenon of speech-induced suppression (SIS). However, the authors appear to equate the observed decrease in highfrequency activity as speech coordination increases with the phenomenon of SIS (in lines 456-457), which is quite a speculative leap. I would encourage the authors to temper this discussion by referring to SIS as a potentially related phenomenon, with a need for more experimental work to determine if this is indeed the same phenomenon as the decreases in high-frequency power observed here. I believe that the authors are arguing here for an interpretation of SIS as reflecting internal modelling of sensory input regardless of whether this is self-generated or other-generated; if this is indeed the case, I would ask the authors to be more explicit here that these ideas are not a standard part of the traditional account of SIS, which only includes internal modelling of self-produced sensory feedback.
As stated in the public review, we thank both reviewers for raising thoughtful concerns about our interpretation of the observed neural suppression as related to speaker-induced suppression (SIS). We agree that our study lacks a passive listening condition, which limits direct comparisons to the original SIS effect, traditionally defined as the suppression of neural responses to self-produced speech compared to externally-generated speech (Meekings & Scott, 2021).
In response, we have reconsidered our terminology and interpretation. In the revised discussion, we refer to our findings as a "SIS-related phenomenon specific to the synchronous speech context." Unlike classic SIS paradigms, our interactive task involves simultaneous monitoring of self- and externally-generated speech, introducing additional attentional and coordinative demands.
The revised discussion also incorporates findings by Ozker et al. (2024, 2022), which link SIS and speech monitoring, suggesting that suppressing responses to self-generated speech facilitates error detection. We propose that the decrease in high-frequency activity (HFa) as verbal coordination increases reflects reduced error signals due to closer alignment between perceived and produced speech. Conversely, HFa increases with reduced coordination may signify greater prediction error.
Additionally, we relate our findings to the "rubber voice" effect (Zheng et al., 2011; Lind et al., 2014; Franken et al., 2021), where temporally and phonetically congruent external speech can be perceived as self-generated. We speculate that this may occur in synchronous speech tasks when the participant's and VP's speech signals closely align. However, this interpretation remains speculative, as no subjective reports were collected to confirm this perception. Future studies could include participant questionnaires to validate this effect and relate subjective experience to neural measures of synchronization.
Overall, our findings extend the study of SIS to dynamic, interactive contexts and contribute to understanding internal forward models of speech production in more naturalistic scenarios.
We have now added these points to the discussion as follows:
“The observed negative correlation between verbal coordination and high-frequency activity (HFa) in STG BA22 suggests a suppression of neural responses as the degree of synchrony increases. This result aligns with findings on speaker-induced suppression (SIS), where neural activity in auditory cortex decreases during self-generated speech compared to externally-generated speech (Meekings & Scott, 2021; Niziolek et al., 2013). However, our paradigm differs from traditional SIS studies in two critical ways: (1) the speaker's own voice is always present and predictable from the forward model, and (2) no passive listening condition was included. Therefore, our findings cannot be directly equated with the original SIS effect.
Instead, we propose that the suppression observed here reflects a SIS-related phenomenon specific to the synchronous speech context. Synchronous speech requires simultaneous monitoring of self- and externally generated speech, a task that is both attentionally demanding and coordinative. This aligns with evidence from Ozker et al. (2024, 2022), showing that the same neural populations in STG exhibit SIS and heightened responses to feedback perturbations. These findings suggest that SIS and speech monitoring are related processes, where suppressing responses to self-generated speech facilitates error detection.
In our study, suppression of HFa as coordination increases may reflect reduced prediction errors due to closer alignment between perceived and produced speech signals. Conversely, increased HFa during poor coordination may signify greater mismatch, consistent with prediction error theories (Houde & Nagarajan, 2011; Friston et al., 2020).”
(11) Within this section, you also speculate in line 460 that "Moreover, when the two speech signals come close enough in time, the patient possibly perceives them as its own voice." I would recommend citing studies on the 'rubber voice' effect to back up this claim (e.g. Franken et al., 2021; Lind et al., 2014; Zheng et al., 2011).
We are grateful to the Reviewer for this interesting suggestion. Directly following the previous comment, the section now states:
“Furthermore, when self- and externally-generated speech signals are temporally and phonetically congruent, participants may perceive external speech as their own. This echoes the "rubber voice" effect, where external speech resembling self-produced feedback is perceived as self-generated (Zheng et al., 2011; Lind et al., 2014; Franken et al., 2021). While this interpretation remains speculative, future studies could incorporate subjective reports to investigate this phenomenon in more detail.”
(12) As noted in my public review, since your methods are correlational, you need to be careful about inferring the causal role of any brain areas in supporting a specific aspect of functioning e.g. line 501-504: "By contrast, in the inferior frontal gyrus, the coupling in the high-frequency activity is strongest with the input-output phase difference (input of the VP - output of the speaker), a metric that reflects the amount of error in the internal computation to reach optimal coordination, which indicates that this region optimises the predictive and coordinative behaviour required by the task." I would argue that the latter part of this sentence is a conclusion that, although consistent with, goes beyond the current data in this study, and thus needs tempering.
We agree with the Reviewer and changed the sentence as follows:
“By contrast, in the inferior frontal gyrus, the coupling in the high-frequency activity is strongest with the inputoutput phase difference (input of the VP - output of the speaker), a metric that could possibly reflect the amount of error in the internal computation to reach optimal coordination. This indicates that this region could have an implication in the optimisation of the predictive and coordinative behaviour required by the task.”
