Time-frequency EEG markers of word boundaries in speech production

15 Speech production requires orchestration of multiple brain systems, including cortical and subcortical 16 areas that support the unfolding of the spoken message across hierarchical linguistic levels, such as 17 phonemes, syllables, words or phrases. Transitions between levels are critical for fluent speech, yet 18 the neural dynamics of, for example, syllable-level and word-level transitions remain unknown. In this 19 electroencephalography (EEG) study, we use time -frequency analysis and source localization to 20 determine differences associated with word-boundary vs. within-word syllable transitions. To this end, 21 pseudoword pairs comprising six consonant -vowel (CV) syllables with different word-boundary 22 positions were used. Fluent human adults produced the utterances at the rhythm of a learned visual 23 metronome (i.e., syllable-by-syllable), such that each syllable was uttered at matching times 24 independently of its relative word position. Accordingly, a target syllable could be either a within-word 25 syllabic transition or a between-word transition, while other linguistic properties, including articulation, 26 stress pattern , co-articulation or prosody, were matched. EEG time-frequency analyses of neural 27 sources successfully revealed sensitivity to hierarchical structure. Neural sources in left and right 28 inferior frontal lobes, as well as left superior temporal lobe were differentially recruited when producing 29 the same exact syllables, in the same exact utterance position, but under different word boundary 30 contexts. A right inferior frontal source showed a robust time-frequency modulation in word transitions 31 that included elevated event-related synchronization in the theta and beta range. Interestingly, despite 32 our efforts to control speech pace across conditions using metronome-based guidance, small, albeit 33 significant timing delays emerged, confirming higher cognitive demands at word boundaries. 34 35 36

speech production; electroencephalography; hierarchical structure 37 38 39 40 41 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint 1. Introduction 42 43 Language is an organised communication system in which smaller linguistic units are embedded 44 within larger structures: phonemes form syllables, which combine to form words, and words can 45 be used to build phrases and sentences (Rauschecker and Scott, 2009; Peelle and Davis, 2012). 46 This hierarchical structure is reflected in the neurobiology of both speech perception (Ding et al., 47 2015) and production (Sengupta and Nasir, 2016; Tremblay, Deschamps and Gracco, 2016; de 48 Heer et al., 2017) and supported by temporally distinct patterns of neural activity (Giraud and 49 Poeppel, 2012; Peelle and Davis, 2012) . A speech perception study (Ding et al., 2015) used 50 isochronous streams of syllables that could be grouped into words , phrases or sentences at 51 different fixed rates, and showed that cortical electrophysiological activity synchronised (i.e., 52 entrained) not only to the underlying syllabic rhythm but also to the word , phrase and sentence 53 rhythms. This frequency -tagging paradigm demonstrates that neural oscillations within posterior 54 temporal and inferior frontal cortices track the hierarchical linguistic structure of speech perception. 55 In other words, the brain may have special neural signatures at different levels of speech hierarchy. 56 Similarly, continuous speech production involves transforming abstract linguistic representations 57 into precisely timed articulatory actions, engaging cortical and subcortical regions . Foundational 58 neurocognitive models, such as the ‘Blueprint of the Speaker’ (Indefrey and Levelt, 2004) or the 59 dual-stream framework (Hickok and Poeppel, 2007; Hickok, 2012) also describe speech production 60 using a similar hierarchy. Overall, brain regions including motor and premotor cortices are 61 associated with motor planning and execution, auditory cortices with sensory-feedback prediction 62 and monitoring of speech outcomes (Houde and Jordan, 1998; Tian and Poeppel, 2010; 63 Christoffels et al., 2011), and basal ganglia and cerebellar structures with timing and sequencing 64 of motor programmes (Knolle et al., 2019; Archila-Meléndez et al., 2020; Jorge et al., 2022). These 65 regions may play different roles at different hierarchical levels of speech. 66 67 In motor-speech control, an efference copy (i.e., a forward model) is an internal representation of 68 the sensory consequences of a planned motor command (Crapse and Sommer, 2008) , which 69 enables the rapid evaluation of incoming sensory feedback for error detection and correction (Ford, 70 Roach and Mathalon, 2010; Niziolek, Nagarajan and Houde, 2013; Simmonds et al., 2014) . At 71 present, a key challenge of speech production is disentangling different levels of planning from 72 production proper and monitoring of predicted sensory consequences (Tremblay, Deschamps and 73 Gracco, 2016) . Producing any utterance involves a continuously updated planning scope that 74 spans upcoming segments and often the next word. Consequently, neural activity around a 75 speech-production boundary (e.g., a syllable or word) may reflect (i) syllabic sequencing and 76 articulatory chunking, (ii) lexical selection and word-level planning, and (iii) sensory predictions and 77 error monitoring, all partially overlapping in time. Hence, to study the full-scale neural processes 78 involved in speech production requires continuous speaking paradigms, while experimentally 79 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint isolating the different neural processes involved. Interestingly, debate remains on what are the 80 basic units of articulatory planning. Some studies suggest that syllables are the units of articulatory 81 planning (Laganaro, Valente and Perret, 2012; Bürki et al., 2016), whereas others postulate that 82 speech planning can proceed at the word level (Piai et al., 2015; Assaneo and Poeppel, 2018) . 83 84 Distinguishing between speech planning, monitoring and execution processes may also shed light 85 on the neural disruptions present in motor-speech disorders, including stuttering, where fluent 86 transitions between speech units are compromised (Indefrey, 2011; Chang and Guenther, 2020). 87 Interestingly, the relative position of speech gestures within words affects stuttering behaviour. 88 Studies have shown considerably higher dysfluency rates at word, phrase and sentence boundary 89 transitions, compared to within-word syllabic transitions (Bloodstein and Ratner, 2008; Buhr and 90 Zebrowski, 2009) . This dysfluency disproportion could arise from either encoding, prediction or 91 monitoring (Loucks et al., 2007; Ghazanfar and Eliades, 2014). Furthermore, the basal ganglia play 92 a critical role in motor control by modulating the cortex for motor initiation/execution, based on the 93 direct and indirect striatum -based pathways (Schulz et al., 2005) . These cortico-basal ganglia-94 thalamo-cortical circuits (aka, the cortico-basal ganglia loop) are essential for the execution of 95 motor commands, including motor speech gestures. Their role in stuttering has been suggested, 96 as well as, possible parallels between stuttering and Parkinson’s disease (PD) (Alm, 2004; Giraud 97 et al., 2008). Beyond the possible role of the direct and indirect pathways in stuttering, a more 98 recent link between stuttering and the basal ganglia has been proposed based on failures within a 99 faster cortico-basal ganglia loop that short -cuts the striatum, and instead stimulates the sub -100 thalamic nucleus directly - the hyperdirect pathway (HDP) (Neef, Anwander and Friederici, 2015). 101 The HDP leads to hyper-fast inhibitory signals to the cortex (Nambu, Tokuno and Takada, 2002). 102 Precision in inhibitory motor control via the HDP is suggested as necessary to stop or correct 103 ongoing motor actions, thereby contributing to sensorimotor control in speech production (Whillier 104 et al., 2018; Usler, 2022) . Importantly, the HDP may also be critically involved in motor/speech 105 initiation (Usler, 2022; Nambu et al., 2023) by resetting the primary motor cortex (M1) prior to the 106 arrival of the excitatory direct pathway that ultimately leads to muscle contraction (Nambu, 2011). 107 In turn, cortical inhibition by the HDP may be particularly important at higher hierarchical levels of 108 motor planning requiring M1 reset, including transitions between words, phrases and sentences. 109 Importantly, impairments along the HDP has been implicated in stuttering (Usler, 2022) . 110 111 Transitions between linguistic units, whether within a word ( e.g., syllable-to-syllable) or between 112 words (e.g., word-to-word), are hypothesized to engage distinct neural oscillatory mechanisms 113 (Cao, Thut and Gross, 2017; Cao et al. , 2024; Orpella et al. , 2024) . Prior studies have used 114 delayed-production paradigms (Kittilstved et al., 2018; Archila-Meléndez et al., 2020) , in which 115 speech planning and production proper (i.e., speech execution) are separated by an experimental 116 cue. While this approach helps to dissociate planning- from execution-related activity, it limits our 117 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint understanding of how neural dynamics support online transitions during continuous speech 118 production. In addition, while it is possible to instruct participants to produce isolated speech units 119 of different hierarchical levels (e.g., isolated syllables versus words) (Silbert et al., 2014), probing 120 these mechanisms in sequential/continuous speaking tasks is crucial. For instance, producing the 121 syllable ba in different contexts: i) in isolation; ii) as the first syllable in the word banana; or iii) as 122 the final syllable in the word scuba, may involve distinct neural processes (from planning to 123 monitoring) for the ba component, despite requiring the same broad articulatory gesture s. 124 However, these conditions hold co-articulatory differences that confound how the syllable ba, in 125 this example, is produced in the sequence context. If, instead, target syllables are embedded in a 126 syllabic sequence with matching co-articulatory conditions, but under different word-boundaries, it 127 becomes possible to study hierarchical speech production processes in a controlled manner. We 128 attempt to provide an analogy of how we to test this question using English language: if the word 129 ‘scu’ and ‘nana’ would be legal english words, the sequences ‘scuba nana’ and ‘scu banana’ would 130 be valid phrases allowing to study how the target syllable ba plays different hierarchical roles in the 131 sequences, while preserving its phonological/phonetic form and co-articulatory effects. In practice, 132 working with pseudoword pairs enable s the construction of controlled stimulus sequences (that 133 occur across different syllable positions) across experimental conditions, which we use in this 134 study. 135 EEG time-frequency analyses, such as event-related spectral perturbations (ERSP) allow tracking 136 transient power changes across frequency bands in response to time-locked experimental events 137 (Delorme and Makeig, 2004), which can be used to separate hierarchical transitions and explore 138 the neural dynamics of speech production (Vos et al., 2010; Jenson et al., 2014). Beta oscillatory 139 activity (~15-30 Hz) over frontal cortices - often prominent over right inferior frontal and premotor 140 regions - has been associated with preparatory set/maintenance and sequential control during 141 speech and orofacial actions (Pfurtscheller and Lopes da Silva, 1999; Weiss and Mueller, 2012). 142 Converging evidence indicates a functional subdivision whereby low beta (~1 5-20 Hz) is more 143 closely associated with the indirect basal ganglia pathway whereas high beta (~20-30 Hz) reflects 144 hyperdirect cortico-subthalamic signals, coupling the premotor cortex (e.g., inferior frontal regions 145 - IFG, specially the right IFG) and the supplementary motor areas (SMA) with the STN (Oswal et 146 al., 2021; Herz et al., 2023; Cao et al., 2024). Changes in time-frequency provide neural signatures 147 for different aspects of speech production, which may help unveil the neurobiological failures that 148 contribute to stuttering, mainly around word-boundaries. Clarifying how oscillatory signals encode 149 word-to-word transitions in continuous speech may therefore open new venues to understand 150 different neural dysfunctions that can contribute to future research on motor-speech disorders 151 (Neef et al., 2018). 152 In the present study we investigate EEG time-frequency changes during paced overt speech 153 production to compare within-word and between -word syllabic transitions , in the context of 154 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint pseudowords. Fluent p articipants produced pseudo-word pairs forming a sequence of six 155 consonant-vowel (CV) syllables, at the pace of a visual metronome. Importantly, pseudo-word pairs 156 consisted of either a two-syllable word followed by a four-syllable word (i.e., condition 2+4) or two 157 three-syllable words (i.e., condition 3+3). This experimental design enabled the direct comparison 158 of neural activity at syllable-to-syllable and word-to-word transitions in a controlled and time-locked 159 manner. In turn, p articipants produced the same exact syllables, in the same exact sequence 160 position, but under different relative positions within a word ( i.e., first versus last syllable). 161 Accordingly, we focus on EEG time-frequency changes time-locked to the production of the third 162 syllable, as this syllable is an event that corresponds to a word -to-word transition in the 2+4 163 condition and a within-word syllable-to-syllable transition in 3+3 condition. 164 165 2. Methodology 166 167 2.1 Participants 168 Twenty-one fluent Portuguese-speaking adults (14 females, 5 males), aged between 18 and 53 169 years (Mean ± Standard Deviation (SD) = 29.4 ± 9.6), participated in the study. Two participants 170 were excluded due to technical errors during the EEG-audio recordings. All participants were right-171 handed, had completed secondary education or higher education (bachelor’s, master’s, or doctoral 172 level) and reported normal hearing and no history of psychiatric, neurological, or language-related 173 disorders. Participation was voluntary, and no compensation was given for taking part in the study. 174 The study was approved by the ethical committee of University of Algarve, Portugal , and 175 performed in accordance with the Declaration of Helsinki and Oviedo Convention. All participants 176 provided written informed consent prior to testing. 177 178 2.2 Stimuli 179 Stimuli consisted of sequences of two pseudo-words designed to isolate neural activity related to 180 syllable- and word-level transitions (see Figure 1B). Each sequence contained two pseudo-words 181 and conformed to one of two structures: condition 2+4 (e.g., náfa dacalána), in which a disyllabic 182 pseudo-word was followed by a tetrasyllabic pseudo-word, and condition 3+3 (e.g., náfada calána), 183 in which two trisyllabic pseudo-words were presented. Both structures were matched for length (six 184 syllables per trial). All syllables followed a CV structure ( ‘ba’, ‘da’, ‘fa’, ‘ma’, ‘na’, ‘sa’, ‘ta’, ‘ca’, ‘la’, 185 and ‘ga’) corresponding to the International Phonetic Alphabet (IPA) consonant sounds /b/, /d/, /f/, 186 /m/, /n/, /s/, /t/, /k/,/l/, and /g/ followed by the vowel /a/. We minimised bilabial closures at and before 187 the target syllable 3 to reduce perioral EMG while maintaining clear tongue -tip versus tongue-188 dorsum articulations. Labial consonants were restricted to speech positions after syllable 3; i.e., no 189 labial articulations occurred at syllable position 2 or 3 to reduce EMG artifacts on the EEG signal 190 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint (Whitham et al., 2007; Stepp, 2012). The third syllable was always tongue-driven (using /d/ or /k/ 191 consonants). The pseudo-words were phonotactically legal in Portuguese and carried stress on 192 the first syllable of the first pseudo-word and the penultimate syllable of the second pseudo-word, 193 mirroring natural stress patterns while avoiding lexical or semantic familiarity effects (Arruda et al., 194 2023). Hence, the stress pattern of the individual syllables was the same for all stimuli (and across 195 conditions). Finally, written stimuli were presented in lowercase text using a uniform sans-serif font 196 (size 24), centred on the screen with a grey background using Presentation® software 197 (Neurobehavioral Systems, Inc., Berkeley, CA, USA; www.neurobs.com). 198 2.3 Experimental Procedure 199 The experiment took place in an electrically shielded EEG laboratory at the Cognitive Neuroscience 200 Group of University of Algarve. EEG data were recorded using a 128-channel BioSemi ActiveTwo 201 system (BioSemi B.V., Amsterdam, the Netherlands) with electrodes positioned according to the 202 international 10/20 system. The vertex sensor (i.e., Cz) was centred between the inion and the 203 nasion anatomical fiducials on the sagittal plane, and the midpoint between the ears on the coronal 204 plane. Conductive gel was applied to each electrode site using a blunt-tipped plastic syringe, and 205 electrode impedance was verified to minimize offsets, in line with BioSemi specifications. Once 206 EEG preparation was completed, participants seated comfortably in front of a computer screen at 207 approximately 70 cm viewing distance. 208 209 The experiment consisted of 160 trials in total, divided into 4 blocks of 40 trials each. Each block 210 contained five repetitions of 4 sequences from condition 2+4 and 4 of condition 3+3, fully 211 randomised. Within a block, the same sequence was never presented under both conditions. 212 Hence, blocks differed in how syllables were grouped, providing variation while preserving the 213 same underlying syllabic material across blocks. The order of blocks was counterbalanced across 214 participants, ensuring that all participants completed both conditions under comparable exposure 215 and repetition constraints . Following EEG cap placement and impedance checks, participants 216 began each trial with a practice phase followed by a paced production using a visual metronome 217 (see Figure 1A). The practice phase had the target pseudo -word pair written on the screen, and 218 participants repeatedly uttered the target sequence three times. Upon completion of the trial 219 practice phase, participants pressed the ‘space bar’ button to signal their readiness to proceed to 220 the paced production phase. In the paced production phase, the written sequence disappeared 221 from the screen, and a green fixation cross appeared, flashing every 700 ms (1.43 Hz). Participants 222 were instructed to produce one syllable per flash, aligning their speech with the visual rhythm. 223 Because the lexical boundary shifted across sequence structures, the transitions flanking syllable 224 3 differed by condition. In condition 2+4, the transitions between syllables 2 and 3 were between-225 word transitions and in condition 3+3, were within-word transitions. Conversely, the transitions 226 between syllables 3 and 4 were within -word transitions in condition 2+4 and between -word 227 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint transitions in condition 3+3. This counter balancing enabled within-item contrast of between-word 228 vs within-word syllabic transitions while controlling segmental content and pacing . 229 230 231 Figure 1 232 A) Experimental design. ITI – inter-trial-interval. B) Stimuli. Two sequence types were used: Condition 2+4 233 and Condition 3+3. C) Behavioural results corresponding to speech onsets flanking syllable 3. Red and green 234 dots depict individual trial syllable speech timings, relative to syllable 3, per subject (red = syllable 2; green 235 = syllable 4); vertical lines mark the metronome timings. Mean (SD) speech timings are shown at the top. D) 236 Averaged speech-envelope responses. Group speech envelope traces time-locked to syllable 3 (grey = 237 individual participants; coloured = condition averages). E) Regions of interest. Bilateral inferior frontal gyrus 238 (IFG), superior temporal gyrus (STG), and supplementary motor areas (SMA) used for hypothesis-driven 239 visualisation in ICA cluster selection. 240 241 242 243 244 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint 2.4 Data Recording and Pre -processing 245 Speech responses were simultaneously recorded with EEG at 16384 Hz (16-bit resolution) using 246 a Shure SM58 microphone placed approximately 15 cm from the participants’ mouth. Continuous 247 EEG was recorded from the 128-channel BioSemi ActiveTwo system (BioSemi B.V., Amsterdam, 248 the Netherlands), with a sampling rate of 512 Hz. Electrodes were arranged according to the 249 international 10/20 system. The BioSemi CMS/DRL system served as the reference and ground 250 during acquisition. Audio data were processed to identify speech response onsets. First, speech 251 recordings were subjected to a Hilbert transformation, and the resulting amplitude envelope was 252 low-passed filtered at 8 Hz and resampled to 256 Hz. Within each trial, six local envelopes (one 253 per syllable) were detected to provide syllable-level temporal markers for identifying the production 254 of the 6 syllables composing our stimuli. Custom Matlab routines were used for audio processing. 255 EEG preprocessing was performed in EEGLAB (Delorme and Makeig, 2004), including custom-256 made MATLAB scripts (version R2021b; MathWorks, Natick, MA). The continuous EEG recordings 257 were i) resampled to 256 Hz and ii) high-pass filtered at 1 Hz to remove slow signal drifts. iii) Noisy 258 and flat-line channels were identified and removed using the ‘clean_rawdata’ EEGLAB function, 259 after which data were iv) re-referenced to the average of all remaining channels. Across datasets, 260 a mean of 97.9 channels ( SD = 9.6) remained following bad -channel rejection. Next, v) 261 independent component analysis (ICA; INFOMAX) was performed, and components reflecting 262 ocular, muscular, or other stereotypical artefacts were automatically identified and removed using 263 ‘ICLabel’ and ‘ICFlag’ EEGLAB functions (Pion-Tonachini, Kreutz-Delgado and Makeig, 2019). vi) 264 Channels previously excluded were subsequently interpolated back into the dataset using the 265 ’interp’ method. vii) Following ICA cleaning, the data were low-pass filtered at 70 Hz and a band-266 stop (48-52 Hz) filter was applied to further suppress line-power noise (i.e., 50 Hz energy). Next, 267 we segmented the continuous EEG into epochs from -1 to +1 s ecs centred on the third -syllable 268 production event. Time-locking our EEG analysis to the third syllable minimises contamination from 269 sequence-onset preparatory motor activity, while still providing symmetric context for preceding 270 and following syllable transitions. Although the speech envelope peak can lag the true acoustic 271 onset by a small amount, the same alignment procedure and fixed pacing were used in both 272 conditions, and behavioural timing (Figure 1D) confirmed consistent alignment across participants. 273 Event-related spectral perturbations (ERSPs) were computed for each trial and independent 274 component (IC), providing a time-frequency representation of oscillatory time-frequency changes 275 across conditions. 276 277 278 2.5 Data Analysis and Statistics 279 280 Independent Component Analysis and Source Clustering 281 282 Equivalent current dipoles were estimated for each independent component using a standard 283 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint boundary element head model. Components with residual variance greater than 15% were 284 excluded. Next, dipoles were computed for each IC component and participant. Finally, 20 clusters 285 were made by matching dipole locations and spectral activity between 3 and 40 Hz (see Figure 2). 286 Cluster sizes ranged from 19 to 197 ICs (M = 93.1, SD = 40.0). With this approach, we took 287 advantage of the high-density EEG montage (i.e., 128 channels), by performing group statistics at 288 the EEG source level instead of relying on EEG -channel signals, which are known to have low 289 spatial specificity (Delorme & Makeig, 2004 ). 290 291 292 Figure 2 293 Independent Component (IC) clusters, and centroid location (red sphere) in Talairach coordinates. Blue 294 spheres are individual ICs forming each cluster. 295 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint Time-Frequency Analysis 296 297 EEG was segmented into epochs from -1 to +1 secs aligned to the speech response time of the 298 third syllable (target syllable) of the sequence. Time-locking our EEG analysis to the third syllable 299 minimises contamination from sequence -onset preparatory motor activity, while still providing 300 symmetric context for preceding and following syllable transitions. Although the speech-envelope 301 peaks may lag true acoustic onset by a small amount, the same alignment procedure and fixed 302 pacing were used in both conditions, and behavioural timing (Figure 1D) confirmed consistent 303 alignment across participants. Event-related spectral perturbations (ERSPs) were computed for 304 each trial and independent component (IC), providing a time -frequency representation of time-305 frequency power changes across conditions. 306 307 ERSPs were averaged within each cluster and assessed statistically using a permutation test 308 (based on 2,000 condition -label permutations). For hypothesis -driven visualisation, we report 309 clusters corresponding to bilateral IFG, bilateral STG, and bilateral SMA (Figure 1E); example 310 ERSPs are shown for L-IFG (Cluster 1), R-IFG (Cluster 3), and L-STG (Cluster 8) (see figure 3). 311 ERSPs for all 20 clusters are provided in figure 5. Multiple comparisons over the time-frequency 312 bins were corrected using FDR (q = 0.05) (Benjamini and Hochberg, 1995) (i.e., 200 time-points 313 by 68 frequencies). Behavioural contrasts were evaluated with two-tailed, paired t-tests. We refer 314 to decreases in band-limited power as event-related desynchronisation (ERD) and increases as 315 event-related synchronisation (ERS), reporting frequency bands and time windows relative to the 316 epoch average (Delorme and Makeig, 2004). 317 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint 318 319 Figure 3 320 ERSPs in the three clusters matching the selected ROIs (Clusters 1, 3, and 8). First and second left time-321 frequency plots depict the grand -averaged ERSPs per condition. Third and fourth plots depict ERSP 322 statistics: uncorrected effects (p<0.05) and FDR-corrected (q = .05). A) left-IFG cluster. B) right-IFG cluster. 323 C) left-STG cluster. 324 325 326 3 Results 327 328 3.1 Behavioural Results 329 330 Participants successfully aligned their speech production with the visual metronome (see Figure 331 1C-D). Timing between successive syllables around syllable 3 differed slightly, but significantly by 332 condition. The interval between syllable 2 and 3 was significantly shorter in the 3+3 condition in 333 comparison to the 2+4 condition, mean (SD) = 17 ms (34.49), t(18) = 2.15, p = .045. The interval 334 between syllable 3 and 4 was longer in 3+3 than in 2+4, mean (SD) = 29 ms (43.75), t(18) = -2.90, 335 p = 0.010. Group speech-envelope traces time-locked to syllable 3 confirmed consistent alignment 336 of the third -syllable envelope peak across participants and conditions (Figure 1D). 337 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint 338 339 3.2 Time-frequency results (Event -Related Spectral Perturbations) 340 341 Time-frequency differences were analysed within the three ROI clusters (Clusters 1, 3, and 8; see 342 Figure 2). EEG statistics are based on permutation -based condition tests (2,000 label 343 permutations) with FDR for multiple comparisons correction, across time-frequency bins (q = .05). 344 Consistent with our analysis plan, bilateral IFG, bilateral STG, and bilateral SMA were treated as 345 confirmatory ROIs (FDR correction within time-frequency bins). No other ROI clusters yielded FDR-346 corrected effects. For completeness, ERSPs for all 20 clusters are shown in figure 5. 347 348 The left IFG (Cluster 1; tal: -41, 46, 6) showed largely similar ERSPs across conditions, with only 349 a small, isolated FDR -corrected significant alpha patch (~ 10 -14 Hz) about 450 ms prior to the 350 target syllable 3 (Figure 3A). Broader uncorrected differences (p<0.05) did not survive FDR 351 correction. In the right IFG (Cluster 3; tal: 49, 39, 9), a robust pre -pk3 modulation survived FDR 352 correction, spanning ~ 4 -30 Hz (theta/alpha into low beta; Figure 3B, FDR panel). Power 353 modulations were greater for the 2+4 condition in comparison to the 3+3 condition, with additional 354 weaker patches evident at the uncorrected level in the early post-syllable-3 window. In the left STG 355 (Cluster 8; tal: -59, -39, 3), we observed two brief FDR-significant time-frequency patches (figure 356 3C): a pre-syllable 3 alpha patch (~ 9-11 Hz, [-400 to -300] msecs) and a short post-syllable 3 high-357 beta patch (~ 26-30 Hz, [0 to +100] msecs). Outside these windows, no sustained corrected effects 358 were present. The right STG showed no FDR -corrected differences (see figure 3 and 4 ). 359 Considering all 20 clusters, the clearest FDR-corrected modulation emerged in right IFG, indicating 360 a right-lateralised inferior frontal preparatory signature preceding the target syllable 3. Left IFG and 361 left STG contributed only limited, and no other clusters yielded FDR-corrected effects. See figure 362 4 for a summary of the significant ERSP contrasts in the confirmatory ROIs. 363 364 365 366 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint 367 368 Figure 4 369 ERSP summary per frequency band (delta, theta, alpha, low beta and high beta) per cluster (cluster 1 – left 370 IFG, cluster 2 – right IFG, cluster 8 – left STG). Blue line is condition 2+4 and red line is condition 3+3. 371 Shaded areas around the blue and red lines represent standard errors from the group mean. Black horizontal 372 bars on top of each cluster-band plot depict uncorrected p values (p<0.05), and the yellow line depict FDR 373 corrected p values (q<0.05). 374 375 376 377 4 Discussion 378 379 Although language is a hierarchically organized communication system, segmental boundaries in 380 natural speech are often concealed due to coarticulation. The perceptual challenge for a listener is to 381 recover speech boundaries from continuously presented speech , which involves specific neural 382 oscillations (Ding et al., 2015). In speech production , the same hierarchical levels become relevant 383 for understanding motor speech disorders such stuttering, but their neural signatures remain 384 underspecified (Neef et al., 2015). Here, we investigated differences in cortical oscillations across 385 different levels of hierarchy -syllable forming word boundary vs. within -word syllable boundary . 386 Clustering of ICA components , obtained from high-density EEG, identified three clusters within our 387 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint regions of interest , which showed significantly different time-frequency signatures when comparing 388 within-word and between-word syllabic transitions: the left and right IFG and the left STG (Figure 2). 389 The importance of these clusters has been previously identified in the literature, as a distinguishing 390 part of the language and speech network that supports speech planning versus monitoring, 391 respectively (Hickok, 2012; Tremblay, Deschamps and Gracco, 2016). Across these clusters, the most 392 robust differences in modulation were in the right IFG spanning both theta and beta oscillatory bands 393 (Figure 3). 394 We observed a robust effect spanning theta and beta oscillations around 500 ms prior to the target 395 syllable, which was a between-word transition in the 2+4 condition and a within-word syllabic transition 396 in the 3+3 condition. This effect indicates that frontal control networks adjust preparatory states 397 according to the hierarchical role of the upcoming syllable, maintaining within-word chunking versus 398 reconfiguration at a boundary. This pattern aligns with reports that right-frontal beta oscillations index 399 preparatory and sequential control during speech (Pfurtscheller and Lopes da Silva, 1999; Weiss and 400 Mueller, 2012; Piai et al., 2015). Furthermore, theta oscillations may contribute to joint operations with 401 beta rhythms, namely in sensory prediction (Arnal and Giraud, 2012) or cortico-basal ganglia circuits 402 (Oswal et al., 2021) . Importantly, the right inferior frontal cortex is central in inhibitory control via 403 hyperdirect projections to the subthalamic nucleus, which is expressed in beta neural rhythms 404 (Nambu, 2011; Jorge et al., 2022). 405 406 These findings have important implications for future research on motor speech disorders, particularly 407 stuttering (Rocha, Carmona and Correia, 2025) . In stuttering, sensorimotor processes that support 408 fluent transitions between speech units are often disrupted, leading to higher dysfluency rates at word, 409 phrase, and sentence boundaries compared to within-word syllabic transitions (Bloodstein and Ratner, 410 2008; Buhr and Zebrowski, 2009). Neuroimaging studies indicate that atypical function in the cortico-411 basal ganglia circuits, including both the direct and indirect pathways, may contribute to speech 412 disruptions (Alm, 2004; Giraud et al., 2008), and recent work highlights the HDP as crucial for inhibitory 413 control and preparatory motor regulation in speech (Nambu, 2011; Neef et al., 2015) . The HDP 414 conveys rapid signals from frontal inferior cortical regions to the STN, enabling stopping and 415 adjustment of ongoing motor commands, which supports sensorimotor integration and accurate 416 coordination of speech movements (Whillier et al., 2018; Usler, 2022) . Dysfunction in these circuits 417 may impair the ability to reset or prepare the motor cortex at motor planning boundaries (e.g., word 418 boundaries), providing a possible explanation for why stuttering often emerges at transitions between 419 hierarchical speech units. 420 421 Although the focus of this study was on the segmental level of speech production, prosodic information 422 can also define word boundaries. In this study, the syllables marking a word boundary also mark a 423 difference in prosody. Other studies implicate the right IFG in the perceptual categorization of 424 emotionally-valenced prosodic information (Sammler et al., 2015; Zhang et al., 2026). The right STG 425 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint is also implicated in prosody and modulated by its emotional salience (Sammler et al., 2015), however, 426 right STG condition-differences did not emerge in this study. One possibility is that the right IFG is part 427 of a dorsal stream (Hickok and Poeppel, 2007; Ausilio, Craighero and Fadiga, 2012; Correia, Jansma 428 and Bonte, 2015) associated with sensorimotor transformations, which can include information about 429 prosody and may therefore inherently contribute to decisions about word boundaries. Nevertheless, 430 the right IFG emerged in the comparison between the 2+4 and 3+3 conditions, where stress pattern 431 was held constant. Hence, prosody differences cannot fully account for the right IFG activation 432 differences found between conditions, here. The results are more likely consistent with a role for the 433 right IFG in a basal-ganglia cortical loop that supports segmental sequences forming larger units, 434 namely word units. 435 Several possible limitations should be considered. First, the sample size (N=19), although typical for 436 EEG studies may limit sensitivity to smaller effects between condition 2+4 and 3+3, especially given 437 the need to correct for multiple time -frequency bins tested per IC cluster. Second, within -word 438 transitions were 17 milliseconds shorter between syllable s 2 and 3, and 29 m illiseconds shorter 439 between syllables 3 and 4. These timing differences may also contribute to the ERSP estimates. Third, 440 paced speech constrains generalization of the findings to more natural and prosodically richer speech 441 production; but strict pacing and pseudo-words enhance control and timing, allowing a more controlled 442 analysis of EEG data in epochs less likely contaminated by speech responses (i.e., speech -related 443 EMG).. Finally, while the most parsimonious interpretation of the results in the right IFG cluster is 444 related to the cortico-basal ganglia circuit underlying the HDP, we cannot fully rule out other factors 445 confirming this as its sole role. Scalp EEG has limited coverage for subcortical regions (i.e., reduced 446 signal-to-noise-ratio from deep brain sources) and thus does not provide a direct measure . In the 447 future, simultaneous extracranial and intracranial EEG recordings that more directly target subcortical 448 regions (e.g., from the STN in PD patients undergoing deep -brain stimulation - DBS) can provide 449 additional insights into the interpretation of the results. Altogether, the behavioural and neural data 450 indicate that hierarchical structure shapes both the timing and pre-speech cortical dynamics of speech 451 production. The right -lateralised IFG oscillatory signature preceding the target syllable suggests 452 predictive control mechanisms tuned to whether the upcoming syllable completes or starts a word, 453 complementing accounts of hierarchical temporal scaffolding in speech perception and production 454 (Ghitza, Giraud and Poeppel, 2013; Ding et al., 2015; Piai et al., 2015). 455 456 457 458 459 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint 460 461 462 Figure 5 463 ERSPs in all the 20 clusters (see Figure 2 for cluster locations). First and second left time-frequency plots 464 depict the grand -averaged ERSPs per condition. Third and fourth plots depict ERSP statistics based on 465 permutation testing (2,000 label permutations): third plot is uncorrected effects (p<0.05), and fourth plot is 466 FDR-corrected ( q = .05). 467 468 469 470 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint

471 Alm, P. A. (2004) ‘Stuttering and the basal ganglia circuits : a critical review of possible relations’, 472 37, pp. 325 –369. doi: 10.1016/j.jcomdis.2004.03.001. 473 Archila-Meléndez, M. E. et al. (2020) ‘Combining Gamma With Alpha and Beta Power Modulation 474 for Enhanced Cortical Mapping in Patients With Focal Epilepsy’, Frontiers in Human Neuroscience, 475 14, p. 555054. doi: 10.3389/fnhum.2020.555054. 476 Arnal, L. H. and Giraud, A. L. (2012) ‘Cortical oscillations and sensory predictions’, Trends in 477 Cognitive Sciences. Elsevier Ltd, 16(7), pp. 390 –398. doi: 10.1016/j.tics.2012.05.003. 478 Assaneo, M. F. and Poeppel, D. (2018) ‘The coupling between auditory and motor cortices is rate-479 restricted: Evidence for an intrinsic speech -motor rhythm’, Science Advances, 4(2). doi: 480 10.1126/sciadv.aao3842. 481 Ausilio, A. D., Craighero, L. and Fadiga, L. (2012) ‘The contribution of the frontal lobe to the 482 perception of speech’, Journal of Neurolinguistics . Elsevier Ltd, 25(5), pp. 328 –335. doi: 483 10.1016/j.jneuroling.2010.02.003. 484 Benjamini, Y. and Hochberg, Y. (1995) ‘Controlling the False Discovery Rate: A Practical and 485 Powerful Approach to Multiple Testing’, Journal of the Royal Statistical Society Series B: Statistical 486 Methodology , 57(1), pp. 289 –300. doi: 10.1111/j.2517 -6161.1995.tb02031.x. 487 Bloodstein, O. and Ratner, N. B. (2008) A Handbook on Stuttering . New York: Delmar. 488 Buhr, A. and Zebrowski, P. (2009) ‘Sentence position and syntactic complexity of stuttering in early 489 childhood : A longitudinal study’, 34, pp. 155 –172. doi: 10.1016/j.jfludis.2009.08.001. 490 Bürki, A. et al. (2016) ‘Sequential processing during noun phrase production’, Cognition, 146, pp. 491 90–99. doi: 10.1016/j.cognition.2015.09.002. 492 Cao, C. et al. (2024) ‘Low-beta versus high-beta band cortico-subcortical coherence in movement 493 inhibition and expectation’, Neurobiology of Disease , 201, p. 106689. doi: 494 10.1016/j.nbd.2024.106689. 495 Cao, L., Thut, G. and Gross, J. (2017) ‘The role of brain oscillations in predicting self -generated 496 sounds’, NeuroImage , 147. doi: 10.1016/j.neuroimage.2016.11.001. 497 Chang, S. E. and Guenther, F. H. (2020) ‘Involvement of the Cortico -Basal Ganglia-498 Thalamocortical Loop in Developmental Stuttering’, Frontiers in Psychology , 10(January). doi: 499 10.3389/fpsyg.2019.03088. 500 Christoffels, I. K. et al. (2011) ‘The sensory consequences of speaking: Parametric neural 501 cancellation during speech in auditory cortex’, PLoS ONE, 6(5). doi: 502 10.1371/journal.pone.0018307. 503 Correia, J. M., Jansma, B. M. B. and Bonte, M. (2015) ‘Decoding articulatory features from fmri 504 responses in dorsal speech regions’, Journal of Neuroscience , 35(45), pp. 15015 –15025. doi: 505 10.1523/JNEUROSCI.0977-15.2015. 506 Crapse, T. B. and Sommer, M. A. (2008) ‘Corollary discharge across the animal kingdom’, Nature 507 Reviews Neuroscience, 9(8), pp. 587 –600. doi: 10.1038/nrn2457. 508 Delorme, A. and Makeig, S. (2004) ‘EEGLAB: an open source toolbox for analysis of single -trial 509 EEG dynamics’, Journal of Neuroscience Methods , 13, pp. 9–21. 510 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint Ding, N. et al. (2015) ‘Cortical tracking of hierarchical linguistic structures in connected speech’, 511 Nature Neuroscience , 19(1), pp. 158 –164. doi: 10.1038/nn.4186. 512 Ford, J. M., Roach, B. J. and Mathalon, D. H. (2010) ‘Assessing corollary discharge in humans 513 using noninvasive neurophysiological methods’, Nature Protocols, 5(6), pp. 1160 –1168. doi: 514 10.1038/nprot.2010.67. 515 Ghazanfar, A. A. and Eliades, S. J. (2014) ‘The neurobiology of primate vocal communication’, 516 Current Opinion in Neurobiology. Elsevier Ltd, 28, pp. 128–135. doi: 10.1016/j.conb.2014.06.015. 517 Ghitza, O., Giraud, A. L. and Poeppel, D. (2013) ‘Neuronal oscillations and speech perception: 518 Critical-band temporal envelopes are the essence’, Frontiers in Human Neuroscience, 6(JAN), pp. 519 1–4. doi: 10.3389/fnhum.2012.00340. 520 Giraud, A. L. et al. (2008) ‘Severity of dysfluency correlates with basal ganglia activity in persistent 521 developmental stuttering’, Brain and Language , 104(2), pp. 190 –199. doi: 522 10.1016/j.bandl.2007.04.005. 523 Giraud, A. L. and Poeppel, D. (2012) ‘Cortical oscillations and speech processing: Emerging 524 computational principles and operations’, Nature Neuroscience . Nature Publishing Group, 15(4), 525 pp. 511–517. doi: 10.1038/nn.3063. 526 de Heer, W. A. et al. (2017) ‘The hierarchical cortical organization of human speech processing’, 527 Journal of Neuroscience , 37(27), pp. 6539 –6557. doi: 10.1523/JNEUROSCI.3267 -16.2017. 528 Herz, D. M. et al. (2023) ‘Dynamic modulation of subthalamic nucleus activity facilitates adaptive 529 behavior’, PLOS Biology. Edited by A. Gail, 21(6), p. e3002140. doi: 10.1371/journal.pbio.3002140. 530 Hickok, G. (2012) ‘Computational neuroanatomy of speech production’, Nature Reviews 531 Neuroscience, 13(2), pp. 135 –145. doi: 10.1038/nrn3158. 532 Hickok, G. and Poeppel, D. (2007) ‘The cortical organization of speech processing’, Nature 533 Reviews Neuroscience, 8(5), pp. 393 –402. doi: 10.1038/nrn2113. 534 Houde, J. F. and Jordan, M. I. (1998) ‘Sensorimotor adaptation in speech production’, Science, 535 279(5354), pp. 1213 –1216. doi: 10.1126/science.279.5354.1213. 536 Indefrey, P. (2011) ‘The spatial and temporal signatures of word production components: A critical 537 update’, Frontiers in Psychology , 2(OCT), pp. 1 –16. doi: 10.3389/fpsyg.2011.00255. 538 Indefrey, P. and Levelt, W. J. M. (2004) ‘The spatial and temporal signatures of word production 539 components’, Cognition, 92(1–2), pp. 101 –144. doi: 10.1016/j.cognition.2002.06.001. 540 Jenson, D. et al. (2014) ‘Temporal dynamics of sensorimotor integration in speech perception and 541 production: Independent component analysis of EEG data’, Frontiers in Psychology, 5(JUL), pp. 1–542 17. doi: 10.3389/fpsyg.2014.00656. 543 Jorge, A. et al. (2022) ‘Hyperdirect connectivity of opercular speech network to the subthalamic 544 nucleus’, Cell Reports. The Author(s), 38(10), p. 110477. doi: 10.1016/j.celrep.2022.110477. 545 Kittilstved, T. et al. (2018) ‘The Effects of Fluency Enhancing Conditions on Sensorimotor Control 546 of Speech in Typically Fluent Speakers: An EEG Mu Rhythm Study’, Frontiers in Human 547 Neuroscience, 12(April), pp. 1 –15. doi: 10.3389/fnhum.2018.00126. 548 Knolle, F. et al. (2019) ‘Auditory Predictions and Prediction Errors in Response to Self -Initiated 549 Vowels’, Frontiers in Neuroscience , 13(October), pp. 1 –11. doi: 10.3389/fnins.2019.01146. 550 Laganaro, M., Valente, A. and Perret, C. (2012) ‘Time course of word production in fast and slow 551 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint speakers: A high density ERP topographic study’, NeuroImage . Elsevier Inc., 59(4), pp. 3881 –552 3888. doi: 10.1016/j.neuroimage.2011.10.082. 553 Loucks, T. M. J. et al. (2007) ‘Human brain activation during phonation and exhalation: Common 554 volitional control for two upper airway functions’, NeuroImage , 36(1), pp. 131 –143. doi: 555 10.1016/j.neuroimage.2007.01.049. 556 Nambu, A. (2011) ‘Somatotopic organization of the primate basal ganglia’, Frontiers in 557 Neuroanatomy , 5(APRIL), pp. 1–9. doi: 10.3389/fnana.2011.00026. 558 Nambu, A. et al. (2023) ‘Dynamic Activity Model of Movement Disorders: The Fundamental Role of 559 the Hyperdirect Pathway’, Movement Disorders, 38(12), pp. 2145–2150. doi: 10.1002/mds.29646. 560 Nambu, A., Tokuno, H. and Takada, M. (2002) ‘Functional significance of the cortico -subthalamo-561 pallidal “hyperdirect” pathway’, Neuroscience Research, 43(2), pp. 111–117. doi: 10.1016/S0168 -562 0102(02)00027 -5. 563 Neef, N. E. et al. (2015) ‘Speech dynamics are coded in the left motor cortex in fluent speakers but 564 not in adults who stutter’, Brain, 138(3), pp. 712 –725. doi: 10.1093/brain/awu390. 565 Neef, N. E. et al. (2018) ‘Structural connectivity of right frontal hyperactive areas scales with 566 stuttering severity’, Brain, 141(1), pp. 191 –204. doi: 10.1093/brain/awx316. 567 Neef, N. E., Anwander, A. and Friederici, A. D. (2015) ‘The Neurobiological Grounding of 568 Persistent Stuttering: from Structure to Function’, Current Neurology and Neuroscience Reports , 569 15(9). doi: 10.1007/s11910 -015-0579-4. 570 Niziolek, C. A., Nagarajan, S. S. and Houde, J. F. (2013) ‘What does motor efference copy 571 represent? evidence from speech production’, Journal of Neuroscience, 33(41), pp. 16110–16116. 572 doi: 10.1523/JNEUROSCI.2137 -13.2013. 573 Orpella, J. et al. (2024) ‘Reactive Inhibitory Control Precedes Overt Stuttering Events’, 574 Neurobiology of Language , 5(2), pp. 432 –453. doi: 10.1162/nol_a_00138. 575 Oswal, A. et al. (2021) ‘Neural signatures of hyperdirect pathway activity in Parkinson’s disease’, 576 Nature Communications . Springer US, 12(1), pp. 1 –14. doi: 10.1038/s41467 -021-25366-0. 577 Peelle, J. E. and Davis, M. H. (2012) ‘Neural oscillations carry speech rhythm through to 578 comprehension’, Frontiers in Psychology , 3(SEP), pp. 1–17. doi: 10.3389/fpsyg.2012.00320. 579 Pfurtscheller, G. and Lopes da Silva, F. H. (1999) ‘Event -related EEG/MEG synchronization and 580 desynchronization: basic principles’, Clinical Neurophysiology, 110(11), pp. 1842 –1857. doi: 581 10.1016/S1388 -2457(99)00141 -8. 582 Piai, V. et al. (2015) ‘Beta oscillations reflect memory and motor aspects of spoken word 583 production’, Human Brain Mapping , 36(7), pp. 2767 –2780. doi: 10.1002/hbm.22806. 584 Pion-Tonachini, L., Kreutz-Delgado, K. and Makeig, S. (2019) ‘ICLabel: An automated 585 electroencephalographic independent component classifier, dataset, and website’, NeuroImage , 586 198, pp. 181 –197. doi: 10.1016/j.neuroimage.2019.05.026. 587 Rauschecker, J. P. and Scott, S. K. (2009) ‘Maps and streams in the auditory cortex: Nonhuman 588 primates illuminate human speech processing’, Nature Neuroscience , 12(6), pp. 718 –724. doi: 589 10.1038/nn.2331. 590 Rocha, M. F., Carmona, J. and Correia, J. M. (2025) ‘EEG responses to auditory cues predict 591 fluency variability and stuttering intervention outcome’. doi: 10.1101/2025.02.21.635719. 592 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint Sammler, D. et al. (2015) ‘Report Dorsal and Ventral Pathways for Prosody’, pp. 3079 –3085. doi: 593 10.1016/j.cub.2015.10.009. 594 Schulz, G. M. et al. (2005) ‘Functional neuroanatomy of human vocalization: An H215O PET 595 study’, Cerebral Cortex , 15(12), pp. 1835 –1847. doi: 10.1093/cercor/bhi061. 596 Sengupta, R. and Nasir, S. M. (2016) ‘The predictive roles of neural oscillations in speech motor 597 adaptability’, Journal of Neurophysiology , 115(5), pp. 2519 –2528. doi: 10.1152/jn.00043.2016. 598 Silbert, L. J. et al. (2014) ‘Coupled neural systems underlie the production and comprehension of 599 naturalistic narrative speech’, Proceedings of the National Academy of Sciences of the United 600 States of America, 111(43), pp. E4687 –E4696. doi: 10.1073/pnas.1323812111. 601 Simmonds, A. J. et al. (2014) ‘Sensory -motor integration during speech production localizes to 602 both left and right plana temporale’, Journal of Neuroscience , 34(39), pp. 12963 –12972. doi: 603 10.1523/JNEUROSCI.0336-14.2014. 604 Tian, X. and Poeppel, D. (2010) ‘Mental imagery of speech and movement implicates the dynamics 605 of internal forward models’, Frontiers in Psychology , 1(OCT), pp. 1 –23. doi: 606 10.3389/fpsyg.2010.00166. 607 Tremblay, P., Deschamps, I. and Gracco, V. L. (2016) ‘Neurobiology of Speech Production’, in 608 Hickok, G. and Small, S. L. (eds) Neurobiology of Language . Academic Press, pp. 741–750. doi: 609 10.1016/B978 -0-12-407794-2.00059-6. 610 Usler, E. R. (2022) ‘Why Stuttering Occurs: The Role of Cognitive Conflict and Control’, Topics in 611 Language Disorders , 42(1), pp. 24 –40. doi: 10.1097/TLD.0000000000000275. 612 Vos, D. M. et al. (2010) ‘Removal of muscle artifacts from EEG recordings of spoken language 613 production’, Neuroinformatics , 8(2), pp. 135 –150. doi: 10.1007/s12021 -010-9071-0. 614 Weiss, S. and Mueller, H. M. (2012) ‘“Too many betas do not spoil the broth”: The role of beta 615 brain oscillations in language processing’, Frontiers in Psychology , 3(JUN), pp. 1 –15. doi: 616 10.3389/fpsyg.2012.00201. 617 Whillier, A. et al. (2018) ‘Adults who stutter lack the specialised prespeech facilitation found in non-618 stutterers’, PLoS ONE, 13(10), pp. 1 –26. doi: 10.1371/journal.pone.0202634. 619 Zhang, Y. et al. (2026) ‘More than words : word predictability , prosody , gesture and mouth 620 movements in natural language comprehension’, (January). doi: 621 10.1098/rspb.2021.0500/879281/rspb.2021.0500.pdf. 622 623 624 625 626 627 628 629 630 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted March 25, 2026. ; https://doi.org/10.64898/2026.03.23.713128doi: bioRxiv preprint