Prosodic recitation training selectively enhances sensorimotor connectivity during poetry reading: a longitudinal fNIRS study

Abstract

Whether different modes of language engagement induce distinct neural changes remains poorly understood. Here we provide longitudinal evidence that vocal recitation training—but not semantic analysis—is associated with changes in sensorimotor networks that correlate with emotional responses. We assigned 70 participants to prosody training (rhythmic recitation), comprehension training (semantic analysis), or transcription control, all engaging with identical poetic materials. Using fNIRS before and after a four-day intervention, a linear mixed-effects model revealed a significant group × training interaction confirming that only prosody training enhanced connectivity between the middle frontal gyrus (MFG) and postcentral gyrus (PoCG)—regions supporting motor planning and sensorimotor integration. This was accompanied by improved rhythmic performance and increased physiological arousal. These effects extended to the untrained portion of the same poem, suggesting near transfer to structurally similar materials. Inter-subject representational similarity analysis revealed that training-induced neural changes in sensorimotor regions (MFG, IPL, PoCG) correlated with individual differences in emotional arousal, indicating a brain-behavior correspondence with affective processing. In contrast, comprehension training did not produce comparable sensorimotor connectivity changes, while transcription produced minimal effects. These findings suggest that engagement mode can shape sensorimotor connectivity, providing convergent behavioral, physiological, and neural evidence for sensorimotor involvement in prosodic recitation.

A four-day fNIRS training study in language learners shows that reciting poetry aloud, but not analyzing its meaning, strengthens sensorimotor brain connectivity linked to emotional arousal during reading.

Introduction

Poetry is a literary form in which semantic content is uniquely fused with musical structure through prosodic features, such as rhythm and rhyme, engaging both linguistic and aesthetic processing. A key question for cognitive neuroscience is whether different modes of engaging with such materials—reading silently, analyzing meaning, or reciting aloud—differentially shape neural organization and emotional responses. Here we report a longitudinal training study examining whether prosodic recitation training, compared with semantic analysis and passive transcription, selectively strengthens sensorimotor brain networks during poetry reading.

Emerging research reveals that poetry activates brain networks extending beyond canonical language regions. Neuroimaging studies demonstrate that poetry reading engages not only the inferior frontal and temporal language areas but also regions associated with rhythm, music, and emotion^1,2. Chinese poetry specifically activates the orbitofrontal cortex and insula—regions linked to aesthetic evaluation—alongside action simulation areas including the supplementary motor area and precentral gyrus³. Recent work shows that poetry composition strengthens connections between language, sensorimotor, and executive networks compared to descriptive writing⁴. These findings suggest that poetry naturally recruits simulation networks, with language rich in imagery and rhythm engaging the inferior frontal gyrus (IFG), precentral gyrus (PreCG), and postcentral gyrus (PoCG)—regions central to motor planning, sensorimotor integration, movement execution, and action simulation^5–8. Behaviorally, poetic devices enhance memory and comprehension^9,10, while recited poetry elicits physiological responses comparable to music, including chills and reward-related neural activity¹¹. Such activations support readers’ vivid reenactment of poetic scenes and foster engagement with the speaker’s emotional state¹².

We hypothesize that vocal recitation may represent a particularly effective mode for engaging sensorimotor networks. Poetry’s metrical structure imposes specific motor constraints on articulation, while its condensed imagery may promote sensorimotor simulation. When readers vocalize poetry, the physical act of rhythmic speech production could activate motor planning regions through articulation demands, potentially strengthening connections with somatosensory areas that encode bodily sensations. This sensorimotor coupling could provide a neural substrate for experiencing verse through sensorimotor engagement rather than merely decoding semantic content. If this account is correct, then training in prosodic recitation should specifically enhance connectivity between motor planning (e.g., MFG) and somatosensory (e.g., PoCG, PreCG) regions^13,14, while leaving pure semantic processing networks relatively unchanged.

Despite these theoretical advances, significant methodological limitations persist. First, most prior studies rely on cross-sectional contrasts between poetry and prose, confounding genre effects with differences in content and emotional valence^1,3. Second, nearly all neuroimaging research to date has employed fMRI^4,11,15, constraining participants to silent reading or passive listening in noisy scanner environments—conditions that fail to capture the natural experience of reciting poetry aloud. Third, no study has yet tested whether different modes of engagement are associated with differential neural organization, leaving open the critical question of whether recitation training can actually strengthen sensorimotor networks.

To address these gaps, we employ a longitudinal learning paradigm with functional near-infrared spectroscopy (fNIRS), capitalizing on two key advantages. Such designs reduce confounds, provide stronger inferential leverage, and have successfully revealed language learning effects at lexical and passage levels^16–18. Even brief interventions can induce measurable cortical changes, highlighting the brain’s sensitivity to structured practice. Additionally, fNIRS’s robustness to motion artifacts enables reliable measurement during overt speech¹⁹, making it well suited for studying poetry recitation in its natural, vocalized form. Previous fNIRS studies confirm that speech-related motor artifacts minimally affect hemodynamic responses in non-motor regions²⁰. We focus on Chinese EFL (English as a Foreign Language) learners, who process rhythm and prosody less automatically than native speakers, offering a potentially more sensitive model for detecting training-induced neural changes—though this population-specific advantage also constrains the generalizability of our findings to native speakers or other learner populations (see Limitations).

We conducted a longitudinal experiment with seventy Chinese EFL learners randomly assigned to prosody training (focused on rhythmic recitation and prosodic awareness), comprehension training (focused on semantic understanding), or transcription practice (mechanical copying as control). All groups engaged with identical poetic materials drawn from Edgar Allan Poe’s the Raven, ensuring strict content control. We employed behavioral ratings, acoustic analysis, skin conductance responses (SCR), and fNIRS to capture multifaceted training effects. Furthermore, we applied functional connectivity analysis and inter-subject representational similarity analysis (IS-RSA) to link neural dynamics with emotional responses.

Based on our theoretical framework, we hypothesized that prosodic training would induce experience-dependent changes in sensorimotor networks. Specifically, we predicted that prosody training would: (1) enhance behavioral rhythmic performance; (2) strengthen functional connectivity between motor planning (e.g., MFG) and somatosensory regions (e.g., PoCG); and (3) reveal brain-behavior correspondence wherein neural changes predict individual differences in emotional arousal. We tested prediction (1) using behavioral and acoustic measures, prediction (2) using phase locking value (PLV)-based functional connectivity with a group × training interaction test, and prediction (3) using IS-RSA. By contrast, comprehension training was expected to strengthen language-processing networks, while transcription training would exert minimal effects. This design enables us to examine whether vocal recitation represents an effective mode for engaging with poetry by recruiting sensorimotor networks more strongly than semantic analysis alone.

Results

Behavioral performance demonstrates selective prosody training effects

Prosody knowledge

Linear mixed-effects model analysis revealed a significant interaction between group and training (χ² =  34.392, df = 2, p < 0.001). Post hoc tests (see Fig. 1a) showed that the prosody training group scored significantly higher on prosody knowledge after training compared to before training (β = 2.916, t(67) = 9.629, p < 0.001), significantly higher than the comprehension group after training (β = 2.481, t(103) = 5.468, p < 0.001), and significantly higher than the transcription group after training (β = 2.982, t(103) = 6.215, p < 0.001). Neither the comprehension nor the transcription groups showed significant improvements in prosody knowledge.

Fig. 1. Training-induced changes in behavioral, acoustic, and physiological measures.

a Prosody knowledge test scores (0–7 scale) showing selective and significant improvement in the prosody group after training. b Self-reported appreciation ratings (1–5 scale) for Poem 1 demonstrating enhanced aesthetic appreciation following prosody and comprehension training but not transcription. c Comprehension ratings for trained Poem 1 revealing improvements across all groups, with comprehension training producing the largest gains as expected. d Comprehension ratings for untrained Poem 2 showing significant generalization effects only for prosody and comprehension groups. e Emotional arousal self-ratings indicating enhanced subjective emotional engagement following prosody and comprehension training. f Skin conductance response event counts (number of SCR peaks >0.01 μS during 4-min reading) showing objective increases in autonomic arousal following prosody training. g Speaking rate (syllables per minute) for Poem 1 demonstrating increased fluency specifically after prosody and comprehension training. h Articulation rate showing selective improvement in the prosody group that generalized across both poems. i Intensity-based stressed-to-unstressed syllable ratios (S/W) revealing enhanced rhythmic differentiation uniquely following prosody training. Statistical significance determined by linear mixed-effects models with post hoc Tukey corrections.

Appreciation

For Poem 1 appreciation self-ratings (see Fig. 1b), a significant interaction emerged between group and training (χ² = 8.603, df = 2, p = 0.014). Post hoc tests showed that the prosody training group rated appreciation significantly higher after training than before (β = 1.084, t(67) = 7.035, p < 0.001). Similarly, the comprehension group rated appreciation significantly higher after training (β = 1.121, t(67) = 7.295, p < 0.001). However, no significant difference was found in the transcription group appreciation before versus after training (β = 0.501, t(67) = 2.913, p = 0.061).

Comprehension

For Poem 1 comprehension self-ratings (see Fig. 1c), a significant interaction was observed between group and training (χ² = 44.166, df = 2, p < 0.001). Post hoc tests indicated that the prosody training group reported significantly higher comprehension after training than before (β = 1.081, t(67) = 7.271, p < 0.001). The comprehension group also showed a significant increase in comprehension ratings after training (β = 2.443, t(67) = 16.428, p < 0.001), as did the transcription group (β = 0.954, t(67) = 5.721, p < 0.001). For Poem 2 comprehension (see Fig. 1d), a marginally significant interaction emerged between group and training (χ² =  5.014, df = 2, p = 0.082). Post hoc tests revealed that the prosody training group showed significantly higher comprehension ratings after training (β = 0.601, t(67) = 4.125, p = 0.002), and the comprehension group also showed significant improvement (β = 0.884, t(67) = 6.050, p < 0.001). However, no significant difference was found in the transcription group’s comprehension before versus after training (β = 0.400, t(67) = 2.460, p = 0.151).

Emotional arousal

For Poem 1 emotional arousal self-ratings (see Fig. 1e), a significant interaction was found between group and training (χ² = 10.053, df = 2, p = 0.005). Post hoc tests showed that the prosody training group reported significantly higher emotional arousal after training (β = 1.003, t(67) = 5.654, p < 0.001), and the comprehension group also showed a significant increase (β = 1.320, t(67) = 7.464, p < 0.001). However, no significant difference was found in the transcription group’s emotional arousal before versus after training (β = 0.450, t(67) = 2.276, p = 0.219). For Poem 2, emotional arousal ratings showed no significant interaction between group and training, suggesting that emotional engagement effects were specific to the trained material.

Acoustic analysis reveals enhanced rhythmic performance

Global fluency measures

For speaking rate (see Fig. 1g), the final fitted model was: lmer(speaking_rate ~ training × text type × group  + (1 + training + text type | subject)). Results showed a significant three-way interaction among training, group, and text type (χ² = 25.103, df = 2, p < 0.001). Post hoc tests revealed that for the prosody group, speaking rate significantly increased after training when reading Poem 1 (β = 0.402, t(98) = 9.884, p < 0.001). Similarly, for the comprehension group, speaking rate significantly increased after training when reading Poem 1 (β = 0.209, t(98) = 4.931, p < 0.001). However, for the transcription group, no significant difference in speaking rate emerged before versus after training when reading Poem 1 (β = 0.133, t(98) = 2.833, p = 0.181). For Poem 2, no significant differences in speaking rate before versus after training were observed for the prosody group (β = 0.066, t(98) = 1.61, p = 0.907), comprehension group (β = 0.098, t(98) = 1.640, p = 0.887), or transcription group (β = 0.092, t(98) = 1.95, p = 0.723).

For articulation rate (see Fig. 1h), the final fitted model was: lmer(articulation_rate ~ training × group +  training × text type + (1 + training | subject)). Results showed a significant interaction between training and group (χ² = 19.278, df = 2, p < 0.001). Post hoc tests indicated that the prosody group exhibited a significant increase in articulation rate after training (β = 0.518, t(61) = 6.437, p < 0.001). However, for the comprehension group (β = 0.183, t(61) = 2.180, p = 0.262) and transcription group (β = 0.039, t(61) = 0.428, p = 0.998), no significant difference in articulation rate emerged.

Local prosodic measures

We calculated the intensity-based stressed-to-unstressed syllable ratio (S/W ratio) to examine whether participants’ reading became more rhythmic after training (see Fig. 1i). The final fitted model was: lmer(sw_intensity ~ training × group + text type + (1 | subject) + (1 | item)). Results revealed a significant interaction between training and group (χ² = 6.622, df = 2, p = 0.037). Follow-up analyses showed that only in the prosody group did the intensity-based S/W ratio significantly increase after training compared to before (β = 0.011, t(3744) = 3.857, p = 0.002). For the comprehension group (β = 0.001, t(3744) = 0.468, p = 0.997) and transcription group (β = 0.002, t(3744) = 0.654, p = 0.986), no significant difference in intensity-based S/W ratio emerged. Furthermore, the post-training S/W ratio in the prosody group was significantly higher than the comprehension group after training (β = 0.018, t(99.7) = 3.969, p = 0.002) and higher than the transcription group after training (β = 0.017, t(99.7) = 3.558, p = 0.007).

Physiological responses index enhanced emotional engagement

To objectively assess emotional arousal levels during poetry reading before and after training, we measured and analyzed participants’ SCRs (see Fig. 1f). The final fitted model was: glmer(Event_count ~  training × group + text type + (1 | subject), family = poisson). Results showed a significant interaction between training and group (χ² = 69.513, df = 2, p < 0.001). Post hoc tests indicated that compared to pre-training, emotional arousal frequency significantly increased in the prosody group after training (β = 0.244, z = 6.519, p < 0.001), whereas it significantly decreased in the transcription group after completing the tasks (β = −0.230, z = −5.311, p < 0.001). However, for comprehension group participants, no significant difference was found in emotional arousal frequency before versus after completing the tasks (β = 0.006, z = 0.180, p = 0.998).

Functional connectivity analysis reveals training-specific sensorimotor changes

To directly test whether training-induced connectivity changes differed across groups, we fitted linear mixed-effects models with group × training interaction terms, controlling for baseline language proficiency (CET6, LexTALE).

For the primary MFG-PoCG analysis during Poem 1 reading, the LMM revealed a significant group × training interaction (χ² = 15.402, df = 2, p < 0.001). Post-hoc comparisons at post-training showed that the prosody group had significantly higher MFG-PoCG connectivity than both the comprehension group (β = 0.167, t(111) = 3.026, p.adj = 0.023) and the transcription group (β = 0.161, t(110) = 2.722, p.adj = 0.036). No significant difference was found between the comprehension and transcription groups (β = 0.006, t(111) = 0.105, p.adj =  0.922). Within-group paired t-tests confirmed that the prosody group showed significantly enhanced MFG-PoCG connectivity after training (t(24) = 3.670, SD = 0.173, p.adj = 0.034), while neither the comprehension nor the transcription group showed significant changes. See Fig. 2a, d.

Fig. 2. Training-specific enhancement of functional connectivity between cortical regions.

a Prosody training selectively strengthened connectivity between left middle frontal gyrus (MFG) and left postcentral gyrus (PoCG) during reading of trained Poem 1. b The MFG-PoCG connectivity enhancement persisted during reading of untrained Poem 2. c Comprehension training increased connectivity between left inferior frontal gyrus (IFG) and left MFG exclusively for trained Poem 1. Heatmaps display group-averaged phase locking values (PLV) difference scores (post-training minus pre-training) with FDR-corrected significant connections highlighted in warm colors. Brain renderings illustrate the anatomical locations of the involved ROIs. d Functional connectivity (PLV) between MFG and PoCG is shown for the prosody, comprehension, and transcription groups at pre- and post-training for the trained poem (Poem 1) and the untrained poem (Poem 2). LMMs revealed a significant group × training interaction in both conditions (Poem 1: χ² =  15.402, df = 2, p_adj < 0.001; Poem 2: χ² = 15.012, df = 2, p_adj < 0.001). Post hoc comparisons at post-training indicated higher connectivity in the prosody group relative to the comprehension and transcription groups, with no significant difference between the latter two. These effects were observed for both trained and untrained materials.

For the untrained Poem 2, the group × training interaction for MFG-PoCG connectivity was again significant (χ² = 15.012, df = 2, p < 0.001). Post-hoc comparisons at post-training confirmed that the prosody group had significantly higher connectivity than both the comprehension group (β = 0.134, t(111) = 2.482, p.adj = 0.043) and the transcription group (β = 0.170, t(110) = 2.943, p.adj = 0.029), with no difference between the latter two groups (β = 0.036, t(111) = 0.627, p.adj = 0.854). Within-group paired t-tests showed enhanced MFG-PoCG connectivity in the prosody group for Poem 2 (t(24) = 3.825, SD = 0.165, p.adj = 0.012), indicating that the training-related connectivity change extended to structurally similar untrained material. See Fig. 2b, d.

In the comprehension training group, within-group paired t-tests found significantly enhanced IFG-MFG connectivity during Poem 1 reading after training (t(24) = 3.665, SD = 0.242, p.adj = 0.040; see Fig. 2c). However, the LMM revealed no significant group × training interaction for IFG-MFG connectivity (χ² = 2.473, df = 2, p.adj = 0.294), indicating that this within-group effect did not differ statistically from changes observed in the other groups. No significant enhancement in functional connectivity was observed in the comprehension group during Poem 2 reading, indicating that comprehension training effects remained specific to trained material.

The transcription group showed no significant changes in functional connectivity after task completion for either poem. Furthermore, no significant differences in functional connectivity were found across all three groups during expository text reading after training (all ps > 0.05), confirming that training effects were specific to poetic materials.

Neural-behavioral correspondence: individual differences in emotional responses

Inter-subject representational similarity analysis (IS-RSA) examined whether training-induced neural change patterns corresponded with physiological emotional response patterns derived from SCR-recorded emotional frequency (Event_count). This analysis tests the hypothesis that participants with more similar neural changes also exhibit more similar emotional responses.

In the prosody group (see Fig. 3b), for the trained Poem 1, neural RDMs of the left MFG (r = 0.411, p.adj =  0.017), left inferior parietal lobule (IPL) (r = 0.368, p.adj = 0.028), and left PoCG (r = 0.354, p.adj = 0.044) were significantly positively correlated with the physiological emotional RDM. These correlations extended to the untrained Poem 2: neural RDMs of left MFG (r = 0.386, p.adj = 0.021), left IPL (r = 0.328, p.adj = 0.039), and left PoCG (r = 0.364, p.adj = 0.037) remained significantly correlated with physiological emotional responses. This pattern indicates that prosody training-induced neural changes in sensorimotor regions correlated with individual differences in emotional arousal.

Fig. 3. RSA and regression results.

a Data processing pipeline: this panel illustrates data processing steps for RSA. For each participant in the prosody, comprehension, and transcription groups, preprocessed HbO time-series data were extracted for Poem 1, Poem 2, and expository text trials before and after training. Difference vectors (post-training minus pre-training) were computed for each ROI and trial type. Euclidean distances between participants’ difference vectors were calculated to generate neural RDMs for each ROI and training condition. Similarly, SCR-recorded emotional frequency differences were used to create behavioral RDMs. Spearman correlation coefficients were then computed using Mantel tests to assess representational similarity between neural and physiological RDMs. b RSA between neural RDMs and SCR-recorded emotional RDMs: in the prosody group, left MFG, IPL, and PoCG showed significant similarity for both Poem 1 and Poem 2. In the comprehension group, left IFG and SMG showed significant similarity for Poem 1 only. No effects were observed in the transcription group or for expository texts. c Regression results for neural RDMs: only the prosody group showed a significant effect of comprehension level in left MTG for Poem 1, with higher comprehension differences linked to greater neural dissimilarity. The icons used in (a) are self-made, original, and fully compliant with CC BY licensing and commercially reusable content requirements.

In the comprehension group, for the trained Poem 1 (see Fig. 3b), neural RDMs of Left IFG (r = 0.401, p.adj  = 0.020) and Left supramarginal gyrus (SMG) (r = 0.441, p.adj = 0.009) were significantly positively correlated with the physiological emotional RDM. However, no significant correlation was found between neural RDMs and physiological emotional RDM for the untrained Poem 2 in the comprehension group. For the transcription group, no significant correlations between neural RDMs and physiological emotional RDM were observed for either Poem 1 or Poem 2 (all ps > 0.05). Additionally, for all three training groups, no significant correlations were found between neural RDMs for expository texts and physiological emotional RDM (all ps > 0.05).

Regression analysis

We fitted the following model: lmer(ROI_RDM ~ prosody_knowledge + appreciation + comprehension +  emotional_arousal + (1 | subject1) + (1 | subject2)). Results revealed a significant main effect of comprehension level in left middle temporal gyrus only for Poem 1 in the prosody group (χ² = 7.518, df = 1, p.adj = 0.039; see Fig. 3c). This indicates that participants with greater differences in their understanding of the poem exhibited less similar neural activity patterns in left MTG, suggesting that semantic processing becomes more deeply encoded in neural representations specifically under prosodic engagement conditions. No other significant main effects emerged in other ROIs or conditions.

Discussion

This study provides convergent longitudinal evidence that vocal recitation of poetry—rather than semantic analysis or mechanical transcription—is selectively associated with enhanced connectivity in sensorimotor networks that correlates with emotional responses. By integrating behavioral ratings, acoustic analyses, physiological measures, and fNIRS data, we obtained converging evidence across four levels of analysis. First, prosody training significantly improved rhythmic reading performance and prosodic awareness, as indexed by faster articulation rates, stronger stressed-to-unstressed syllable ratios, and enhanced prosody knowledge test scores. Second, it selectively strengthened functional connectivity between the middle frontal gyrus (MFG) and postcentral gyrus (PoCG), as confirmed by a significant group × training interaction in linear mixed-effects modeling, a fronto-parietal network supporting sensorimotor integration. Third, it enhanced both subjective (self-rated appreciation and emotional arousal) and objective (SCR event frequency) indices of emotional engagement. Fourth, IS-RSA revealed that neural change patterns in sensorimotor regions correlated with individual differences in physiological emotional responses. Critically, neither comprehension nor transcription training produced comparable multidimensional effects. These findings provide convergent evidence consistent with the hypothesis that vocal recitation engages sensorimotor networks more extensively than other modes of poetic engagement, aligning with theoretical predictions about the relationship between prosodic performance and sensorimotor processing. We emphasize that, at the neural level, the robust and statistically confirmed effect is the training-specific enhancement of MFG-PoCG connectivity; the broader contribution of this study lies in the convergence of this connectivity change with behavioral, acoustic, physiological, and brain-behavior evidence, rather than in a claim of widespread network reorganization.

The central finding of this study is that prosody training selectively strengthened functional connectivity between the MFG and postcentral gyrus (PoCG)—a neural pathway linking motor planning to somatosensory integration. Importantly, this selectivity was confirmed not only by within-group pre-post comparisons but also by a significant group × training interaction, providing direct statistical evidence that the MFG-PoCG connectivity change was specific to prosody training. The MFG plays crucial roles in motor planning, working memory, and action-related processing^5,13,14, while the postcentral gyrus serves as the primary somatosensory cortex, integrating bodily sensations with motor commands^6,7. The strengthened coupling between these regions is consistent with our hypothesis that prosodic recitation activates sensorimotor integration networks during engagement with the poem’s rhythmic structure.

An important methodological consideration is whether the observed MFG-PoCG connectivity changes reflect training-related sensorimotor engagement rather than motion artifacts from vocal articulation. Three lines of evidence argue against this alternative. First, participants read expository text aloud, yet this condition showed no significant training-induced sensorimotor modulation. Second, the transcription group also engaged in vocal reading during assessment but exhibited no significant sensorimotor connectivity changes. Third, our pre-test to post-test analytical approach effectively subtracts baseline motor signals. The selective MFG-PoCG enhancement following prosody training is more consistent with training-specific sensorimotor engagement rather than general vocalization-related motor processes. While motion artifacts can thus be reasonably excluded, we acknowledge that other aspects of the prosody training condition—including greater vocal practice time, rehearsal demands, and potential memorization involvement—may also have contributed to the observed effects, and we discuss these factors in the Limitations section.

This interpretation is supported by converging evidence across multiple measurement levels. Behaviorally, participants in the prosody group demonstrated mastery of prosodic form through improved prosody knowledge test scores, faster articulation rates, and stronger intensity-based stressed-to-unstressed syllable ratios—acoustic markers of more rhythmic and fluent oral performance. The co-occurrence of these behavioral improvements and neural connectivity changes is consistent with the proposal that the physical act of rhythmic speech production engages sensorimotor networks more strongly than semantic processing alone^4,21,22.

The IS-RSA provides additional insight by demonstrating that individual differences in neural changes within sensorimotor regions (MFG, IPL, and PoCG) correlated with physiological emotional responses. Participants who showed greater neural changes in these regions following prosody training also exhibited stronger autonomic arousal during poetry reading, as indexed by SCR event frequency. This brain-behavior correspondence was absent in comprehension and transcription groups, suggesting a selective association from prosodic engagement through sensorimotor processing to emotional engagement. We note, however, that this correlational pattern does not in itself establish a causal pathway from neural change to emotional response; alternative explanations, including a common underlying factor driving both neural and emotional changes, remain possible.

The extension of prosody training effects from trained to untrained stanzas is noteworthy. Despite never practicing the second half of the Raven, participants in the prosody group showed enhanced MFG-PoCG connectivity, improved acoustic performance, and increased emotional arousal when reading these novel stanzas. Because the trained and untrained portions share the same meter, rhyme scheme, and stylistic features, this transfer is best characterized as near transfer to structurally similar material²³ rather than evidence of domain-general skill acquisition. Whether prosodic training effects would generalize to poems with different metrical structures or to non-poetic rhythmic texts remains an open question for future research.

By contrast, comprehension training engaged classical language-processing pathways without significantly enhancing sensorimotor connectivity. The enhanced IFG-MFG connectivity observed within this group aligns with established roles of the IFG in semantic retrieval and syntactic processing^24,25. However, we note that the IFG-MFG group × training interaction was not statistically significant, indicating that this within-group change did not differ reliably from changes in other groups. This finding should therefore be interpreted with caution. Although comprehension training improved semantic understanding and enhanced appreciation ratings for the trained poem, it did not significantly strengthen sensorimotor networks, extend effects to untrained material, or produce the physiological arousal changes observed with prosody training. We also acknowledge that the comprehension condition did not employ action-related semantic content, which might have been more likely to engage motor regions. Future studies could examine whether action-sentence comprehension training produces different sensorimotor outcomes. An additional observation from the regression analysis is that comprehension level differences predicted neural dissimilarity in left MTG, but only for the prosody group reading the trained poem. This suggests that under conditions of prosodic engagement, individual differences in semantic understanding may be more distinctly reflected in neural activity patterns—a finding consistent with the view that form and meaning interact during poetry processing, though the specificity of this effect to a single ROI and condition warrants cautious interpretation. The transcription group’s lack of neural or behavioral changes confirms that passive engagement is insufficient for either semantic or sensorimotor processing.

The convergence of subjective and objective emotional measures provides evidence for recitation’s affective effects. Prosody training not only increased self-reported aesthetic appreciation and emotional arousal but also enhanced SCR event frequency—a well-established index of autonomic arousal independent of conscious report²⁶.

We propose that rhythmic vocalization may create a state of motor-respiratory entrainment wherein motor, respiratory, and cardiovascular systems synchronize to the poem’s temporal structure. This hypothesis extends beyond traditional cognitive accounts by proposing that the act of vocalization itself—not just the cognitive processing of rhythm—contributes to emotional engagement^27,28. This synchronization may facilitate emotional engagement with the poetic speaker’s voice, similar to mechanisms observed in music perception where rhythmic entrainment enhances emotional engagement²⁹.

Integrating our findings, we propose a descriptive framework summarizing how different modes of poetic engagement are associated with distinct patterns of neural and behavioral change (Fig. 4). Three observations emerge from this framework. First, prosody training was uniquely associated with enhanced sensorimotor connectivity (MFG-PoCG), as confirmed by a significant group × training interaction (Fig. 4a), suggesting that rhythmic recitation engages motor planning and somatosensory systems in a manner not observed with semantic analysis or passive copying. Second, this sensorimotor connectivity change extended to structurally similar untrained material (near transfer), whereas comprehension training effects remained limited to trained material—a dissociation consistent with the possibility that prosodic engagement may strengthen more generalizable processing routines than semantic analysis alone. Third, IS-RSA revealed that individual differences in sensorimotor connectivity changes correlated with physiological emotional responses in the prosody group, suggesting a brain-behavior correspondence between sensorimotor processing and emotional engagement. However, we emphasize that this framework is descriptive rather than mechanistic: the correlational nature of the brain-behavior findings does not establish a causal pathway from neural change to emotional response, and alternative explanations remain possible (see Limitations). By contrast, comprehension training was associated with within-group IFG-MFG connectivity enhancement, although this effect did not reach significance in the between-group interaction test and was limited to trained material (Fig. 4b).

Fig. 4. Descriptive framework of training-specific effects on neural connectivity during poetry reading.

The framework summarizes how three training conditions are associated with distinct patterns of cognitive processing, neural connectivity change, and behavioral outcomes. Prosody training (a) is associated with enhanced sensorimotor connectivity between MFG and PoCG, confirmed by a significant group × training interaction, with near transfer to structurally similar untrained material. Comprehension training (b) shows within-group IFG-MFG connectivity enhancement that did not reach significance in the between-group interaction test; effects were limited to trained material. Transcription control (c) shows no significant neural or behavioral changes. Line thickness reflects effect strength (strong, moderate, and weak). Solid circle borders indicate effects confirmed by a significant group × training interaction; dashed circle borders indicate within-group effects only (interaction not significant). Statistical annotations (green = significant; orange = non-significant) are shown beneath each network label.

These findings advance our understanding of how engagement mode shapes neural processing during poetry reading by providing longitudinal evidence consistent with the hypothesis that prosodic recitation drives training-related neural changes. Previous cross-sectional studies identified poetry’s neural correlates^1,3,11 but could not control for content differences or examine training-induced change. Our longitudinal design with strict material control demonstrates that engagement mode, not content alone, is associated with differential patterns of neural change.

Our findings also speak to debates about the role of sensorimotor systems in language comprehension. While some theorists emphasize language as abstract symbolic manipulation^30,31, others propose that linguistic meaning is grounded in sensorimotor experience^{21,22,32–34}. Our results are consistent with a nuanced position: while semantic comprehension can proceed without strong sensorimotor engagement, prosodic recitation appears to recruit sensorimotor networks more extensively, and this recruitment correlates with enhanced emotional responses. However, the present data cannot determine whether sensorimotor engagement is a necessary condition for emotional responses to poetry or one of several possible pathways. Notably, the comprehension group also showed enhanced appreciation and emotional arousal self-ratings for the trained poem despite no significant sensorimotor connectivity changes, suggesting that semantic processing alone can contribute to subjective affective engagement—albeit without the accompanying physiological arousal changes or transfer effects observed in the prosody group. This pattern is more consistent with a multiple-pathways account in which sensorimotor and semantic routes each contribute to different dimensions of emotional experience.

The implications extend beyond literary neuroscience to educational practice. While our findings cannot be directly generalized to classroom settings, they suggest that integrating structured recitation may simultaneously develop prosodic competence and enhance emotional engagement—an approach that could be particularly valuable for second-language learners struggling with native-like prosody³⁵. Our finding that even brief prosody training is associated with measurable neural changes suggests that relatively modest training interventions could yield detectable neural effects, though the long-term stability and practical significance of such changes remain to be established.

Several important limitations warrant consideration. First, our sample consisted exclusively of advanced Chinese EFL learners from university settings. While L2 learners may offer a sensitive model for detecting training effects due to their less automatized prosodic processing, this is also a substantial limitation. The observed effects may partly reflect prosodic training in a non-native language rather than general mechanisms of poetry processing³⁵. Future research should examine whether similar effects emerge in native speakers and across different language backgrounds. Second, this is a study of prosodic recitation training using poetry as the material, not a study of poetry per se. Although the null findings for expository text reading (no significant FC or RSA effects across all groups) confirm that training effects were specific to poetic materials at the assessment level, we did not include a non-poetry prosodic recitation training condition (e.g., rhythmic prose or song lyrics). We therefore cannot determine whether the observed training effects reflect properties unique to poetry or would also arise from prosodic training with other rhythmic materials. Our conclusions should be understood as pertaining to the effects of prosodic training during poetry reading, not to claims about what is unique to poetry itself. Third, our training protocol comprised only two 1-h sessions over 4 days. While sufficient to produce detectable neural changes, this brief intervention cannot address whether effects persist long-term or whether extended training might yield qualitatively different outcomes. The term “neural plasticity” as used in this study refers to short-term training-related changes and should not be interpreted as evidence of lasting structural reorganization. The absence of follow-up assessments leaves open critical questions about the stability of neural changes. Fourth, the prosody training condition differs from the other conditions not only in prosodic emphasis but also in the amount of vocal practice, rehearsal demands, and potential memorization involvement. We cannot fully separate the contributions of prosodic processing, motor practice, and memory to the observed effects. Future designs could include conditions that match vocal practice time without prosodic instruction to isolate these components. Finally, the fNIRS montage provided limited and asymmetric spatial coverage, with denser sampling over left-lateralized regions. This design precludes conclusions about right-hemisphere involvement, which prior literature suggests may be important for prosodic processing^36,37. The spatial resolution of fNIRS (~3 cm) also limits the precision of ROI-level interpretations, and the cortical renderings in our figures should not be taken to imply whole-cortex coverage.

In conclusion, we demonstrate that vocal recitation is associated with enhanced sensorimotor connectivity during poetry reading, and that these connectivity changes correlate with emotional responses, suggesting that engagement mode shapes both neural organization and affective experience. Our descriptive framework (Fig. 4) integrates these findings into a framework proposing how different modes of engagement may produce distinct patterns of neural change, with sensorimotor engagement serving as a potential pathway linking prosodic performance to emotional responses. While these findings contribute to our understanding of how language engagement recruits sensorimotor networks, important questions remain regarding the specificity of these effects to poetry, their generalizability beyond L2 learners, and their stability over time.

Methods

Participants

A priori power analysis was conducted using G*Power 3.1.9.7³⁸. For the primary 2 × 3 mixed design (group  × training), assuming f = 0.25 (medium effect), α = 0.05, and power = 0.80, a total of 42 participants were required. For paired-sample t-tests comparing pre-post connectivity changes within each group, we assumed a medium-to-large effect (d = 0.60) based on prior neuroimaging training studies^16–18, yielding a required sample of 19 participants per group. To ensure sufficient power for both the omnibus interaction test and within-group comparisons, 70 participants were recruited (25 prosody, 25 comprehension, and 20 transcription). The transcription group served as a passive control with expected minimal variance; its slightly smaller sample (n = 20) reflects this design consideration. All were advanced Chinese EFL learners (18 males; mean age = 24.0 years, range = 21–32; see Table 1) recruited from university communities. None were English literature majors or had prior training in English poetic prosody. Eligibility was confirmed through the College English Test Band 6 (CET-6: M = 512.2, SD = 52.5) and LexTALE vocabulary assessment (M = 67.6, SD = 10.0). Participants with prior exposure to the Raven or formal prosody instruction were excluded. They were randomly assigned to the three training groups. No participants were excluded during preprocessing, and no significant group differences were found in age, CET-6, or LexTALE scores (all ps > 0.10). All participants provided written informed consent and received monetary compensation. The study was approved by the Institutional Review Board of Nanjing Normal University (IRB No. NNU202501005). All ethical regulations relevant to human research participants were followed.

Table 1.

Participant characteristics by training group

Group		Age			CET-6			LexTALE
		M	SD		M	SD		M	SD
Prosody group Comprehension group Transcription group		24.00 24.20 23.65	2.18 1.85 1.09		522.36 513.92 497.25	56.53 47.96 51.84		69.48 66.56 66.45	12.31 8.60 8.44

Experimental design and materials

We employed a mixed factorial design with training (pre- vs. post-training) and text type (trained poem, untrained poem, and expository text) as within-subject factors, and group (prosody, comprehension, and transcription) as a between-subject factor.

Experimental materials comprised two poetic excerpts and one expository passage. The trained poem (Poem 1) consisted of the first six stanzas of Edgar Allan Poe’s the Raven (353 words). The untrained poem (Poem 2) comprised the subsequent six stanzas (362 words), used to assess whether training effects extend to novel but structurally comparable material. Because Poem 2 is drawn from the same work and shares identical prosodic structure (trochaic octameter, ABCBBB rhyme scheme, frequent internal rhymes, and alliterations), any observed transfer is best characterized as near transfer to structurally similar material rather than domain-general skill acquisition (see Discussion). To control for cortical changes arising from repeated exposure or training per se, we included an expository text for vocal reading: a 399-word passage from a CET-6 reading comprehension test on artificial intelligence. This design ensured that observed effects could be attributed to training type rather than material-specific differences.

Training procedures

Training consisted of two 1-h individual sessions conducted on consecutive days with a 1-day interval between sessions (see Fig. 5a). All training was delivered by the same experimenter to ensure consistency.

Fig. 5. Experimental design and fNIRS measurement setup.

a Overview of the 4-day experimental protocol. b Schematic illustration of the three training conditions: prosody training, comprehension training, and transcription control. c Detailed timeline of pre- and post-training assessment procedures. Following pronunciation familiarization and a 2-min resting-state recording, participants read three texts aloud. Each reading was limited to 4 min, with simultaneous recording of fNIRS, voice, and skin conductance data. Self-ratings of appreciation, comprehension, and emotional arousal were collected immediately after each poem. d Spatial distribution of fNIRS measurement channels over the cortical surface. Red dots indicate source optodes (emitters), blue dots represent detector optodes (receivers), and numbered labels (1–44) denote measurement channels formed by source-detector pairs. The probe configuration covered bilateral prefrontal, left temporal, and parietal regions implicated in language, aesthetic, and sensorimotor processing. All icons and schematic elements in (a–c) are self-made, original, and meet CC BY licensed, commercially reusable content standards.

Prosody training

Participants received instruction in fundamental English prosody concepts, including meter, rhyme schemes, and alliteration, then practiced rhythmic recitation of the Raven stanza-by-stanza (reciting 45 min in training session 1 and 50 min in training session 2). Each session combined listening to professional recordings, guided discussion with experimenters, and supervised oral practice with feedback until rhythmic accuracy was achieved (For training details, see Fig. 5b). We note that prosodic recitation inherently involves some degree of semantic processing and memorization; the training conditions are therefore best understood as differing in emphasis rather than achieving complete dissociation of prosodic from semantic processing (see Discussion for further consideration).

Comprehension training

This active control focused on semantic rather than prosodic aspects. Participants received contextual information about the poet and text, learned unfamiliar vocabulary, discussed meaning and emotional undertones of each stanza, and performed oral translations into Chinese with experimenter feedback. We note that this training condition did not specifically employ action-related content; the absence of motor-region engagement in this group may partly reflect the non-action nature of the semantic material.

Transcription training

This passive control required participants to copy the poem by hand in two different handwriting styles (Apple Chancery and Bradley Hand) without prosodic or semantic instruction, ensuring comparable material exposure while minimizing interpretive engagement.

Experimental procedure

The experiment comprised four stages across 4 days: pre-training assessment (day 1), two training sessions (days 2 and 3), and post-training assessment (day 4). Pre- and post-training sessions followed identical procedures to enable within-subject comparisons (see Fig. 5a).

Each assessment session began with pronunciation familiarization to prevent reading disruptions. Participants then received task instructions emphasizing attention to rhythm and rhyme when reading poems, with explicit permission to read at their natural pace. We acknowledge that instructing all participants to attend to rhythm and rhyme may have encouraged prosodic processing across all groups, potentially reducing the specificity of between-group contrasts. However, this instruction was held constant across groups, meaning that differential effects can still be attributed to the training manipulation.

Following probe placement and signal calibration, participants completed a 2-min resting-state recording (see Fig. 5c), then performed the main reading tasks in counterbalanced order (poems followed by expository texts for half of the participants; reverse order for the other half). For each poem, participants had 4 min to read aloud the six stanzas, with instructions that finishing early or not completing all stanzas was acceptable. Actual reading times were recorded. The expository text followed identical procedures but without rating scales. Comparable self-report measures were not collected for the expository text because the rating scales were specifically designed for aesthetic/emotional responses to poetry; applying them to expository text would have lacked construct validity.

Data acquisition

Behavioral measures

Participants completed prosody knowledge tests before pre-training and after post-training sessions, assessing understanding of English poetic meter, rhyme, rhythm, and alliteration. Each version contained seven comparable-difficulty items; the number of correct responses was recorded. Following each poem reading in pre- and post-training sessions, participants completed three 5-point Likert scales assessing appreciation, comprehension, and emotional arousal.

Voice recording

Oral reading performance was recorded using a directional shotgun microphone (Rode, NSW, Australia) connected to a Zoom F8n Pro recorder (Zoom, NY, USA) at 48 kHz, 16-bit resolution. The microphone was positioned directly before each participant, with adjustments ensuring consistent sound quality.

Skin conductance responses

Electrodermal activity indexed objective emotional arousal during pre- and post-training sessions. Signals were acquired at 2000 Hz using a BIOPAC MP150 system (BIOPAC Systems Inc., Goleta, CA) with isotonic gel electrodes attached to the left hand’s index and middle fingertips. Data were collected with AcqKnowledge 5.0 software.

fNIRS recording

Cortical hemodynamic responses were measured using a LABNIRS continuous-wave system (Shimadzu, Kyoto, Japan) with 16 sources and 16 detectors arranged at 3 cm separation, yielding 44 channels. The system operated with three wavelengths (780, 805, and 830 nm) at 120 ms temporal resolution (8.33 Hz sampling rate). Probes targeted bilateral prefrontal, left temporal, and parietal regions previously implicated in aesthetic experience, rhythm, and language processing (see Fig. 5d)^3,4,39. The montage provided asymmetric hemisphere coverage, with denser sampling over left-lateralized language and sensorimotor regions and more limited right-hemisphere coverage. This design choice was motivated by our focus on left-lateralized language-motor networks; however, it precludes strong conclusions about right-hemisphere involvement or whole-brain network organization (see Supplementary Table S1 for complete channel-to-ROI mapping with MNI coordinates and Brodmann area assignments). Channel positions were digitized with a FASTRAK 3D system (Polhemus, Vermont, USA); MNI coordinates were estimated with NIRS-SPM⁴⁰.

Data analysis

Statistics and reproducibility

All statistical analyses were conducted using R (version 4.2.1). For behavioral and acoustic measures, linear mixed-effects models (LMMs) were fitted using lmer() from the lmerTest package. For SCR count data, generalized linear mixed-effects models (GLMMs) with a Poisson distribution were fitted using glmer(). All tests were two-tailed with α = 0.05. Post hoc pairwise comparisons used the Tukey adjustment for multiple comparisons. Effect sizes are reported as beta coefficients (β) from model estimates. Functional connectivity analyses used linear mixed-effects models testing group × training interactions, with FDR correction across all ROI pairs. Representational similarity analyses employed Mantel tests with 1000 permutations. Randomization was performed by random assignment to training groups. Experimenters were not blinded to group assignment during training sessions, but acoustic and fNIRS data analyses were conducted using automated pipelines.

Behavioral data

For prosody knowledge tests and self-rating scales (appreciation, comprehension, and emotional arousal), we fitted linear mixed-effects models (LMMs) using the lmer() function from the lmerTest package in R (version 4.2.1). Training (pre- vs. post-training) was included as a within-subjects factor, and group (prosody, comprehension, transcription) as a between-subjects factor. The fitted model structure was: score ~ training × group + (1 | subject). Post hoc analyses were conducted using emmeans() for pairwise comparisons with Tukey adjustment for multiple comparisons.

Acoustic data

To assess whether participants’ reading became more rhythmic and fluent after training, all audio recordings were split into Poem 1 and Poem 2 using Audacity (v3.7.0), pre-segmented with Web-based MAUS⁴¹, then manually annotated in Praat, marking boundaries of stressed and unstressed syllables in the last line of the first five stanzas—selected as regions of interest (ROI) due to incomplete final stanza readings by some participants.

For global fluency analysis, we extracted total speaking time (onset to offset), number and duration of silent pauses, and calculated articulation time by subtracting total pause duration from speaking time. We derived: speaking rate (syllables per speaking time), articulation rate (syllables per articulation time), pause proportion (total pause time divided by speaking time), pause rate (syllables per pause), and average pause duration.

For local prosodic analysis, we focused on the same regions across the first five stanzas to extract syllable-level duration, mean pitch, and mean intensity, then computed stressed-to-unstressed (S/W) ratios—comparing stressed to unstressed syllables in duration, intensity, and pitch within each prosodic foot—as rhythmicity indicators, with higher ratios reflecting more distinct prosodic patterning.

We fitted LMMs using lmer(). Fixed effects included training (pre- vs. post-training), group (prosody, comprehension, transcription), and text type (Poem 1, Poem 2), plus interactions. Random effects initially included by-subject and by-item random intercepts and slopes for interactions. Models were simplified stepwise when non-convergent, first removing interaction terms, then random slopes. Model comparisons used AIC and likelihood ratio tests via anova(), with lower AIC indicating better fit.

SCR data

Acqknowledge 5.0 software (BIOPAC, Goleta, CA) was utilized for SCR analysis. We first conducted linear smoothing and low-pass filtering at 1 Hz to remove noise and high-frequency signals, then reduced the sampling rate from 2000 Hz to 7.8 Hz to increase processing speed. We constructed phasic EDA using the smoothing baseline removal method with a 5-s baseline estimation window width and 0.01 μS SCR threshold level. We extracted skin conductance response peak number (Event_Count) during 4-min reading periods, defined as total SCR events exceeding the threshold, as an emotional arousal indicator: higher Event_Count indicates more frequent physiological arousal and more active emotional responses during reading.

To examine whether training significantly influenced emotional arousal frequency, we fitted a GLMM in R using glmer(), employing Poisson regression accounting for the count data nature. Fixed effects included training (pre- vs. post-training), group (prosody, comprehension, and transcription), and text type (Poem 1, Poem 2, and expository text), plus interactions. Random effects initially included by-subject and by-item random intercepts and slopes for interactions. Post hoc pairwise comparisons of estimated marginal means used the emmeans package with Tukey-adjusted p-values.

fNIRS data

During preprocessing, raw fNIRS data were converted from text to snirf format, then processed in Homer3 (v1.86)⁴², a MATLAB-based open-source fNIRS analysis toolbox. Channels with excessively weak or strong signals were flagged using hmrR_PruneChannels (SNR threshold = 2). Raw intensity signals were transformed to optical density using hmrR_Intensity2OD. Motion artifacts were identified using hmrR_MotionArtifactByChannel (tMotion = 0.5 s, tMask = 1.0 s, STDEVthresh = 50, AMPthresh = 5) and corrected via cubic spline interpolation (hmrR_MotionCorrectSpline)⁴³. Bandpass filtering (0.01–0.1 Hz) was applied using hmrR_BandpassFilt to remove physiological noise. Optical density was converted to oxyhemoglobin (HbO) and deoxyhemoglobin (HbR) concentrations using hmrR_OD2Conc. Further residual motion artifact correction used correlation-based signal improvement (hmrR_MotionCorrectCbsi)⁴⁴. For each trial, the 30-s pre-reading onset period defined the baseline. The mean baseline value was subtracted from the reading period data. Participants received 4 min to complete each reading task, but due to variable completion times, analyzing the full 4-min window was inappropriate. Average reading duration across participants was 230 s; therefore, 0–230 s intervals were used for subsequent analyses. Complete preprocessing parameters and data quality metrics are reported in Supplementary Table S2.

Data quality and exclusion rationale

Our preprocessing strategy prioritized correction over rejection for three reasons. First, PLV requires continuous, uninterrupted time series; deleting segments would introduce artificial phase discontinuities. Second, each condition provided only ~230 s of data, making segment deletion costly for statistical power. Third, dual-correction (spline interpolation followed by CBSI) has been validated for overt speech fNIRS paradigms^43,44. The CBSI method leverages the expected negative HbO-HbR correlation to identify and correct motion-contaminated segments without data deletion.

NIRS-statistical parametric mapping partitioned 44 channels into 10 ROI based on the anatomical automatic labeling template (ICBM152)⁴⁵: left inferior frontal gyrus (left IFG), left middle frontal gyrus (left MFG), left dorsolateral superior frontal gyrus (left SFGdor), right dorsolateral superior frontal gyrus (right SFGdor), left inferior parietal lobule (left IPL), right inferior parietal lobule (right IPL), left postcentral gyrus (left PoCG), right postcentral gyrus (right PoCG), left supramarginal gyrus (left SMG), and left middle temporal gyrus (left MTG).

Functional connectivity analysis

We conducted functional connectivity (FC) analysis using preprocessed HbO concentration data. FC analysis was performed in MATLAB (R2020b). We selected the PLV⁴⁶ as our connectivity metric for three reasons. First, PLV captures phase synchronization, which directly reflects neural oscillatory coordination⁴⁷, and prior fNIRS studies have applied PLV to slow HbO signals to characterize functional connectivity^7,48,49. Second, PLV focuses exclusively on phase consistency and disregards amplitude fluctuations, making it inherently more robust to motion- and articulation-related amplitude noise than correlation-based measures—a property particularly advantageous for our overt speech task⁴⁶. Third, PLV tolerates constant cross-regional hemodynamic delays that would bias Pearson correlation estimates. Phase was extracted using the Hilbert transform on bandpass-filtered (0.01–0.1 Hz) HbO signals; the 230-s recording at 8.33 Hz (1917 time points) exceeds the recommended minimum for reliable PLV estimation⁴⁶.

PLV was computed for all possible ROI pairs, yielding ROI×ROI matrices (10 × 10 pairs) representing functional connectivity strength between each ROI pair for each participant across three training groups. To test whether training-induced connectivity changes differed across groups—the critical test of training specificity—we fitted linear mixed-effects models (LMMs) for theoretically motivated ROI pairs. For the primary MFG-PoCG analysis, the model was specified as: PLV ~ group × training + CET6 + LexTALE + (1 | Subject), where group (prosody, comprehension, transcription) was a between-subject factor, training (pre, post) was a within-subject factor, and CET6 and LexTALE scores were included as covariates to control for baseline language proficiency. A significant group × training interaction indicates differential training-induced connectivity changes across groups. Post-hoc pairwise comparisons at post-training were conducted with FDR correction. Within-group pre/post paired t-tests were also conducted for each ROI pair, with resulting p-values FDR-corrected across all ROIs. All statistical tests were two-tailed with α = 0.05. The analysis code, including the PLV computation function with full documentation and the FDR correction implementation (Benjamini–Hochberg procedure), is available at https://osf.io/9h5v2/.

Inter-subject representational similarity and regression analysis

To determine whether participants who showed greater training-induced neural changes also exhibited stronger emotional responses—a test of brain-behavior correspondence that cannot be addressed by standard subject-level regression—we employed representational similarity analysis (RSA)⁵⁰. IS-RSA preserves the full multivariate structure of individual differences by computing dissimilarity matrices from neural and behavioral data and testing whether these matrices are correlated, rather than collapsing across subjects^51–53.

First, as illustrated in Fig. 3a, we extracted preprocessed HbO time-series data (0–230 s) for each trial (Poem 1, Poem 2, and expository text) and each ROI for each participant under three training conditions, both before and after training. For each training group, we subtracted pre-training from post-training HbO time-series data to obtain vectors representing differences (post-pre) for Poem 1, Poem 2, and expository text under each training condition. For each ROI, we paired post-pre difference vectors for all participants and computed Euclidean distances, generating one representational dissimilarity matrix (RDM) per ROI and training condition. This yielded 90 neural RDMs (3 training groups × 3 text types × 10 ROIs), each represented as a participant × participant matrix.

Second, for SCR-recorded emotional frequency (Event_count), we similarly calculated differences (post-training minus pre-training) for Event_count under three text types (Poem 1, Poem 2, and expository text) to obtain pre-post differences in physiological emotional responses across three training groups. We computed pairwise Euclidean distances between all participants, generating 9 behavioral RDMs (3 training groups × 3 text types)⁵³. To compare RDMs, only upper (or lower) triangular elements were retained; diagonal values were removed to avoid redundancy.

Representational similarity was assessed using the Mantel test⁵⁴, a permutation-based approach evaluating correlation between two distance matrices. Specifically, two target RDMs were vectorized and their correlation coefficient computed (Spearman correlation was used). One or both vectors were then permuted, and the correlation coefficient was recalculated. After 1000 permutations, the observed correlation coefficient (r) was compared against the null distribution of permuted values. A significantly higher r than the permutation-based distribution indicates reliable representational similarity between two RDMs. Mantel tests were implemented using the vegan package in R.

To further investigate behavioral factors associated with neural representational similarity, we conducted linear mixed-effects modeling using lmer() from the lmerTest package in R. Our goal was to determine whether individual differences in specific behavioral dimensions—including prosody knowledge, appreciation level, comprehension level, and emotional arousal level—were associated with neural representational dissimilarity in ROIs. The dependent variable was the pre-to-post training representational difference for subject pairs within each ROI. Fixed effects included pairwise Euclidean distances between subjects in prosody level, appreciation ability, comprehension ability, and emotional arousal. Random intercepts for subject1 and subject2 were included to account for non-independence of dyadic data and repeated measures across pairings. Resulting p-values were FDR-corrected across all ROIs.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conceptualization, X.X. and X.Z. Methodology, X.X., C.J., and X.Z. Investigation, X.X., C.J., and X.Z. Formal analysis, C.J. Visualization, C.J. and X.X. Writing-original draft, X.X. and X.Z. Writing-review and editing, X.X. Supervision, X.X. Funding acquisition, X.X.

Peer review

Peer review information

Communications Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary handling editors: Zenas Chao and Benjamin Bessieres. A peer review file is available.

Funding

This work was supported by the National Social Science Foundation of China under Grant Nos. 24AYY021 (Key Project) and 21&ZD288 (Major Project).

Data availability

All stimuli (the full poem The Raven and the expository text), and behavioral, acoustic, physiological, and fNIRS data are available on the Open Science Framework at https://osf.io/9h5v2/. This study is not pre-registered.

Code availability

Custom analysis scripts for fNIRS data preprocessing, functional connectivity analysis (including the fully documented PLV computation function and FDR correction implementation), and inter-subject representational similarity analysis are available on the Open Science Framework at https://osf.io/9h5v2/.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains Supplementary material available at https://doi.org/10.1038/s42003-026-10664-4.