Concurrent processing of the prosodic hierarchy is supported by cortical entrainment and phase-amplitude coupling Authors: Chantal Oderbolz, Elisabeth Stark, Sebastian Sauppe, Martin Meyer Published: 2024 ABSTRACT AbstractModels of phonology posit a hierarchy of prosodic units that is relatively independent from syntactic structure, requiring its own parsing. It remains unexplored how this prosodic hierarchy is represented in the brain. We investigated this foundational question by means of an electroencephalography (EEG) study. Thirty young adults listened to German sentences containing manipulations at different levels of the prosodic hierarchy. Evaluating speech-to-brain cortical entrainment and phase-amplitude coupling revealed that prosody’s hierarchical structure is maintained at the neural level during spoken language comprehension. The faithfulness of this tracking varied as a function of the hierarchy’s degree of intactness as well as systematic interindividual differences in audio-motor synchronization abilities. The results underscore the role of complex oscillatory mechanisms in configuring the continuous and hierarchical nature of the speech signal and situate prosody as a structure indispensable from theoretical perspectives on spoken language comprehension in the brain. Introduction The speech signal contains hierarchically organized acoustic information. As a consequence of cochlear processing (Moore and Glasberg 1981; Decorsière et al. 2015), what becomes available of this signal to the auditory cortex are amplitude modulations (AMs) at different frequencies (Joris et al. 2004; Ghitza et al. 2012). Most pronounced are amplitude modulations at around 4 to 8 Hz, but energy also accumulates at the lower modulation rates of ~2 and ~1 Hz. In turn, these acoustic features correspond to syllables (Chandrasekaran et al. 2009), stress or feet (Rosen 1992), and intonation phrases (Inbar et al. 2020). Phonological theories capture these modulations, positing that the speech signal is linguistically organized into units of increasing sizes. These suprasegmental units (spanning over multiple speech sounds) form a prosodic hierarchy that ranges from syllables to intonation phrases (Beckman and Pierrehumbert 1986; Beckman 1996). Larger units such as feet are composed of and dominate smaller units such as syllables (see Fig. 1). Importantly, these models emphasize that prosody “is a complex grammatical structure that must be parsed in its own right” (Beckman 1996). Indeed, the information encoded in the prosodic hierarchy provides the listener with cues crucial for creating phonological (Cho et al. 2007), semantic (Nygaard et al. 2009), syntactic (Millotte et al. 2008), and pragmatic (Cheang and Pell 2008) representations of an utterance. Thus, by extension, comprehending speech requires that the brain efficiently and concurrently process the amplitude modulations associated with the levels of the prosodic hierarchy. Nevertheless, despite accruing evidence for prosody’s significance for language comprehension, this conjecture remains at best sparsely explored. We address this gap by means of an EEG study using a simple design involving spoken sentences with prosodic manipulations. Theoretical framework of the current study. A) The speech signal accumulates energy at certain modulation frequencies shown here in a plot of the temporal modulation spectrum (left bar: 0.6 to 1.1 Hz; middle bar: 1.8 to 2.8 Hz; right bar: 3.5 to 4.7 Hz). B) Phonological models posit that prosody, like syntax, is hierarchically structured (Liberman and Prince 1977; Beckman and Pierrehumbert 1986; Selkirk 1986). This so-called prosodic hierarchy describes prosodic units and their relationships as relatively independent from syntax and thus as a structure that must be “parsed in its own right” (Beckman 1996). The specific prosodic units that constitute the prosodic hierarchy vary slightly from model to model with consensus at the very top (i.e. concerning constituents that are defined by a coherent intonation contour) and at the very bottom (concerning constituents within a lexical word) (Shattuck-Hufnagel and Turk 1996). Here, we mainly subscribe to Beckman and colleagues’ work (Beckman and Pierrehumbert 1986; Beckman 1996) since it is acoustically motivated. Nonetheless, prosody interacts with syntax in important ways (Bennett and Elfner 2019), but this interaction is beyond the scope of this paper. C) Recent models propose a role for multifrequency (nested) oscillations in dealing with the hierarchical and continuous nature of the speech signal. Recent neurophysiological and computational models invoke oscillatory mechanisms through which hierarchical information (such as prosody) can be extracted from continuous speech. On the one hand, the brain employs multifrequency cortical oscillations that align with the quasi-periodic temporal features of the speech signal (e.g. amplitude modulations) in a process called cortical entrainment (Ghitza 2011; Giraud and Poeppel 2012; Lakatos et al. 2019). The cyclical nature of cortical oscillations creates temporal windows through which the speech signal is segmented into chunks. Temporal prediction is a natural result of these cycles wherein the oscillatory phase is shifted to remain approximately constant across time, allowing for optimal alignment with and segmentation of the speech signal. On the other hand, cross-frequency coupling (CFC) is a phenomenon ubiquitous in dynamic (neural) systems and different architectures of cross-frequency interactions between oscillations are thought to give rise to different functions (Canolty and Knight 2010; Giraud and Poeppel 2012; Hyafil et al. 2015b). One putative function linked to phase-amplitude coupling (PAC), or nesting (a form of CFC), is the temporal parsing of continuous stimuli like speech. PAC coordinates the timing and organization of neural activity between the phase of slower and the amplitude of faster oscillations, helping to identify and group relevant acoustic and linguistic units in the speech signal (Hyafil et al. 2015a). Specifically, PAC is defined as the unidirectional modulation of the power of a faster oscillation by the phase of a slower oscillation (Hyafil et al. 2015b). A growing body of evidence demonstrates that cortical entrainment and PAC are of mechanistic relevance for speech perception (Gross et al. 2013; Hyafil et al. 2015a; Keitel et al. 2018; Molinaro and Lizarazu 2018; Riecke et al. 2018; Zoefel et al. 2018; Rimmele et al. 2021). However, while there is some research investigating the neural signatures of prosodic processing that shows that the brain is sensitive to and tracks prosodic cues, studies are so far relatively scarce and have exclusively concentrated on isolated features, such as intonation (Bourguignon et al. 2013; Teoh et al. 2019; Inbar et al. 2023). It is therefore an open question as to how the hierarchical structure of prosody, with its layers of (acoustic) features, modulates neural responses to the speech signal. The current study provides a unified investigation of how the low-frequency amplitude modulations associated with the prosodic hierarchy are concurrently represented in the brain and whether this representation is underpinned by multifrequency cortical entrainment and a nested hierarchy of oscillations. Participants listened to short sentences that contained prosodic manipulations (see Materials and Methods). These manipulations were introduced to probe and amplify the brain’s ability to concurrently represent different levels of the prosodic hierarchy and how its configuration modulates neural responses. Each sentence consisted of 16 syllables. At the foot level, stress patterns adhered to either a regular (where a stressed syllable was always followed by an unstressed syllable) or an irregular (where there was no consistent stress alternation) pattern (see Table 1). In all cases, the stress sequence was based on lexical stress patterns naturally found in German. At the intonation phrase level, sentences either had an unmanipulated (i.e. naturally varying) or a manipulated (i.e. pitch-flattened) intonation contour (Meyer et al. 2004). Based on all sentences’ temporal modulation spectrum, we used pairwise phase consistency (PPC) (Vinck et al. 2010) to evaluate cortical entrainment to the syllable/theta (3.5 to 4.7 Hz), foot/high delta (1.8 to 2.8 Hz), and intonation/low delta (0.6 to 1.1 Hz) levels. Further, we quantified PAC between low (0.5 to 10 Hz) and high (1.5 to 50 Hz) frequencies as a function of prosodic manipulations using nonlinear driven autoregressive (DAR) models (Dupré la Tour et al. 2017). Additionally, we examined one more factor that characterizes differences between listeners. Assaneo et al. (2019) uncovered a bimodal distribution in the general population with listeners belonging to either a group of high or low audio-motor synchronizers. High synchronizers show increased cortical entrainment during passive listening to speech compared to low synchronizers. This is also accompanied by differences in macroscopic neuroanatomy (greater left lateralization of the arcuate fasciculus) and behavior (better word-learning performance) (Assaneo et al. 2019). In consideration of these observations, we grouped participants into high (n = 11) and low synchronizers (n = 19) post hoc, according to their performance on an established behavioral task (Lizcano-Cortés et al. 2022) (see Materials and Methods; Fig. S1). Table 1Examples of stimuli with different, grammatically permissible stress patterns (stressed syllables are printed in boldface).Regular Christa sollte damals einen runden gelben Käfer malen.“Christa should have painted a round yellow bug back then.”IrregularLeonardo soll anschliessend den berühmten Hut aufsetzen.“Leonardo should put on the famous hat afterwards.” Examples of stimuli with different, grammatically permissible stress patterns (stressed syllables are printed in boldface). We predicted the brain to phase-align to each level of the prosodic hierarchy concurrently, such that the influence of prosodic manipulations would be most apparent at the corresponding neural level (i.e. the influence of feet in the high delta and of intonation in the low delta range). However, considering the theoretical framework of the prosodic hierarchy, with larger units dominating smaller units, we likewise expected prosodic manipulations to shape cortical entrainment at lower levels. That is to say, the effect of manipulations at the level of the foot should also be found at the neural level in the theta frequency range. Specifically, we hypothesized a facilitative effect of rhythmic regularity, manifesting as increased phase alignment, compared to rhythmic irregularity. Rhythmic regularity, in both speech and auditory sequences, provides cues regarding both the meaningful grouping of units and their timing (Quené and Port 2005; Domahs et al. 2008; Roncaglia-Denissen et al. 2013). These are features that the brain can take advantage of to anticipate the occurrence of units in the incoming auditory (speech) stream, resulting in optimal alignment of neural oscillations (Giraud and Poeppel 2012). Similarly, the effect of manipulations at the level of the intonation phrase should also be found at subordinate levels. Specifically, we predicted that an unmanipulated intonation contour would coincide with increased phase alignment compared to a lack of variation in the intonation contour. Intonation conveys, to a certain degree, the syntactic structure of an utterance and thus allows words to be grouped into meaning-carrying chunks (Frazier et al. 2006). This slow prosodic rhythmicity is, in turn, tracked by the brain (Bourguignon et al. 2013). Furthermore, having been shown to aid in the parsing and grouping of continuous speech (Gross et al. 2013; Hyafil et al. 2015a), we expected the prosodic hierarchy and manipulations thereof to be neurally represented by the nesting of oscillations, such that the patterns of PAC would differ as a function of experimental conditions. These differences in PAC should be particularly evident at low frequencies, given that the prosodic hierarchy occupies the acoustic space associated with slow temporal modulations (Fig. 1A). Finally, we expected to observe differences between synchronizer groups for both cortical entrainment and PAC (Assaneo et al. 2019). However, since the neural sensitivity of high and low synchronizers to prosodic structure is so far unexplored, we had no directed hypotheses. Materials and methods Participants We recruited 33 native speakers of German (5 males). Participants were right-handed, nonmusicians, and none reported any past or current neurological or severe psychiatric disorders. Three participants were excluded from the study: We had technical issues with one, another was not a native speaker of German, and the last one was not right-handed. This left us with a sample size of 30 participants between the ages of 19 and 29 years (M = 23.2 years, SD = 3.32 years). We assessed and confirmed normal hearing ability using pure tone audiometry (MAICO ST20, MAICO Diagnostics, Berlin, Germany), defined as audiometric thresholds of 30 dB HL at octave frequencies from 0.25 to 8.0 kHz. We determined participants’ handedness using a German version of the Annett questionnaire (Annett 1970). Participants were informed in written and verbal form of the procedure and gave their written consent. Participants received either course credit or monetary compensation for their efforts. The study protocol was approved by the ethics committee of the Faculty of Arts and Social Sciences, University of Zurich (Approval number: 20.6.12), and conducted according to the principles defined in the Declaration of Helsinki. Stimuli The stimulus materials consisted of 480 unique German sentences spoken by a trained, female native speaker who was instructed to speak naturally and clearly. The recorded sentences were between 3,403 and 4,505 ms long (M = 3,954 ms, SD = 177 ms). Sentences followed a constant syntactic structure: a noun phrase (either a proper name or a noun), a verb phrase (a modal verb modified by an adverbial phrase, and a sentence-final main verb), and an object noun phrase (modified by one or two adjectives). Sentences were constructed to be meaningful but low in semantic predictability such that participants would need to attend to and actively parse each word. All words were part of the basic German vocabulary, i.e. among the 10,000 most frequent words (Scharloth et al. 2016). To probe the brain’s ability to concurrently represent different levels of the prosodic hierarchy and its configuration, any given sentence consisted of 16 syllables and contained experimental manipulations at the levels of the foot and intonation phrase. At the level of the foot, sentences either followed a regular (with predictably and consistently alternating stressed and unstressed syllables) or an irregular (i.e. unpredictable but still permissible) stress pattern (see Table 1). At the level of the intonation phrase, sentences either had an unmanipulated (i.e. naturally varying, as produced by the speaker) or a manipulated (i.e. constant) intonation contour. We used Praat (Boersma and van Heuven 2001) to extract the mean fundamental frequency (f0) value across all stimuli (196.61 Hz) and to flatten individual sentences’ fundamental frequencies to this value with the Pitch-Synchronous Overlap-Add (PSOLA) algorithm. The acoustic waveforms, their corresponding speech envelopes, spectra, and spectrograms for four example stimuli can be found in Fig. S11. As part of a larger study investigating the neural signatures of prosodic processing, the stimuli were partially embedded in speech-shaped white noise. The analyses we report here are restricted to the clean part of the sentences. Stimuli were presented at a sampling rate of 44100 Hz. Procedure From the pool of 480 stimuli, we randomly assigned 400 spoken sentences to each participant. Half of these sentences were then randomly assigned to the manipulated intonation condition and the other half to the unmanipulated intonation condition so that a given participant never heard the same stimulus in both intonation conditions. During EEG recording, participants were comfortably seated inside a booth. They listened to the binaurally presented stimuli through insert earphones (Etymotic ER-2, Etymotic Research) attached to foam ear tips (ER1-14A) at a volume of approximately 70 dB Sound Pressure Level (SPL). Stimuli were presented in a randomized order in five blocks of 80 trials, preceded by 8 practice trials. We used Presentation (Neurobehavioral Systems Inc., Berkeley, USA) to control stimulus delivery. Participants took short breaks between blocks as needed. As a behavioral task, participants were asked to type on a keyboard what had been said during noise after listening to each stimulus. There was no time limitation for participants’ responses, but they were encouraged to write their responses spontaneously even if they had not understood every single word. The inter-trial interval following this response was randomly jittered between 2,000 and 2,500 ms. The interval between the offset of speech and onset of response screen was jittered between 300 and 400 ms. The main experiment (excluding EEG application) lasted approximately 80 min. Speech preprocessing For each stimulus, we computed the broadband envelope in Python, using the toolboxes Brian2 (version 2.4.2), Brian2Hears (version 0.9.2) (Fontaine et al. 2011; Stimberg et al. 2019), and Scipy (version 1.5.3). First, to simulate the basilar membrane within the cochlea, we applied a gammatone filterbank (Decorsière et al. 2015; Biesmans et al. 2017): Acoustic waveforms were filtered into 8 subbands in the human hearing range from 20 to 20000 Hz and center frequencies equally spaced on the approximately logarithmic equivalent rectangular bandwidth (ERB) scale. Gammatone filters were implemented as cascades of four second-order infinite impulse response (IIR) filters. The approximated impulse response is based on the exact gammatone implementation as described by Slaney (1997). The parameter values for computation of the ERB bandwidth stem from Glasberg and Moore (1990). To include further nonlinearities of cochlear processing—specifically the transduction properties of inner hair cells—we then applied half-wave rectification and compression with a 6/10 power law to the output of the gammatone filterbank. We recombined the subbands as the average to obtain the broadband envelope. Finally, the signal was downsampled from 44100 to 1000 Hz to match the sampling rate of the EEG data. The continuous f0 envelope was calculated by first extracting the raw pitch values for each stimulus using Praat (Boersma and van Heuven 2001) and then interpolating these using the Wavelet prosody analyzer toolkit (version 3.2.2020) (Suni et al. 2017). The frequency bands of interest for which to compute phase synchronization were then determined based on the peaks in the temporal modulation spectra of the speech stimuli (Fig. 1A; see below) (Ding et al. 2017; Varnet et al. 2017). These acoustically defined ranges corresponded to phonologically defined ranges of the syllable, stress, and intonation phrase rates (i.e. the number of produced syllables, stressed syllables/feet, and intonation phrases divided by sentence duration). The minimum and maximum across all these values determined the lower and upper bound of the respective frequency bands: syllable/theta: 3.5 to 4.7 Hz, stress/high delta: 1.8 to 2.8 Hz, and intonation/low delta: 0.6 to 1.1 Hz. The envelopes were then filtered to these frequencies using a third-order Butterworth bandpass filter (roll-off 18 dB/octave; forward–backward digital filter using cascaded second-order sections). Temporal modulation spectrum The temporal modulation spectrum of a sound signal shows how fast its amplitude fluctuates over time and is neurophysiologically interpreted to approximately reflect the tuning of neurons to modulation frequencies (Shamma 2001; Theunissen et al. 2001; Ding et al. 2017). For its calculation, we followed the procedure described in Ding et al. (2017). First, for each speech stimulus, we extracted its amplitude envelope (as described above). Then, each envelope’s discrete Fourier transform (DFT) is computed by means of the fast Fourier transform. The modulation spectrum is equal to the root mean square of the DFT of all envelopes. We ran these calculations for speech stimuli with an unmanipulated vs. manipulated intonation contour separately (see Fig. 1A). E‌EG EEG was recorded using BrainVision Recorder (version 1.23.0003, BrainProducts GmbH, Gilching, Germany) and a 64-channel actiCHamp Plus system (Brain Products GmbH, Gilching, Germany) arranged according to an extended 10/20 system with the following Ag/AgCl active electrodes: Fp1, Fp2, AF7, AF3, AFz, AF4, AF8, F7, F5, F3, F1, Fz, F2, F4, F6, F8, FT9, FT7, FC5, FC3, FC1, FC2, FC4, FC6, FT8, FT10, T7, C5, C3, C1, Cz, C2, C4, C6, T8, TP9, TP7, CP5, CP3, CP1, CPz, CP2, CP4, CP6, TP8, TP10, P7, P5, P3, P1, Pz, P2, P4, P6, P8, PO7, PO3, POz, PO4, PO8, O1, Oz, and O2. The signal was online referenced to FCz, and we used FPz as a ground electrode. Throughout the recording, electrode impedance was kept below 30 kΩ using electroconductive gel. The signal was recorded continuously with a bandwidth from direct current (DC) to 280 Hz and digitized at a sampling rate of 1000 Hz. Preprocessing of the data was performed in Python (version 3.8.6) using MNE Python (version 1.0.0) (Gramfort et al. 2013; Gramfort et al. 2014). We first filtered the data from 1 to 30 Hz (for phase synchronization analyses) and from 1 to 50 Hz (for phase-amplitude coupling analyses) using a one-pass, zero-phase, noncausal bandpass filter (roll-off 6 dB/octave) and applied a one-pass, zero-phase noncausal bandstop filter at 50 Hz and harmonics (roll-off 6 dB/octave). Ocular artifacts such as blinks and saccades and heartbeat artifacts were corrected using independent component analysis (ICA) (Jung et al. 2000). Components were manually selected for removal by visually inspecting their topographies and time courses. Artifactual channels were identified visually and excluded from the ICA decomposition. The ICA solution was then applied to data filtered between 0.3 to 30 Hz and 0.3 to 50 Hz, respectively (one-pass, zero-phase, noncausal bandpass filter; roll-off 6 dB/octave), and previously excluded channels were interpolated using spherical spline interpolation (Perrin et al. 1989). We re-referenced the cleaned data to the average across all channels and segmented the continuous data into epochs from −2,000 ms prior to speech onset to 300 ms postspeech offset. Finally, epochs with a maximum and/or minimum peak-to-peak signal amplitude (PTP) of 150 and 1 μV, respectively, were dropped. PTP is defined as the absolute difference between the lowest and highest signal value and indicates the peakedness/flatness of the signal. Quantifying cortical entrainment: phase synchronization EEG signals were first bandpass-filtered to the same frequencies as the speech signal (3.5 to 4.7, 1.8 to 2.8, and 0.6 to 1.1 Hz, respectively) using a third-order Butterworth bandpass filter (roll-off 18 dB/octave; forward–backward digital filter using cascaded second-order sections). For each signal, we computed the Hilbert transform and extracted the instantaneous phase as the phase angle of the resulting analytic signal. To quantify the extent of phase synchronization, we calculated single-trial PPC (Vinck et al. 2010) between each speech stimulus and each EEG channel. An advantage of PPC is that it is not biased by sample size. We omitted the first 500 ms of the sentences from this calculation to avoid spurious phase-locking induced by evoked potentials in response to speech onset. PPC is calculated as follows: First, the relative phase between the EEG and speech signal at each timepoint within a trial is calculated. Then, the angular distance between each observation of the relative phase is calculated pairwise. If the phase relationship between the two signals is synchronous across time, the angular distances between pairs of relative phases will be small. Finally, averaging the dot product across all pairs of these observations results in the PPC. PPC values of 0 indicate nonexistent phase synchronization while values of 1 indicate perfect phase synchronization. Because this measure compares all observations of relative phase in a pairwise manner, it is computationally less efficient than other measures of phase synchrony (Lachaux et al. 1999). To counter this, we further downsampled all involved signals to 100 Hz after extracting the instantaneous phase. Quantifying phase-amplitude coupling: nonlinear driven autoregressive models To test for PAC and whether it varies as a function of experimental manipulations, we computed comodulograms on the basis of nonlinear DAR models using the Python toolbox pactools (version: 0.3.1) (Dupré la Tour et al. 2017). This is a statistical signal modeling approach (as opposed to a biophysical one), and it avoids some of the methodological pitfalls associated with other PAC measures such as applying the Hilbert transform to broadband signals or incorrectly choosing the bandwidth of bandpass filters. DAR models can furthermore handle the nonlinearity and nonstationarity characteristic of neural signals and do not expect these to be strictly sinusoidal. Specifically, the (low) driver frequency is first extracted using a bandpass filter. This component represents the phase-modulating frequency. To identify the high-frequency component, the driver frequency is subtracted from the original raw signal. This isolates the part of the signal that is expected to be amplitude-modulated by the driver frequency. To maintain the continuity of the signal, the resulting gap in the signal is filled with Gaussian white noise. By repeating this procedure for a range of driver frequencies, we thus obtain a comodulogram that represents PAC between pairs of frequencies, showing how the phase of low-frequency drivers modulates the amplitude of high-frequency components. Applied to our data, for each participant, electrode, and condition, a comodulogram was calculated across trials for low frequencies between 0.5 and 10 Hz and high frequencies between 1.5 and 50 Hz using an autoregressive model order of 20. In the absence of actual PAC, a measure of PAC can yield values close to but not exactly zero. Thus, to distinguish spurious from actual PAC, we additionally computed surrogate comodulograms by adding random time shifts, thereby removing any possible coupling (Dupré la Tour et al. 2017). Because DARs are computationally expensive, we downsampled the data to 250 Hz. S‌SS test To allow a classification into high and low synchronizers, prior to the experiment participants completed the spontaneous speech synchronization (SSS) task (Assaneo et al. 2019) where they hear a stream of syllables and are asked to whisper the syllable “ta” simultaneously and continuously. Analysis of the degree of synchronization between the produced and perceived speech stream and the categorization into high and low synchronization was done using the openly available code (Lizcano-Cortés et al. 2022). This resulted in 11 high and 19 low synchronizers (high synchronizers: MPLV = 0.59, SDPLV = 0.05; low synchronizers: MPLV = 0.28, SDPLV = 0.06; see Supplementary Fig. S1). Statistical analyses Regions of interest For statistical analysis of phase synchronization, channels were grouped into nine regions of interest (ROIs) (see Supplementary Fig. S2): left frontal (Fp1, AF7, AF3, F7, F5, F3), middle frontal (AFz, F1, Fz, F2), right frontal (Fp2, AF4, AF8, F4, F6, F8), left central (FT9, FT7, FC5, FC3, T7, C5, C3, TP9, TP7, CP5, CP3), middle central (FC1, FCz, FC2, C1, Cz, CP1, CPz, CP2), right central (FC4, FC6, FT8, FT10, C4, C6, T8, CP4, CP6, TP8, TP10), left posterior (P7, P5, P3, PO7, PO3, O1), middle posterior (P1, Pz, P2, POz, Oz), and right posterior (P4, P6, P8, PO4, PO8, O2). Mixed models: glmertrees For the statistical analysis of phase synchronization, we used the R package glmertree (Fokkema et al. 2018), which implements a decision tree–based algorithm for fitting mixed effects models. Predictors (fixed and random effects), their higher-order effects, and interactions are used to split the data into subgroups based on the similarity of response patterns. The model is fit to each subgroup such that different effects can be estimated for different subgroups. The relationship between dependent and independent variables is described by a binary decision tree, which has the advantage of increased interpretability (Fokkema et al. 2021). The splitting variable at each node is selected based on the Bonferroni-corrected P-value of an association test. All data splits are ordered by the amount of variance explained in the data where the root node represents the independent variable that explains the most variance, i.e. it is the strongest predictor of the dependent variable. We fit a beta regression (logit link function) for each of the three frequency bands separately with regularity (regular/irregular), intonation (unmanipulated/manipulated), SSS (high/low), and ROI (9 ROIs) as partitioning variables. The maximal random effects structure (Barr et al. 2013; Singmann and Kellen 2019) consisted of by-participant and by-item random intercepts and slopes and was simplified if there were convergence or singular fit issues (Singmann and Kellen 2019). For the sake of interpretability, the maximal number of splits in each model was restricted to five. Decoding of PAC comodulograms With the goal of investigating the extent to which information about the speech signal is represented in PAC, we performed group-level decoding of PAC comodulograms across the scalp. We trained a machine learning classifier to distinguish between experimental conditions and synchronizer groups and their associated PAC comodulograms at each electrode. The resulting decoding accuracy is indicative of the information available in PAC about the discriminability of experimental conditions and their respective neural representations (Grootswagers et al. 2017) at the sensor level. The data consisted of 2D frequency–frequency matrices per electrode, per condition and per participant. These were unraveled to vectors and associated with their respective target label (regularity: regular, irregular; intonation: unmanipulated, manipulated; regularity × intonation: regular/unmanipulated, regular/manipulated, irregular/unmanipulated, irregular/manipulated, regularity × intonation × SSS: regular/unmanipulated/high, regular/unmanipulated/low, regular/manipulated/high, regular/manipulated/low, irregular/unmanipulated/high, irregular/unmanipulated/low, irregular/manipulated/high, irregular/manipulated/low). We performed decoding of comodulograms using scikit-learn (version 1.2.0) (Pedregosa et al. 2011) in Python (version 3.10.8). After z-transforming the data, we classified PAC comodulograms using a linear Support Vector Machine (SVM; regularization parameter C = 1). We assessed decoding accuracy with a repeated stratified 5-fold cross-validation scheme (100 iterations). This means that, for every iteration, the data was split into five roughly equal-sized folds while ensuring that each fold contained a proportion of target labels representative of the whole dataset. Classifier training was done on four folds and testing on the remaining fold. This was repeated until each fold had been used for testing once. Decoding accuracy corresponded to the mean correct classification across all folds and iterations. Although decoding accuracy above chance level indicates that classification is not purely random, these levels may have limited informative value with datasets of small sample size (Combrisson and Jerbi 2015). We thus provide statistical significance levels for decoding accuracy by relying on an empirical solution, i.e. permutation testing. Specifically, we tested whether decoding accuracies of actual PAC comodulograms are significantly higher than decoding accuracies obtained from classifying surrogate PAC comodulograms (10,000 permutations). Figure 6 shows electrodes that had above-chance-level decoding accuracy and were additionally significantly different from surrogate decoding accuracies (P < 0.05, false discovery rate–corrected [Benjamini and Hochberg 1995] for 64 tests). A visualization of electrode-wise decoding accuracies across iterations (observed vs. surrogate PAC comodulograms) can be found in Supplementary Figs. S3–S6. Furthermore, a visualization of the null distributions for each electrode used to determine significance can be found in Supplementary Figs. S7–S10. Next, to determine which features (i.e. PAC frequency pairs) contributed to the decoding accuracy, we extracted the support vector weights associated with each feature. These weights indicate the contribution of each feature to the classification decision, revealing which features provided the most discriminative information for the classifier regarding the experimental conditions and groups. To facilitate a neurophysiological interpretation, these weights were transformed into activation patterns, in our case PAC strength (Haufe et al. 2014). They were inverse-transformed according to the transformer steps described above. Results Cortical entrainment to the prosodic hierarchy Theta/syllable Modeling neural synchronization in the theta range (3.5 to 4.7 Hz), i.e. to amplitude modulations associated with syllables, revealed as a first split a facilitative effect of rhythmic regularity (Fig. 2). Listening to sentences with a regular stress pattern yielded higher PPC values than listening to irregularly stressed sentences. Within the regular and irregular sentences, respectively, an unmanipulated intonation contour resulted in higher neural synchronization than a manipulated intonation contour. Additionally, we found (albeit somewhat varied) differences between synchronizer groups: Whereas high synchronizers exhibited stronger neural synchronization for sentences that are irregular with a manipulated intonation contour and sentences that are regular with an unmanipulated intonation contour, low synchronizers showed higher PPC for sentences that are irregular with an unmanipulated intonation contour and sentences that are regular with a manipulated intonation contour. In other words, high synchronizers showed increased neural synchronization compared to low synchronizers in response to sentences that were either highly predictable (i.e. with a regular stress pattern and unmanipulated intonation contour) or hard to predict (i.e. with an irregular stress pattern and manipulated intonation contour). Finally, although prosodic manipulations affected neural synchronization across the whole scalp, topographical differences emerged between synchronizer groups. High synchronizers showed differences between centro-frontal and lateral-posterior electrode sites, while low synchronizers showed effects spanning the whole scalp. Regression tree for pairwise phase consistency in the theta frequency range (3.5 to 4.7 Hz), corresponding to the syllable level in the prosodic hierarchy. Topographic maps show the electrode positions of ROIs included in the respective grouping. All splits are statistically significant at P < 0.05 (Bonferroni-corrected). Terminal nodes show box plots of PPC values with the distribution (kernel density estimate) in gray. SSS = spontaneous speech synchronization. High delta/foot Modeling neural synchronization in the high delta range (1.8 to 2.8 Hz), i.e. to amplitude modulations associated with prosodic feet, revealed as a first split a difference between synchronizer groups (Fig. 3). Contrary to our expectations, regardless of the particular sentence type they were listening to, low synchronizers showed increased PPC compared to high synchronizers. Nonetheless, the next and most prominent prosodic split revealed that sentences with a manipulated intonation contour elicited decreased neural synchronization in both groups. While rhythmic regularity was associated with increased PPC in the theta frequency range, the opposite was the case for high delta, where rhythmic irregularity in both synchronizer groups and intonation conditions led to higher PPC values compared to rhythmic regularity. Finally, topographical differences emerged between rhythm conditions but especially among low synchronizers who mainly showed variation between electrode sites over the left and right hemispheres. Regression tree for pairwise phase consistency in the high delta frequency range (1.8 to 2.8 Hz), corresponding to the foot level in the prosodic hierarchy. Topographic maps show the electrode positions of ROIs included in the respective grouping. All splits are statistically significant at P < 0.05 (Bonferroni-corrected). Terminal nodes show box plots of PPC values with the distribution (kernel density estimate) in gray. SSS = spontaneous speech synchronization. Low delta/intonation phrase Modeling neural synchronization in the low delta range (0.6 to 1.1 Hz), i.e. to amplitude modulations associated with intonation, revealed as a first split a facilitative effect of rhythmic regularity (Fig. 4). Listening to sentences with a regular stress pattern resulted in stronger neural synchronization than listening to sentences with an irregular stress pattern. The next split revealed that an unmanipulated intonation contour was associated with higher PPC values in both regular and irregular sentences. Within sentences with a regular stress pattern and an unmanipulated intonation contour, high synchronizers exhibited effects across the whole scalp, while low synchronizers showed differences between left frontal electrode sites and the rest of the scalp. Low synchronizers also showed PPC differences between centro-parietal electrode sites and the rest of the scalp for sentences with an irregular stress pattern and a manipulated intonation contour. In this condition with unpredictable stress and intonation patterns, high synchronizers showed higher neural synchronization compared to low synchronizers whereas the opposite was true for the fully predictable condition (i.e. a regular stress pattern and unmanipulated intonation contour). Finally, listening to irregular sentences with an unmanipulated intonation contour was associated with topographical differences mainly between left frontal, temporal, and posterior electrode sites and right frontal, central, and posterior electrode sites (across both synchronizer groups). Regression tree for pairwise phase consistency in the low delta frequency range (0.6 to 1.1 Hz), corresponding to the intonation phrase level in the prosodic hierarchy. Topographic maps show the electrode positions of ROIs included in the respective grouping. All splits are statistically significant at P < 0.05 (Bonferroni-corrected). Terminal nodes show box plots of PPC values with the distribution (kernel density estimate) in gray. SSS = spontaneous speech synchronization. Overall, the results from the three models imply that the amplitude modulations associated with the levels of the prosodic hierarchy modulate the brain’s response, i.e. cortical entrainment, in accordance with prosody’s hierarchical structure (Fig. 5). Acoustic manipulations at higher levels in the prosodic hierarchy mainly influenced cortical entrainment at lower levels: This is evident in the fact that entrainment in the theta band was primarily influenced by manipulations at the foot level, while entrainment in the high delta band was primarily modulated by manipulations at the intonation phrase level. This implies a hierarchy not only in the acoustic and phonological space but also a hierarchy embodied in neural activity. Nevertheless, effects also manifested in the other direction such that manipulations at the foot level (high delta) impacted entrainment in the low delta band. Moreover, the results indicate that listeners are differentially sensitive to the speech signal’s prosodic properties and accordingly vary in their neural responses to it. Overview of entrainment findings. Taking into account the results from all three regression models reveals that the acoustic information associated with the prosodic hierarchy top–down modulates cortical entrainment. As we illustratively show in this overview figure, the main influence on a given level originates from prosodic manipulations at the next higher, dominating level (bold lines). Influence from manipulations at other levels are also present (dashed lines). Decoding of PAC comodulograms To further explore this pattern of hierarchical relationships, we tested the extent to which information about the speech signal is represented neurally by the nesting of oscillations (see Materials and Methods). We reasoned that if the speech signal has distinct neural representations between conditions in the form of PAC, the corresponding comodulograms should be reliably discriminable by a classifier (Grootswagers et al. 2017). We therefore trained a linear classifier to discriminate between experimental conditions and their associated PAC comodulograms. Figure 6A–C shows that, across the scalp, there are electrodes that reflect neurally distinct PAC between conditions and synchronizer groups. The SVM classifier was able to distinguish whether PAC resulted from trials with a regular vs. irregular rhythm, an unmanipulated vs. manipulated intonation contour, or whether comodulograms stemmed from a high or a low synchronizer. Significant electrodes for regularity (Fig. 6A) and synchronizer group (Fig. 6C) were more numerous over the left hemisphere than over the right hemisphere, whereas they were more distributed across the whole scalp for intonation (Fig. 6B). The classifier was also able to distinguish the combination of prosodic manipulations and the synchronizer group in a more fine-grained manner (Fig. 6D). Electrodes that showed significant above-chance-level decoding accuracies were distributed across the whole scalp. These results suggest that how oscillations are nested depends on the prosodic hierarchy and manipulations thereof, as well as on the synchronizer group. Topographical patterns of decoding accuracies. Dots denote electrodes identified by permutation testing as showing above-chance-level decoding accuracies significantly different from decoding accuracies stemming from classifying surrogate PAC comodulograms. A) Decoding accuracies achieved by the classifier when distinguishing comodulograms stemming from regular vs. irregular sentences (chance-level = 50%). B) Decoding accuracies achieved by the classifier when distinguishing comodulograms stemming from sentences with an unmanipulated vs. manipulated intonation contour (chance level = 50%) C) Decoding accuracies achieved by the classifier when distinguishing PAC comodulograms stemming from high vs. low synchronizers (chance level = 50%). D) Decoding accuracies achieved by the classifier when distinguishing fully crossed PAC comodulograms (chance level = 12.5%). SSS = spontaneous speech synchronization. To identify which PAC frequency relations were particularly informative for the classifier’s ability to distinguish between experimental conditions and groups, we next extracted inverse-transformed support vector weights from the fitted classifiers. With their magnitude and direction, they indicate the importance and contribution of each feature (i.e. PAC frequency pair) to the classification performance. To enable a neurophysiological interpretation, these weights were extracted as activation patterns (Haufe et al. 2014). Plotting these patterns (averaged across electrodes showing significant above-chance-level decoding accuracy) shows that nesting between the phase of oscillations between ~0.5 and 4 Hz and the amplitude oscillations between ~1.5 and 8 Hz mainly acted as drivers of classification of prosodic condition as well as the synchronizer group (Fig. 7). In other words, coupling between oscillations of ~0.5 to 4 Hz and oscillations of ~1.5 to 8 Hz mainly served to distinguish between experimental conditions and groups. Also important for classification, albeit to a lesser extent, was PAC between the phase of oscillations between ~0.5 and 10 Hz and the amplitude of higher-frequency oscillations around 30 and 45 Hz. On the one hand, this is evidence that the speech signal’s hierarchical information occurring at low modulation frequencies is dealt with by oscillatory cross-frequency interactions. On the other hand, this also shows that these cross-frequency interactions manifest differently depending on whether a person is a high or a low synchronizer. Matrices of inverse-transformed classifier weights extracted as activation patterns. The weights indicate how much a particular feature contributed towards the classifier’s decision and thus which features distinguish between experimental conditions and groups. The main drivers of classification were PAC between the phase of oscillations between ~0.5 and 4 Hz and the amplitude of oscillations between ~1.5 and 8 Hz. Additionally, PAC between the phase of oscillations between ~0.5 and 10 Hz and the amplitude of higher-frequency oscillations ~30 and 45 Hz also contributed to the classification decision. A) Weights indicating which features were driving the classification of PAC comodulograms stemming from trials with a regular vs. irregular rhythm. B) Weights indicating which features were driving the classification of PAC comodulograms stemming from trials with an unmanipulated vs. a manipulated intonation contour. C) Weights indicating which features were driving the classification of PAC comodulograms stemming from high vs. low synchronizers. D) Weights indicating which features were driving classification of fully crossed PAC comodulograms. Discussion Cortical entrainment is modulated by the prosodic hierarchy This study brings into alignment acoustics, phonology, and neurophysiology to probe the processing of spoken language in the human brain. The combined cortical entrainment and PAC responses to the syllable, foot, and intonation phrase levels invite the conclusion that the prosodic hierarchy, as motivated by linguistic theory, has a corresponding neural implementation. Evaluating the extent to which the brain synchronizes to the speech signal showed that listeners concurrently process the amplitude modulations associated with the prosodic hierarchy. We found that this neural response is structured in such a way that reflects this hierarchy: Amplitude modulations associated with a given level of the prosodic hierarchy primarily influenced neural synchronization of subordinate levels. Regular and irregular alternations of stress at the foot level had the greatest impact on cortical entrainment at the syllable level (i.e. in the theta frequency range). Similarly, manipulating the intonation contour had the greatest influence on cortical entrainment at the foot level (i.e. the high delta frequency range). In both instances, we observed a facilitative effect when the speech signal contained predictable prosodic patterns, i.e. a regular (vs. an irregular) stress pattern or an unmanipulated vs. manipulated intonation contour. We suggest that underlying these effects are the predictive mechanisms inherent to entrainment. In synchronizing to the regularities of the speech signal, oscillatory cycles act as temporal windows to chunk discrete linguistic units (Ghitza 2011; Giraud and Poeppel 2012), in the current case syllables, feet, and intonation phrases. It is assumed that the phases of these cycles can be used by bilateral brain areas further along in the processing stream (e.g. the motor system or peri-sylvian areas engaged in higher-level linguistic processing) to predict the occurrence of incoming auditory events (Arnal and Giraud 2012; Rimmele et al. 2018). Top–down predictions stemming from these areas phase-reset cortical oscillations (Rimmele et al. 2018) such that information coinciding with phases of high neuronal excitability are, as a consequence, amplified and easier to perceive (Henry and Obleser 2012), ultimately giving rise to a facilitative effect. Nevertheless, we also observed unexpected patterns of effects such as the influence of rhythmic regularity on neural synchronization in the low delta frequency range. In this regard, there are several possible explanations for why intonation was not the variable that explained the most variance. First, pitch is just one of several cues (along with, for example, pauses, final syllable lengthening, and changes in amplitude) that together contribute to signaling intonation phrases and their boundaries (Wightman et al. 1992; de Pijper and Sanderman 1994). Thus, pitch-flattening alone may not have fully eliminated listeners’ ability to predict and synchronize to the intonation phrase level. Second, it has been shown that listeners project a prosodic contour onto speech that is physically lacking one (Fodor 1998; Fodor 2002)—a phenomenon called implicit prosody—which modulates behavioral and neural responses (Hwang and Steinhauer 2011; Glushko et al. 2022). Third, recent research has shown that rising (as compared to falling) boundary contours attract attentional resources to the whole domain (i.e. prosodic phrases) they delimit (Grice et al. 2024). This effect could modulate oscillatory responses in complex ways that remain to be determined. Nevertheless, the results imply that prosody is a structure that significantly modulates the configuration of neural responses in the form of oscillatory activity, underscoring the need to account for prosody in theoretical perspectives (Giraud and Poeppel 2012). PAC reflects manipulations of the prosodic hierarchy Furthermore, delta–theta PAC varied as a function of the prosodic hierarchy and the manipulation thereof. Hierarchically coupled low-frequency oscillations (<8 Hz) were the main features that differentiated PAC comodulograms. This suggests that the brain employs nested oscillations to concurrently process and represent the acoustic information associated with the prosodic hierarchy. Doing so presumably allows the brain to maintain this hierarchy and the codependent nature of the levels (Lakatos et al. 2005; Giraud and Poeppel 2012; Gross et al. 2013; Hyafil et al. 2015b). Consequently, emerging from this top–down coordination of neural activity are discrete units where syllables are bound inside feet, and feet are bound inside intonation phrases. The finding that PAC comodulograms were distinguishable as a function of experimental condition furthermore indicates that introducing prosodic manipulations to the speech signal led to altered cross-frequency coupling, demonstrating the sensitivity of oscillations to the structure of the prosodic hierarchy. Depending on the intactness of the prosodic hierarchy, this would lead to (im)precisely coupled oscillations and thus (in)accurately chunked units. Although this has otherwise not been tested in the context of prosody, previous research has shown that neural activity in the form of PAC is less coordinated for nonspeech (Mai et al. 2016), unintelligible speech (Gross et al. 2013; Lizarazu et al. 2023), and non-native speech (Lizarazu et al. 2023). These findings show that the robustness of PAC coordination is related to the efficiency with which the brain can segment the speech signal. However, it remains to be seen whether PAC as a function of the prosodic hierarchy is also related to behavior or perception and thus is truly conducive to efficient parsing. We additionally found that theta-gamma coupling likewise (although to a lesser extent) contributed to PAC comodulogram classification. Since we did not systematically vary acoustic information in the gamma frequency range (~30 to 50 Hz), this activity could be indicative of “the operating rhythm of the sensory cortex” (Hyafil et al. 2015b) related to the (de-)activation of neuronal networks (Canolty and Knight 2010) that varied between experimental conditions. In instances where a stimulus contains predictable rhythmic patterns, the brain operates in a rhythmic mode where the lower frequency phase and higher frequency (gamma) amplitude are coupled. Gamma oscillations become more synchronous whenever a task-relevant event is expected (Schroeder and Lakatos 2009), enhancing neural representations. When rhythmic patterns are absent or unpredictable, the operating mode changes to one of continuity, and oscillations become decoupled (Schroeder and Lakatos 2009). Although these conjectures in the context of our study are speculative for the time being, these different operating modes may be what partially drove classification. An oscillatory vs. evoked account of neural responses The burgeoning of research on and theorizing about neural oscillations notwithstanding, there is an ongoing debate about the nature of entrainment and whether it is to be considered epiphenomenal, a methodological artifact, or rather fundamentally causal in the context of language and cognition (Obleser and Kayser 2019). Contributing to this discussion, the top–down hierarchical structure of low-frequency neural responses to the prosodic hierarchy we find here speaks to the third interpretation. Low-frequency (delta-band) oscillations, often found in associative (Donhauser and Baillet 2020) and frontal areas (Fontolan et al. 2014; Park et al. 2015), have been shown to top–down modulate both gamma power (Fontolan et al. 2014) as well as entrainment (Park et al. 2015) in the (left) auditory cortex, presumably harnessing contextual information such as prosody (Ghitza 2011; Ghitza et al. 2012) and conveying predictions from higher to lower levels (Schröger et al. 2015). The observed role of delta-band oscillations in serving linguistic and speech-specific processing (Kayser et al. 2015; Molinaro and Lizarazu 2018; Rimmele et al. 2021) therefore comes as no surprise. Additionally, neural phase synchronization in the theta, high, and low delta frequency bands showed nonredundant effects. If the results were merely reflective of passive bottom–up tracking of acoustic transients, we would have expected to observe identical effects across all three frequency bands and no top–down modulation (Cogan and Poeppel 2011). Finally, it has been argued that it is difficult to distinguish the spectro-temporal patterns of evoked responses from those of genuine oscillations (Gourévitch et al. 2020). However, evoked responses like the N100 are not expected to make any measurable contribution to low-frequency entrainment on account of their short latencies. Long-latency components like the N400 or P600, although plausibly involved, are unlikely to have differed systematically between our conditions since the semantic content was comparable across all stimuli and the syntactic structure was held constant and highly predictable (Zhou et al. 2016). However, it is known that variations in pitch or syllable structure can modulate event-related potential (ERP) components like the mismatch negativity (Tsang et al. 2011; Li and Chen 2018; Emmendorfer et al. 2020; Lialiou et al. 2024) or the P3 (Tsang et al. 2011; Hsu et al. 2015). However, this mainly holds for passive listening paradigms whereas we required participants to engage with the stimuli. Nevertheless, the neural dynamics supporting speech perception are most likely more complex than the juxtaposition of “oscillation or not” will allow (Doelling and Assaneo 2021). Interindividual differences: the potential role of the motor system Language and speech are such integral parts of the human condition that—barring pathology—most humans can be considered experts of their first language. At the same time, variation is a natural consequence of the dynamic and complex processes of life that are expressed in our genetic makeup and behaviors, including language and speech. This observation does not just pertain to variation as found in disorders like developmental language disorders, aphasia, or autism but, as it turns out, also to more subtle differences in the neurotypically developed population as recently discovered by Assaneo et al. (2019). The behavioral audio-motor synchronization task they developed to identify these groups requires listeners to synchronize their own speech output with a syllable stream. This task mainly targets the rhythmic amplitude modulations associated with syllabic information and captures the extent to which individuals can capitalize on efferent motor signals to predict auditory input. However, as shown in Fig. 1A, the speech signal also contains other prominent prosodic patterns. So far, it remains unexplored whether auditory-motor coupling as measured by this task also relates to and is beneficial for the perception of these slower modulations as a function of group membership. While we did find some group differences regarding cortical entrainment, these differences varied across frequency ranges and prosodic levels and showed that high synchronizers are not pervasively more in tune with the rhythmicity of the speech signal (although they tended to show stronger synchronization to conditions with unpredictable rhythms). These varying patterns of cortical entrainment might come about as a consequence of compensatory mechanisms, nonlinear relationships, or differential allocation of resources (Kern et al. 2021). Despite these discrepancies, decoding of PAC patterns revealed more systematic differences in the neural representations of the prosodic hierarchy between the two groups especially in cross-frequency coupling between low-frequency oscillations. As one potential explanation, we propose that at the heart of these findings may lie qualitative differences in how the motor system is engaged during perception such that the top–down phase reset of oscillations elicited by the motor system results in more or less precisely aligned high excitability phases with the incoming sensory input (Morillon et al. 2015; Rimmele et al. 2018; Morillon et al. 2019). What differentiates high from low synchronizers is the extent of entrainment in frontal areas involved in speech production as well as structural differences in the white matter fibers connecting frontal and temporal regions (Assaneo et al. 2019). Although the results differentiating synchronizer groups in our study vary across neural measures and experimental conditions, they imply that the differential engagement of the motor system influences the representation of prosodic information. Indeed, there is research to show that (pre-)motor regions (minimally) support not only the perception of speech more generally (Wilson et al. 2004; Meister et al. 2007; Callan et al. 2010) but prosody in particular (Aziz-Zadeh et al. 2010; Sammler et al. 2015). Thus, it is tenable to extend the account of interindividual differences related to audio-motor synchronization (Assaneo et al. 2019) to include prosodic cues as well. We speculate that prosody’s capacity as an integral structuring device, as the current findings suggest, may have evolved from the close functional connection between the motor and auditory cortex (Poeppel and Assaneo 2020). Once established as part of linguistic structure, prosody would have been subject to the same diversification processes as all other subsystems (Bickel et al. 2024). This would explain why languages show such a wide range of linguistic diversity including in the domain of prosody and intonation (Jun 2005). Further research will elucidate the group differences between high and low synchronizers found in the current study as well as the role of the motor system in the perception of prosody. Limitations and future directions Contextualizing the results and our interpretations thereof, we should also take certain limitations into consideration. First, although not the primary focus of this study, the interpretation that high and low synchronizers differ in their neural representations of the prosodic hierarchy as a function of audio-motor synchronization will need to be corroborated in future research. Even though we used a well-established behavioral proxy, a direct measure of the interaction between auditory and motor regions would be valuable to support this claim. Similarly, while the brain apparently seems to care about the prosodic hierarchy’s intactness, it is as of yet still unsettled whether this is of significance for perception and behavior—a conjecture that could be assessed by probing prosodic chunking (Rimmele et al. 2021). Naturally, these considerations also extend to investigating languages other than German with different prosodic characteristics (cf. Schiering et al. 2010). While hierarchically organized prosodic units are assumed to be a universal trait, languages may nevertheless differ in the physical realization and/or presence of certain levels within the hierarchy (Shattuck-Hufnagel and Turk 1996). Furthermore, while topographical differences are indicative of distinct underlying brain processes, because of the nature of scalp-level EEG, future research will need to flesh out what the underlying neural sources of the current entrainment and decoding results are and how they contribute to the topographical patterns observed at the scalp level. Finally, the study and results are a case in point for attempting to verify the psychological and neural plausibility of theoretical linguistic constructs (cf., e.g. Sauppe 2017). In this sense, we regard establishing the commensurability of the units of linguistic and the units of neurophysiological computation as a fundamental responsibility of interdisciplinary research (Poeppel 2012; Embick and Poeppel 2015). Conclusion We provide evidence that prosody and its hierarchical structure are cues that decisively shape the neural representation of the speech signal and the efficiency with which this representation is formed. By interfacing acoustics, phonology, and neurophysiology, we characterize one of the complex ways in which the rich information in the speech signal is elaborated by the brain. The experiment underscores the involvement of predictive oscillatory mechanisms in this process, and it extends the audio-motor synchronization framework to include prosody. Finally, the results are initial evidence that the assumptions brought to bear in theories about prosodic structure (Beckman and Pierrehumbert 1986; Beckman 1996) are a neural reality: Maintaining the prosodic hierarchy neurally—through a combination of cortical entrainment and PAC—ensures the accurate segmentation and binding of the speech signal into coherent prosodic units. To what extent this holds universally across languages will need to be explored in future research. Overall, this set of findings informs and broadens current theories on spoken language comprehension in the brain as well as phonological research. Supplementary Material