Listen to Yourself: The Medial Prefrontal Cortex Modulates Auditory Alpha Power DuringSpeech Preparation

Nadia Müller1,2, Sabine Leske3, Thomas Hartmann1, Szabolcs Szebényi3,4 and Nathan Weisz1

1Center for Mind/Brain Sciences, Università degli Studi di Trento, 38123 Mattarello (TN), Italy, 2Department of Neurology,Epilepsy Center, University Hospital Erlangen, 91054 Erlangen, Germany, 3Department for Neuropsychology, University ofKonstanz, 78457 Konstanz, Germany and 4Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen,6500 Nijmegen, The Netherlands

Address correspondence to Dr Nadia Müller, Università degli Studi di Trento, Center for Mind/Brain Sciences, Via delle Regole 101,38123 Mattarello (TN), Italy. Email: nadia.mueller@gmail.com

How do we process stimuli that stem from the external world andstimuli that are self-generated? In the case of voice perception it hasbeen shown that evoked activity elicited by self-generated sounds issuppressed compared with the same sounds played-back externally.We here wanted to reveal whether neural excitability of the auditorycortex—putatively reflected in local alpha band power—is modu-lated already prior to speech onset, and which brain regions maymediate such a top-down preparatory response. In the left auditorycortex we show that the typical alpha suppression found when parti-cipants prepare to listen disappears when participants expect a self-spoken sound. This suggests an inhibitory adjustment of auditory cor-tical activity already before sound onset. As a second main finding wedemonstrate that the medial prefrontal cortex, a region known for self-referential processes, mediates these condition-specific alpha powermodulations. This provides crucial insights into how higher-orderregions prepare the auditory cortex for the processing of self-generatedsounds. Furthermore, the mechanism outlined could provide further ex-planations to self-referential phenomena, such as “tickling yourself”.Finally, it has implications for the so-far unsolved question of how audi-tory alpha power is mediated by higher-order regions in a more generalsense.

Keywords: effective connectivity, efference copy, MEG, oscillation,top-down

Introduction

Even if our own voice is often intermingled with externalvoices, the brain can distinguish between speech sounds thatare produced by the brain itself and speech sounds that stemfrom the external world. A vast amount of literature indicatesthat the auditory cortex is inhibited when we process self-generated compared with played-back speech sounds. Most ofthese studies looked at evoked potentials or evoked magneticfields (Curio et al. 2000; Houde et al. 2002; Ford and Mathalon2004; Heinks-Maldonado et al. 2005; Martikainen et al. 2005;Baess et al. 2011) and showed that evoked activity is reducedin amplitude for self-generated speech sounds compared withexternally played-back speech sounds even if they had thesame (or similar) physical characteristics. Most of these resultsare interpreted in the framework of the so-called “efferencecopies”, meaning that the motor system is sending a copy ofthe motor command to the respective sensory area, where cor-ollary discharge elicited by this copy is combined with thesensory feedback (Holst and Mittelstaedt 1950; Sperry 1950;Ford and Mathalon 2004). Beyond that, studies on monkeysshow that self-produced vocalization lead to reduced neuronalfiring rates in a majority of auditory cortical neurons (Ploog

1981; Eliades and Wang 2003). In line with that, recordings inepilepsy patients disclosed a suppression of ongoing activity inmiddle and superior temporal gyrus neurons (Creutzfeldt et al.1989) and a suppression of gamma power in the temporal lobeduring speech production (Towle et al. 2008; Flinker et al.2010). Most interestingly, animal data (Eliades and Wang 2003)and also the data derived from the intracranial recordings byCreutzfeld and colleagues (2003) point to a suppression of brainactivity starting already a few hundred milliseconds before soundonset. These findings suggest that the suppression of neuronalactivity in the auditory cortex results could, in part, result frominternal modulatory mechanisms prior to sound onset.

It has been demonstrated that synchronous oscillatory activ-ity in the alpha frequency band (∼10 Hz) are inversely relatedto the excitability of the respective brain regions (Klimeschet al. 2007; Jensen and Mazaheri 2010), an assumption that hasrecently received strong support from invasive recordings(Haegens et al. 2011). An increase of alpha power in a sensoryregion is associated with a functional inhibition of that regionwhen sensory stimuli are processed (Jensen and Mazaheri2010). This has been shown in the visual modality (Wordenet al. 2000; Thut 2006; Romei et al. 2008; Siegel et al. 2008; vanDijk et al. 2008; Bahramisharif et al. 2010; Hanslmayr et al.2011), in the somatosensory modality (Jones et al. 2010;Haegens et al. 2012; Lange et al. 2012) and recently also in theauditory modality (Gomez-Ramirez et al. 2011; Muller andWeisz 2012; Weisz et al. 2014; Frey et al. 2014). The aim of thepresent study was to investigate if the aforementioned inhib-ition of the auditory cortex prior and during speech productioncan also be explained by a top-down modulation of auditoryalpha power, preceding voice onset. Crucially, any differencesin neuronal activity due to differences in sound characteristics(own voice vs. played-back own voice) can be ruled out, bymeasuring brain signals generated in the time intervals preced-ing sound onset. Our prediction being, that the inhibition ofthe auditory cortex for self-spoken versus played-back voicesbecomes evident in a relative increase in auditory alpha power.Such a finding would give evidence on the processes preced-ing the modulations of evoked activity in the context of voiceperception and would, for the first time, provide evidence on apossible internal mechanism modulating auditory cortex excit-ability when expecting self-generated sensory input.

Materials and Methods

ParticipantsTwenty right-handed volunteers reporting normal hearing participatedin the current study (9 m/11 f, mean age 22.6). Participants were

© The Author 2014. Published by Oxford University Press. All rights reserved.For Permissions, please e-mail: journals.permissions@oup.com

Cerebral Cortex November 2015;25:4029–4037doi:10.1093/cercor/bhu117Advance Access publication June 5, 2014

at Erlangen N

uernberg University on A

ugust 15, 2016http://cercor.oxfordjournals.org/

Dow

nloaded from

recruited via flyers posted at the University of Konstanz and were paidfollowing the experiment. The Ethics Committee of the University ofKonstanz approved the experimental procedure and all participants gavetheir written informed consent prior to taking part in the study. Two par-ticipants had to be excluded due to an excessive amount of artifacts.

Experimental ProcedureFirstly, participants were introduced to the lab facilities and informedabout the experimental procedure, which consisted of 2 phases (voicerecordings and main magnetoencephalography [MEG] experiment).For the voice recordings participants were asked to repeat the sound“Aah” 50 times, while their voice was recorded by means of a micro-phone (Zoom H4 USB-microphone). Then on- and off-set of each“Aah”-sound was determined and cut out automatically by a Matlabscript so that 50 sound files resulted. After verifying manually that thesounds were cut out correctly they were copied to the stimulation com-puter for the subsequent MEG experiment. The voice recordings weredone in order to keep physical characteristics of the self-spoken andexternally played-back sounds as similar as possible. Loudness of thesounds was adjusted later in the MEG so that participants perceived theself-spoken and the externally played-back sounds as equally loud. Forthis purpose a random “Aah”-sound was selected and presented to theparticipant in the MEG scanner. Participants had to rate if the played-back sound was louder or weaker compared with the self-spoken sound,whereupon loudness of the played-back sound was adjusted. This pro-cedure was repeated until the participant rated the played-back and self-spoken sounds as equally loud. After that the root mean square ampli-tude of the other recorded sounds was matched to the selected referencesound.

Subsequently, the individual headshapes were collected and themain experiment, consisting of 4 blocks, started. In half of the 4 blocksparticipants were instructed to say the sound “Aah” after a go-signalwhile in the other half of the blocks they were asked to listen to thesound “Aah” (that was randomly taken from the 50 “Aah”-sounds gen-erated before the experiment). Each experimental trial started with abaseline period of 500 ms, upon which a red fixation cross was shownfor 1.5 s (preparation period). After 1.5 s, the red fixation cross turnedinto a green one, which was the go-signal instructing participants toeither say the sound “Aah” (speak condition) or listen to it (listen con-dition). The next trial started 2–3 s after sound-offset. There were atotal of 200 trials. The presentation of visual and auditory stimulus ma-terial during MEG recordings was controlled using Psyscope X (Cohenet al. 1993), an open-source environment for the design and control ofbehavioral experiments (http://psy.ck.sissa.it/) and R version 2.11.1 forMac OS X (http://www.R-project.org). The procedure of the experimentis illustrated in Figure 1.

Data AcquisitionThe MEG recordings were carried out using a 148-channel whole-headmagnetometer system (MAGNESTM 2500 WH, 4D Neuroimaging, SanDiego, USA) installed in a magnetically shielded chamber (Va-kuumschmelze Hanau). Prior to the recordings, individual headshapes were collected using a digitizer. Participants lay in a comfort-able supine position and were asked to keep their eyes open and tofocus on the fixation cross displayed by a video projector (JVCTM,DLA-G11E) outside of the MEG chamber and projected to the ceiling inthe MEG chamber by means of a mirror system. Participants were in-structed to hold still and to avoid eye blinks and movements as best aspossible. A video camera installed inside the MEG chamber allowedthe investigator to monitor participants throughout the experiment.MEG signals were recorded with a sampling rate of 678.17 Hz and ahardwired high-pass filter of 0.1 Hz. The recorded and RMS–matched“Aah”-sounds (see above) were presented through a tube system witha length of 6.1 m and a diameter of 4 mm (Etymotic Research, ER30).Structural images were acquired with a Philips MRI Scanner (PhilipsGyroscan ACS-T 1.5 T, field of view 256 × 256 × 200 sagittal slices).

Data Analysis

PreprocessingWe analyzed the data sets using Matlab (The MathWorks, Natick, MA,Version 7.5.0 R 2007b) and the Fieldtrip toolbox (Oostenveld et al.2011). From the raw continuous data, we extracted epochs of 5 slasting from 2.5 s before onset of the red fixation cross to 2.5 s afteronset of the red fixation cross. This was done for the 2 conditions sep-arately (self-spoken sound, played-back sound) and resulted in 100trials for each condition. As participants could not avoid blinking suffi-ciently we decided to perform an independent component analysis(ICA) in order to minimize the influence of the blinks. For ICA correc-tion we first did a coarse visual artifact rejection, removing trials includ-ing strong muscle artifacts and dead or very noisy channels. Aftercoarse artifact rejection the data sets (concatenated across conditions)were downsampled to 300 Hz. On a subset of trials an ICA was per-formed (RUNICA, Delorme and Makeig 2004) and the affected com-ponents (eye movements) visually selected. After that ICA was againapplied to the data sets of the 2 original conditions and the raw datawere reconstructed with the respective components removed. Finally,the resulting data sets were again visually inspected for artifacts and theresidual artifactual trials rejected. To ensure a similar signal-to-noise ratioacross conditions, the trial numbers were equalized for the comparedconditions (self-spoken vs. played-back) by random omission (60–90trials remained). Finally, data were downsampled to 500 Hz.

Evoked ActivityIn order to replicate the results of previous studies for quality controlpurposes, we assessed the evoked activity elicited by the soundstimuli. First, data were high-pass filtered by 1 Hz and low-pass filteredby 45 Hz. Evoked activity was obtained by averaging the single trials.This was done for both conditions separately (self-spoken vs. played-back, equal trial numbers). Evoked activity was then tested statisticallyby point-wise 2-tailed paired samples t-tests.

Spectral Power AnalysesTime–frequency distributions of the epochs preceding self-spoken andexternally played-back sounds were compared at the sensor andsource level. We estimated task-related changes in oscillatory powerusing a multitaper FFT time–frequency transformation (Percival 1993)

Figure 1. Experimental design. Each experimental trial began with a baseline periodof 500 ms, upon that a red fixation cross was shown for 1.5 s (preparation period).After 1.5 s, the red fixation cross turned into a green one, upon that participants wereinstructed to either say the sound “Aah” (self-spoken condition) or listen to it(play-back condition). The next trial started 2–3 s after sound-offset. In total therewere 200 trials.

4030 Absence of Alpha Suppression During Speech Preparation • Müller et al.

at Erlangen N

uernberg University on A

ugust 15, 2016http://cercor.oxfordjournals.org/

Dow

nloaded from

with frequency-dependent Hanning tapers (time window: Δt = 4/fsliding in 50 ms steps). We calculated power from 3 to 30 Hz in steps of1 Hz and for both conditions separately. The obtained time–frequencyrepresentations were then baseline normalized (baseline: −400 to−100 ms before onset of the red fixation cross, relative change).

In order to test if power modulations are significantly differentbetween conditions (expecting self-spoken vs. played-back own voice)we performed a nonparametric cluster-based permutation test on thebaseline-normalized time–frequency representations (Maris and Oos-tenveld 2007), test based on 2-tailed paired t-tests). This test waschosen to correct for multiple comparisons.

As a next step, we estimated the generators of the sensor effects insource space using the frequency-domain adaptive spatial filteringalgorithm Dynamic Imaging of Coherent Sources (DICS, Gross et al.2001). For each participant an anatomically realistic headmodel (Nolte2003) was created and leadfields for a 3-dimensional grid covering theentire brain volume (resolution: 1 cm) calculated. Together, with thesensor-level cross-spectral density matrix (2 time intervals early 0.5–1 sand late 1–1.5 s, 13 ± 3 Hz, multitaper analysis, conditions concate-nated), we could estimate common spatial filters, optimally passinginformation for each grid point while attenuating influences fromother regions for the frequency and timewindow of interest (accordingto the cluster permutation test at sensor level: 0.5–1.5 s, 13 ± 3 Hz). Thecommon spatial filters were then applied to the Fourier-transformeddata for both conditions separately (same parameters). After that theresulting activation volumes were interpolated onto the individualMRI. In cases where we could not get a structural scan (5 out of 18), wecreated “pseudo”-individual MRIs that were created based on an affinetransformation of the headshape of an Montreal Neurological Institute(MNI) template and the individually gained headshape points. The in-terpolated activation volumes were then normalized to a template MNIbrain provided by the SPM8 toolbox (http://www.fil.ion.ucl.ac.uk/spm/software/spm8). Finally, source solutions for the 2 conditionswere compared using a voxel-wise dependent samples t-statistic. Fromthat analysis the left auditory cortex (Brodman Areas 21/22 andBrodman Area 41), the right precentral cortex and the medial prefront-al cortex (BA 8) were derived as main regions showing a significant in-crease of alpha power for self-generated versus externally played-backsounds. This is illustrated in the results. To get a better estimate of howalpha power in the auditory cortex is modulated we averaged thepower within the left auditory cortex for each participant and for bothconditions separately. These values were then tested against baselinevalues by 2-tailed paired t-tests and for both conditions separately.Beyond that, we tested the baseline values of the speak condition againstthe baseline values of the listen condition again by a 2-tailed paired t-testto rule out the possibility that the relative effects were determined due tobaseline differences.

Power–Power CorrelationsAfter spectral power analysis we aimed at shedding light onto the ques-tion of how the condition-specific relative alpha increases in the audi-tory cortex are mediated. We therefore correlated left auditory alphapower with low-frequency power (from 2–26 Hz) in all other regionsof the brain (for 1–1.5 s). We did this in MNI grid space.

First, a template grid was created (using a template head modelbased on a segmented template MNI brain provided by the SPM8toolbox). Using this template grid an individual grid was generated bywarping the template grid to the individual MRI for each participantseparately. Importantly, the warped individual grids have an equalnumber of points with equal positions in MNI space, so that the indi-vidual grids of different participants can be compared directly (gridpoints of Subject 1 correspond to grid points of Subject 2).

These individual MNI grids were then used for source analysis.Source analysis was done for the single trials and using the DICS beam-former algorithm (MNI grid, 1–1.5 s after red-trigger onset, 13 Hz ± 3,same settings as for alpha power source analysis despite the use of theindividual MNI grids). We calculated source solutions for frequenciesfrom 2 to 26 Hz in increments of 2 Hz. Thereby, power values for eachparticipant, each condition, each trial, each frequency and each gridpoint were obtained. We then calculated correlations between alphapower at the reference voxel, which was defined as the grid pointbeing closest to the main alpha power effect as derived from source

analysis (MNI coordinates: −55 −28 2, left auditory cortex), and allother grid points. We repeated this for all frequencies (2–26 Hz) andfisher z-transformed the correlation values afterwards. We thereby ob-tained a 2-D matrix for both conditions (grid points × frequencies).Afterwards, the frequency × grid point maps were tested for significantdifferences between conditions across subjects using a nonparametriccluster-based permutation test (Maris and Oostenveld 2007); neighborswere defined as grid points that had a distance of <3 cm resulting inaverage 75 neighbors per grid point, which reflects ∼3% of all gridpoints). This analysis yielded that alpha power in the left auditorycortex is strongly correlated with low-frequency power (6–14 Hz) inthe medial prefrontal cortex when participants expect a self-generatedsound. To get a better estimate of how connectivity between themedial prefrontal cortex and the auditory cortex is modulated in bothconditions separately we averaged the correlation values within thesignificant region for each participant and for both conditions separ-ately. These values were then tested against correlation values (withinthe same region) that were obtained during the baseline period by2-tailed paired t-tests, and for both conditions separately.

Partial Directed Coherence Between Auditory Cortex and MedialFrontalAs a final step we wanted to elucidate the direction of information flowbetween the auditory cortex (MNI coordinates: −55 −28 2) and themedial prefrontal cortex (peak voxel of correlation effect, MNI coordi-nates: −4 44 −6), assessed via partial directed coherence (PDC, Baccalaand Sameshima 2001). PDC is a measure of effective coupling that isbased on multivariate autoregressive (MVAR) modeling. For a pair ofvoxels the information flow can be assessed in both directions. We firstprojected the raw time series into source space by multiplying the rawtime series for both conditions separately with a common spatial filter.The spatial filter was created using the LCMV beamformer (Van Veenet al. 1997) and the concatenated data of both conditions (2–26 Hz,time window including baseline and activation −0.5 to 1.5 s). Wethereby obtained time series for both conditions and both sources(auditory cortex, medial prefrontal cortex) separately. For these timeseries a MVAR model was fitted (“bsmart”). The model order was set to15, according to previous analysis approaches (Supp et al. 2007; Weiszet al. 2014). Then a Fourier transform was performed on the resultingcoefficients of the MVAR model. These Fourier-transformed coeffi-cients were then used to calculate partial directed coherence betweenthe auditory and medial prefrontal cortex. The PDC values were base-line normalized using the baseline interval (−0.5 to 0) by first subtract-ing and then dividing the values by the values of the baseline interval.Finally, the PDC values were tested for differences between conditions(speak vs. listen) using paired t-tests.

Results

The current study aimed at disentangling brain activity precedingthe processing of participants’ own voice that was either self-spoken or played-back externally. We investigated brain activityon a local and on a network level in the time interval beforevoice onset and with a focus on low-frequency oscillatory power.

Evoked ResponsesThe event related response was significantly stronger for theexternally played-back speech sound compared with the self-generated one between 150 and 200 ms after sound onset (un-corrected). This is comparable to the previous literature (Curioet al. 2000; Houde et al. 2002; Flinker et al. 2010). Results areshown in Figure 2 (upper panel).

Pre-voice Power Differences—Sensor LevelIn a first step, we assessed differences in low-frequency powerfor self-spoken versus externally played-back voices on sensorlevel. We found a significant power increase (cluster P < 0.05)peaking between 10 and 16 Hz and encompassing frontal and

Cerebral Cortex November 2015, V 25 N 11 4031

at Erlangen N

uernberg University on A

ugust 15, 2016http://cercor.oxfordjournals.org/

Dow

nloaded from

left temporal sensors for self-generated versus externally played-back speech sounds. Interestingly, at a descriptive level, powermodulations at frontal sensors were most dominant in the firstpart of the preparation period, while left temporal power modu-lations became stronger towards the end of the preparationperiod, shortly before voice onset. For comparison see Figure 2(middle panel).

Source Localization of Alpha Power Differences BeforeVoice OnsetIn order to get a better estimate of where in the brain the low-frequency power modulations take place, we performed asource analysis of the power modulations derived from sensorlevel (10–16 Hz, 0.5–1 s and 1–1.5 s after onset of the red fix-ation cross). Source results indicate that, besides extra-auditoryareas (right precentral, medial dorsolateral prefrontal cortex),the left auditory cortex shows a strong relative increase ofalpha power, becoming most evident in the last part of thepreparation period (1–1.5 s, P < 0.01, including Brodman areas21/22/41). There were no significant differences in auditory ac-tivity during baseline (P = 0.19). Interestingly, when extractingthe power modulations for the 2 conditions separately, itturned out that the relative increase of alpha power is due to adecrease of alpha power compared with baseline when partici-pants prepare to process their externally played-back voice(P < 0.01), while this alpha power decrease seems to be abol-ished (no statistical difference compared with baseline) whenparticipants expect to process their self-spoken voice. For com-parison see Figure 2 (lower panel).

Beside the auditory power modulations, alpha power wasincreased in the medial dorsolateral prefrontal cortex (P < 0.01,BA 9 and 32) and the right precentral cortex (P < 0.01, BA 4).Within these extra-auditory regions alpha power was signifi-cantly increased compared with baseline when participantsprepared to speak (P < 0.01), while we found no significant dif-ferences in alpha power compared with baseline when partici-pants prepared to listen (Fig. 3).

Note, that alpha power in the right precentral region wasalready modulated during baseline (probably due to theblocked design). During baseline alpha power was significant-ly decreased in the “speak blocks” compared with the “listenblocks”. Baseline differences of the entire brain are shown inSupplementary Fig. 1.

Modulations of Connectivity with the Left AuditoryCortex Before Voice OnsetThe second main analysis tackled the question of how thecondition-specific alpha power modulations in the auditorycortex are mediated. Therefore we looked at power–power cor-relations between alpha power in the left auditory cortex andlow-frequency power in the other regions of the brain. Thiswas conducted on a single-trial level and for the time intervalpreceding voice onset when auditory alpha power modula-tions were strongest (1–1.5 s after onset of the red fixationcross). We took strong power–power correlations as an indica-tor of a possible communication between the accordant brainregions (Park et al. 2011). A cluster-based permutation testrevealed the medial prefrontal cortex as the main region dif-ferentially communicating with the left auditory cortex whenparticipants expected to listen to their self-produced versus ex-ternally played-back voice (cluster P < 0.05, BA 11). The effectwas strongest for power correlations between 6 and 14 Hz.When looking at the effect more precisely, it turned out thatcommunication between the left auditory cortex and themedial prefrontal cortex was significantly enhanced comparedwith baseline when participants expected their self-spokenvoice (P < 0.05) and significantly reduced compared with base-line when they expected their played-back voice (P < 0.05). SeeFigure 4 (upper panel) for comparison.

Figure 2. Auditory results. The upper panel shows the event-related activity elicitedby self-generated (blue) versus externally played-back (red) voice sounds. The left sidedisplays global field power at left temporal sensors averaged across participants. Thetopography for the significant time interval is shown at the right side. Black dots showsignificant sensors (uncorrected). The event-related activity following the self-generated sound is clearly diminished compared with the externally played-back one(statistically significant in shaded area between 150 and 200 ms after sound onset andat left and right sensors, uncorrected). This points to an inhibition of the auditory cortex,when participants process self-generated compared with externally played-back speechsounds. The middle panel displays the time–frequency representation of powerdifferences between self-produced and externally played-back speech sounds, in theinterval before voice onset. Shown are t-values. Higher values (red) indicate a relativeincrease of power when participants expect to process their self-spoken versusplayed-back voice. Power between 10–16 Hz is significantly increased at rather frontalsensors in the beginning of the preparation period and frontal and left temporal sensorsversus the end of the preparation period. The left lower panel shows how the alpha powermodulation (10–16 Hz) is distributed in the brain for the early and late preparation period(0.5–1 s and 1–1.5 s). Alpha power is relatively increased in the left auditory cortex.Shown are t-values that are masked with P<0.01. The right lower panel shows meanleft auditory power modulations during preparation (1–1.5 s) versus baseline for the 2conditions. Left auditory cortex alpha power is significantly reduced when participantsprepare to listen to the externally played-back voice (P<0.01), while there is nomodulation of alpha power when participants prepare to listen to a self-generated voice(no statistical difference compared with baseline in “speak”-condition).

4032 Absence of Alpha Suppression During Speech Preparation • Müller et al.

at Erlangen N

uernberg University on A

ugust 15, 2016http://cercor.oxfordjournals.org/

Dow

nloaded from

Finally we wanted to assess the direction of the informationflow between the left auditory and the medial prefrontalcortex. In order to do that, we calculated Partial Directed Co-herence between the 2 regions. Information flow from theauditory cortex to the medial prefrontal cortex showed no sig-nificant differences between conditions (all P > 0.05). In con-trast, information flow from the medial prefrontal cortex to theleft auditory cortex was significantly enhanced when partici-pants prepared for speaking (P < 0.05). The effect was stron-gest for the time interval slightly preceding the main auditoryalpha power modulations (0.7–1 s after onset of the red fix-ation cross). For comparison see Figure 4 (lower panel).

Discussion

In the present study we investigated if and how pre-speechbrain activity is modulated when participants expect a self-spoken sound. We concentrated on modulations in low-frequency power with a focus on the auditory cortex and itscommunication with non-auditory regions. Results show thatthe alpha power suppression, typically present when partici-pants expect sounds, is absent in the left auditory cortex whenparticipants expect their own voice. They further show that themedial prefrontal cortex mediates this effect. The absence ofthe auditory alpha power suppression can be interpreted asinhibition of the auditory cortex when participants expect self-spoken sounds. This is in line with the previous literature pos-tulating an inhibition of the auditory cortex when processingself-spoken sounds and extends the previous literature by

showing that brain activity in the auditory cortex is inhibitedalready before sound onset and on a macroscopic scale. So fara suppression of brain activity in the auditory cortex beforespeech onset has only been shown for ongoing activity insingle neurons (Creutzfeldt et al. 1989; Eliades and Wang2003) and not for local field potentials. In addition, the medi-ation of the auditory cortex’s increase in alpha power viamedial prefrontal cortex suggests a mechanism of how audi-tory cortex excitability is adjusted. This gives new insights intothe processes going on within the tested paradigm and, cru-cially, provides first evidence of how auditory alpha powercould be causally modulated by higher-order regions.

Auditory Alpha Power ModulationsAs described above we found a decrease of auditory alphapower compared with baseline in the “listen” condition and anabolishment of that effect in the “speak” condition. There wereno significant differences of auditory alpha activity during thebaseline. This points to a relative inhibition of that brain region(Klimesch et al. 2007; Jensen and Mazaheri 2010), in this case arelative inhibition of the auditory cortex (Gomez-Ramirez et al.2011; Muller and Weisz 2012; Weisz et al. 2014) and is thereforeconsistent with previous literature postulating an inhibition ofthe auditory cortex when processing self-generated sounds com-pared with externally played-back ones (Creutzfeldt et al. 1989;Curio et al. 2000; Houde et al. 2002; Eliades and Wang 2003;Ford and Mathalon 2004). A growing number of studies in thevisual and somatosensory system (Worden et al. 2000; Thut2006; Romei et al. 2008; Siegel et al. 2008; van Dijk et al. 2008;

Figure 3. Extra-auditory results. The left panel shows that alpha power is increased in the right precentral and the dorsolateral prefrontal cortex when participants expect toprocess self-spoken versus played-back voice. Shown are t-values that are masked with P<0.01. The right panel shows mean power modulations for these regions duringpreparation (1–1.5 s) versus baseline for the 2 conditions separately. In both regions (precentral and dorsolateral prefrontal) alpha power is significantly increased compared withbaseline when participants prepare to listen to their self-generated voice (P<0.01), while there is no modulation of alpha power when participants prepare to listen to externallyplayed-back voice (no statistical difference compared with baseline in “listen-condition).

Cerebral Cortex November 2015, V 25 N 11 4033

at Erlangen N

uernberg University on A

ugust 15, 2016http://cercor.oxfordjournals.org/

Dow

nloaded from

Jones et al. 2010; Händel et al. 2011; Haegens et al. 2012; Langeet al. 2012) and also in the auditory system (Gomez-Ramirezet al. 2011; Muller and Weisz 2012; Müller et al. 2013; Weiszet al. 2014; Frey et al. 2014) have convincingly shown that alphapower modulations have the potential to dynamically adjust theexcitability of brain regions with respect to the according taskdemands (e.g., attention, near-threshold detection, memory),and thereby make neuronal processing adaptive and most ef-fective (Jensen and Mazaheri 2010). Based on that the currentresults can be interpreted as follows: the auditory system is bydefault in an inhibited state to filter out the vast amount of audi-tory information it is exposed to. This is in accord with the ob-servation that alpha in sensory cortices is high when subjectsare awake and not engaged in any task (Basar et al. 1997; Kli-mesch et al. 2007; Jensen and Mazaheri 2010). If participantsexpect to process an external sound they reduce auditory alphapower in order to enhance processing capacities for the incom-ing auditory stimulus, as it is the case in “normal” auditory per-ception (Gomez-Ramirez et al. 2011; Hartmann et al. 2012;Muller and Weisz 2012; Weisz et al. 2014). In contrast, if partici-pants expect a self-generated sound, auditory alpha power iskept high meaning that, in that case, processing capacities arenot enhanced compared with baseline. We thus propose thateven if alpha power is not increased beyond baseline levels, the“relative increase” (i.e., the one arising from the condition con-trast) goes in line with the hypothesis of an inhibition of pro-cessing when participants prepare to listen to their self-spoken

voice. Such a mechanism could explain why we process andperceive self-spoken sounds differentially from played-backones. Interestingly, it has been shown that an increase in alphapower has an impact on neuronal firing (Haegens et al. 2011)and also on event-related responses (Basar and Stampfer 1985;Ergenoglu et al. 2004; Klimesch et al. 2007). The present find-ings could thus be the prerequisite of the reported inhibition ofthe auditory cortex during the processing of self-generatedsounds as reported in literature (Creutzfeldt et al. 1989; Curioet al. 2000; Houde et al. 2002; Eliades and Wang 2003; Ford andMathalon 2004; Heinks-Maldonado et al. 2005), however such adirect relation would have to be tested within further studies.Also findings postulating that the suppression in the auditorycortex is very specific for the self-generated sounds and doesnot block auditory processing in general thereby will have to betaken into account (McGuire et al. 1996; Heinks-Maldonadoet al. 2005; Fu et al. 2006). We here suggest that the abolishmentof the usual alpha power reduction when expecting self-generatedsounds is an active and top-down modulated process helping todifferentiate between self-spoken and externally played-backsounds. According to the connectivity results, which are explainedin more detail below, this seems to be indeed the case.

Left Auditory Alpha Power ModulationsWe found that the condition-specific alpha power modulationsare lateralized to the left auditory cortex. This is in line with

Figure 4. Connectivity results. The upper panel shows the spatial dimension of the significant cluster derived from power–power correlations with the left auditory cortex. Alphapower in the auditory cortex was significantly stronger correlated with low-frequency power in the medial prefrontal cortex (depicted by black circle) when participants expected tolisten to their self-spoken versus externally played-back voice (cluster P< 0.05, 1–1.5 s, 6–14 Hz, medial prefrontal cortex). The effect was due to a significant increase of powercorrelations compared with baseline when participants were expecting their self-produced voice (P<0.05) and a significant decrease of power correlations when they expectedtheir externally played-back voice (P<0.05). The lower panel depicts effective connectivity between the left auditory cortex and the medial prefrontal cortex, quantified by PartialDirected Coherence. Information flow from the auditory cortex to the medial prefrontal cortex did not differ between conditions (left). In contrast, information flow from the medialprefrontal cortex to the left auditory cortex was significantly stronger for the self-spoken versus externally played-back speech sounds (right), with the effect being strongest for thetime interval slightly preceding the main auditory alpha power modulations (0.7–1 s).

4034 Absence of Alpha Suppression During Speech Preparation • Müller et al.

at Erlangen N

uernberg University on A

ugust 15, 2016http://cercor.oxfordjournals.org/

Dow

nloaded from

the literature on the processing of self-generated speechsounds showing that the suppression effects are dominant inthe left auditory cortex (Curio et al. 2000; Houde et al. 2002;Heinks-Maldonado et al. 2005; Kauramaki et al. 2010).

Non-auditory Alpha Power ModulationsDespite the absence of alpha power suppression in the audi-tory cortex we derived an increase of alpha power in a pre-frontal region (BA 8), encompassing the medial dorsolateralprefrontal cortex and a power increase in the right precentralcortex. Brodman area 8 is involved in planning, cognitivecontrol, and maintaining attention (MacDonald 2000; Seamanset al. 2008) and also in guiding decisions (Seamans et al.2008). A modulation of that region during speech preparationcould point to an active disengagement from the expected andto-be-inhibited auditory input. However, the role of alpha powerin higher-order regions is not understood so far so that possiblemodulations within these prefrontal regions during speech prep-aration will have to be tested by further studies.

Concerning the precentral cortex, it is important to clarifythat the effect is due to differences in the baseline (see Supple-mentary Fig. 1 for comparison). During baseline alpha poweris reduced in the left and right precentral cortex for the“speak” compared with the “listen” condition. The left precen-tral alpha power decrease is still present in the preparationinterval what is in line with the modulations we would expectin the precentral cortex during motor preparation (Jasper andPenfield 1949; Pfurtscheller et al. 1996; Sauseng et al. 2009).Interestingly, however, the right precentral power decrease forthe “speak” versus “listen” conditions disappears, leading tothe impression of an increase of alpha power in the precentralcortex when subjects prepare to speak. These hemispheric dif-ferences could be due to the dominance of the left hemispherefor speech production (Llorens et al. 2011; Price et al. 2011).

Auditory Alpha Power Modulation Mediated by theMedial Prefrontal CortexAnother crucial question to answer was how the condition-specific auditory alpha power modulations are mediated. Wecould elucidate the medial prefrontal cortex (BA 11) as themain region showing increased communication in thisprocess. Crucially, this increase in communication was drivenby an increase of unilateral communication from the medialprefrontal cortex to the left auditory cortex. This provides clearevidence for a condition-specific modulation of auditory alphapower communicated by the medial prefrontal cortex. Themedial prefrontal cortex is involved in self-referential thinking(meta-analyses, Johnson et al. 2002; Heatherton et al. 2006;Northoff et al. 2006; van der Meer et al. 2010), in comparingthe self with others (Moore et al. 2013) and self-reflective judg-ments (Macrae et al. 2004). It is thus also from a theoreticalpoint of view very likely that the medial prefrontal cortex has acrucial role in mediating the excitability of the auditory cortexwhen we process our own voice. Interestingly, the increase ofinformation flow from the medial prefrontal cortex to the audi-tory cortex was strongest shortly before the relative increase ofauditory alpha power. All in all, we suggest that the medial pre-frontal cortex triggers alpha power in the auditory cortex so thatself-generated sounds are processed less intensely and we caneasily distinguish between self-generated and external sounds.

Conclusion

With the present study our aims were to disentangling brainactivity associated with the expectation of a self-spoken sound.We concentrated on alpha power modulations in the auditorycortex and on how these modulations are mediated by non-auditory brain regions. We can show that the typical alphapower suppression when participants expect external soundsis absent in the left auditory cortex when participants expectself-spoken sounds. This points to a relative inhibition of theauditory cortex that is already present before speech onset andis in line with the previous literature showing a suppression ofbrain activity mainly during sound production. Importantly,the current findings complement the existing evidence onmodulations of evoked activity and rather local modulations inauditory activity (as derived by ECoG and animal studies) byelucidating that also the state/excitability of the auditory cortexis modulated when processing self-generated sounds, whichbecame evident in the auditory alpha power modulations. Assecond main finding we demonstrate that the medial prefrontalcortex, a region known for self-referential processes, mediatesthese condition-specific alpha power modulations. This pro-vides crucial insights into how higher-order regions preparethe auditory cortex for the processing of self-generated soundsand seems interesting as a mechanism itself having the poten-tial to explain similar phenomena related to self-referentialprocessing like for instance “tickling yourself”. Beyond that,the findings also have implications for the so–far unsolvedquestion of how auditory alpha power is mediated by higher-order regions in a more general sense.

Supplementary MaterialSupplementary material can be found at: http://www.cercor.oxfordjournals.org/.

Funding

This work was supported by the European Research Council(grant number: 283404) and the Deutsche Forschungsge-meinschaft (grant number: 4156/2-1).

NotesWe thank Nick Peatfield for proofreading the manuscript. Conflict ofInterest: None declared.

ReferencesBaccala LA, Sameshima K. 2001. Partial directed coherence: a new

concept in neural structure determination. Biol Cybern. 84:463–474.Baess P, Horváth J, Jacobsen T, Schröger E. 2011. Selective suppression

of self-initiated sounds in an auditory stream: An ERP study. Psy-chophysiology. 48:1276–1283.

Bahramisharif A, van Gerven M, Heskes T, Jensen O. 2010. Covert at-tention allows for continuous control of brain–computer interfaces.Eur J Neurosci. 31:1501–1508.

Basar E, Schurmann M, Basar-Eroglu C, Karakas S. 1997. Alpha oscilla-tions in brain functioning: an integrative theory. Int J Psychophy-siol. 26:5–29.

Basar E, Stampfer HG. 1985. Important associations among EEG-dynamics, event-related potentials, short-term memory and learn-ing. Int J Neurosci. 26:161–180.

Cohen J, MacWhinney B, Flatt M, Provost J. 1993. PsyScope: an inter-active graphic system for designing and controlling experiments in

Cerebral Cortex November 2015, V 25 N 11 4035

at Erlangen N

uernberg University on A

ugust 15, 2016http://cercor.oxfordjournals.org/

Dow

nloaded from

the psychology laboratory using Macintosh computers. Behav ResMethods Instrum Comput. 25:257–271.

Creutzfeldt O, Ojemann G, Lettich E. 1989. Neuronal activity in thehuman lateral temporal lobe. Exp Brain Res. 77:451–475.

Curio G, Neuloh G, Numminen J, Jousmaki V, Hari R. 2000. Speakingmodifies voice evoked activity in the human auditory cortex. HumBrain Mapp. 9:183–191.

Delorme A, Makeig S. 2004. EEGLAB: an open source toolbox for ana-lysis of single-trial EEG dynamics including independent compo-nent analysis. J Neurosci Methods. 134:9–21.

Eliades SJ, Wang X. 2003. Sensory-motor interaction in the primateauditory cortex during self-initiated vocalizations. J Neurophysiol.89:2194–2207.

Ergenoglu T, Demiralp T, Bayraktaroglu Z, Ergen M, Beydagi H, UresinY. 2004. Alpha rhythm of the EEG modulates visual detection per-formance in humans. Cogn Brain Res. 20:376–383.

Flinker A, Chang EF, Kirsch HE, Barbaro NM, Crone NE, Knight RT.2010. Single-trial speech suppression of auditory cortex activity inhumans. J Neurosci. 30:16643–16650.

Ford JM, Mathalon DH. 2004. Electrophysiological evidence of corol-lary discharge dysfunction in schizophrenia during talking andthinking. J Psychiatr Res. 38:37–46.

Frey J, Mainy N, Lachaux JP, Müller N, Bertrand O, Weisz N. 2014. Se-lective modulation of auditory cortical alpha activity in an audiovi-sual spatial attention task. J Neurosci. 34(19):6634–6639.

Fu CH, Vythelingum GN, Brammer MJ, Williams SCR, Amaro E,Andrew CM, Yágüez L, van Haren NEM, Matsumoto K, McGuire PK.2006. An fMRI study of verbal self-monitoring: neural correlates ofauditory verbal feedback. Cereb Cortex. 16:969–977.

Gomez-Ramirez M, Kelly SP, Molholm S, Sehatpour P, Schwartz TH,Foxe JJ. 2011. Oscillatory sensory selection mechanisms during inter-sensory attention to rhythmic auditory and visual inputs: a humanelectrocorticographic investigation. J Neurosci. 31:18556–18567.

Gross J, Kujala J, Hämäläinen M, Timmermann L, Schnitzler A, Salme-lin R. 2001. Dynamic imaging of coherent sources: studying neuralinteractions in the human brain. Proc Natl Acad Sci. 98:694–699.

Haegens S, Luther L, Jensen O. 2012. Somatosensory anticipatoryalpha activity increases to suppress distracting input. J Cogn Neu-rosci 24:677–685.

Haegens S, Nácher V, Luna R, Romo R, Jensen O. 2011. α-Oscillationsin the monkey sensorimotor network influence discrimination per-formance by rhythmical inhibition of neuronal spiking. Proc NatlA-cad Sci. 108:19377–19382.

Händel BF, Haarmeier T, Jensen O. 2011. Alpha oscillations correlatewith the successful inhibition of unattended stimuli. J Cogn Neuros-ci. 23:2494–2502.

Hanslmayr S, Gross J, Klimesch W, Shapiro KL. 2011. The role of alphaoscillations in temporal attention. Brain Res Rev. 67:331–343.

Hartmann T, Schlee W, Weisz N. 2012. It’s only in your head: expectancyof aversive auditory stimulation modulates stimulus-induced auditorycortical alpha desynchronization. NeuroImage. 60:170–178.

Heatherton TF, Wyland CL, Macrae CN, Demos KE, Denny BT, KelleyWM. 2006. Medial prefrontal activity differentiates self from closeothers. Soc Cogn Affect Neurosci. 1:18–25.

Heinks-Maldonado TH, Mathalon DH, Gray M, Ford JM. 2005. Fine-tuning of auditory cortex during speech production. Psychophysi-ology. 42:180–190.

Holst E, Mittelstaedt H. 1950. Das Reafferenzprinzip. Naturwissenschaften.37:464–476.

Houde JF, Nagarajan SS, Sekihara K, Merzenich MM. 2002. Modulationof the auditory cortex during speech: an MEG study. J Cogn Neuros-ci. 14:1125–1138.

Jasper H, Penfield W. 1949. Electrocorticograms in man: Effect of voluntarymovement upon the electrical activity of the precentral gyrus. Archiv fürPsychiatrie und Zeitschrift Nervenkrankheiten. 183:163–174.

Jensen O, Mazaheri A. 2010. Shaping functional architecture by oscilla-tory alpha activity: gating by inhibition. Front Hum Neurosci. 4:186.

Johnson SC, Baxter LC, Wilder LS, Pipe JG, Heiserman JE, PrigatanoGP. 2002. Neural correlates of self reflection. Brain. 125:1808–1814.

Jones SR, Kerr CE, Wan Q, Pritchett DL, Hämäläinen M, Moore CI. 2010.Cued spatial attention drives functionally relevant modulation of the

Mu rhythm in primary somatosensory cortex. J Neurosci. 30:13760–13765.

Kauramaki J, Jaaskelainen IP, Hari R, Mottonen R, Rauschecker JP,Sams M. 2010. Lipreading and covert speech production similarlymodulate human auditory-cortex responses to pure tones. J Neuros-ci. 30:1314–1321.

Klimesch W, Sauseng P, Hanslmayr S. 2007. EEG alpha oscillations: theinhibition–timing hypothesis. Brain Res Rev. 53:63–88.

Lange J, Halacz J, van Dijk H, Kahlbrock N, Schnitzler A. 2012. Fluctua-tions of prestimulus oscillatory power predict subjective perceptionof tactile simultaneity. Cereb Cortex. 22:2564–2574.

Llorens A, Trébuchon A, Liégeois-Chauvel C, Alario F-X. 2011. Intra-cranial recordings of brain activity during language production.Front Psychol. 2:375.

MacDonald AW. 2000. Dissociating the role of the dorsolateral prefrontaland medial dorsolateral prefrontal cortex in cognitive control. Science.288:1835–1838.

Macrae CN, Moran JM, Heatherton TF, Banfield JF, Kelley WM. 2004.Medial prefrontal activity predicts memory for self. Cereb Cortex.14:647–654.

Maris E, Oostenveld R. 2007. Nonparametric statistical testing of EEG-and MEG-data. J Neurosci Methods. 164:177–190.

Martikainen MH, Kaneko K, Hari R. 2005. Suppressed responses toself-triggered sounds in the human auditory cortex. Cereb Cortex.15:299–302.

McGuire PK, Silbersweig DA, Frith CD. 1996. Functional neuroanat-omy of verbal self-monitoring. Brain. 119(Pt 3):907–917.

Moore WE, Merchant JS, Kahn LE, Pfeifer JH. 2013. “Like me?”: ventro-medial prefrontal cortex is sensitive to both personal relevance andself-similarity during social comparisons. Soc Cogn Affect Neurosci.9(4):421–426.

Müller N, Keil J, Obleser J, Schulz H, Grunwald T, Bernays R-L, Hup-pertz H-J, Weisz N. 2013. You can’t stop the music: reduced audi-tory alpha power and coupling between auditory and memoryregions facilitate the illusory perception of music during noise.NeuroImage. 79:383–393.

Muller N, Weisz N. 2012. Lateralized auditory cortical alpha band activ-ity and interregional connectivity pattern reflect anticipation oftarget sounds. Cereb Cortex. 22:1604–1613.

Nolte G. 2003. The magnetic lead field theorem in the quasi-static ap-proximation and its use for magnetoencephalography forwardcalculation in realistic volume conductors. Phys Med Biol. 48:3637.

Northoff G, Heinzel A, de Greck M, Bermpohl F, Dobrowolny H, Pank-sepp J. 2006. Self-referential processing in our brain—a meta-analysisof imaging studies on the self. NeuroImage. 31:440–457.

Oostenveld R, Fries P, Maris E, Schoffelen J-M. 2011. FieldTrip: opensource software for advanced analysis of MEG, EEG, and invasiveelectrophysiological data. Comput Intell Neurosci 2011:1–9.

Park H, Kang E, Kang H, Kim JS, Jensen O, Chung CK, Lee DS. 2011.Cross-frequency power correlations reveal the right superior tem-poral gyrus as a hub region during working memory maintenance.Brain Connectivity. 1:460–472.

Percival DB. 1993. Spectral analyses for physical applications.Cambridge: Cambridge University Press.

Pfurtscheller G, Stancak AJ, Neuper C. 1996. Event-related synchron-ization (ERS) in the alpha band —an electrophysiological correlateof cortical idling: a review. Int J Psychophysiol. 24:39–46.

Ploog D. 1981. Inhibition of auditory cortical neurons during phon-ation. Brain Res. 215:61–76.

Price CJ, Crinion JT, Macsweeney M. 2011. A generative model ofspeech production in Broca’s and Wernicke’s Areas. Front Psychol.2:237.

Romei V, Rihs T, Brodbeck V, Thut G. 2008. Resting electroencephalo-gram alpha-power over posterior sites indexes baseline visualcortex excitability. Neuroreport. 19:203–208.

Sauseng P, Klimesch W, Gerloff C, Hummel FC. 2009. Spontaneouslocally restricted EEG alpha activity determines cortical excitabilityin the motor cortex. Neuropsychologia. 47:284–288.

Seamans JK, Lapish CC, Durstewitz D. 2008. Comparing the prefrontalcortex of rats and primates: insights from electrophysiology. Neuro-tox Res. 14:249–262.

4036 Absence of Alpha Suppression During Speech Preparation • Müller et al.

at Erlangen N

uernberg University on A

ugust 15, 2016http://cercor.oxfordjournals.org/

Dow

nloaded from

Siegel M, Donner TH, Oostenveld R, Fries P, Engel AK. 2008. Neuronalsynchronization along the dorsal visual pathway reflects the focusof spatial attention. Neuron. 60:709–719.

Sperry RW. 1950. Neural basis of the spontaneous optokineticresponse produced by visual inversion. J Comp Physiol Psychol.43:482–489.

Supp GG, Schlögl A, Trujillo-Barreto N, Müller MM. 2007. DirectedCortical Information Flow during Human Object Recognition: Ana-lyzing Induced EEG Gamma-Band Responses in Brain’s SourceSpace. PLoS ONE. 2(8):e684.

Thut G. 2006. -Band Electroencephalographic Activity over OccipitalCortex Indexes Visuospatial Attention Bias and Predicts VisualTarget Detection. J Neurosci. 26:9494–9502.

Towle VL, Yoon HA, Castelle M, Edgar JC, Biassou NM, Frim DM, SpireJP, Kohrman MH. 2008. ECoG gamma activity during a languagetask: differentiating expressive and receptive speech areas. Brain.131:2013–2027.

Van der Meer L, Costafreda S, Aleman A, David AS. 2010. Self-reflectionand the brain: a theoretical review and meta-analysis of neuroimagingstudies with implications for schizophrenia. Neurosci Biobehav Rev34:935–946.

Van Dijk H, Schoffelen JM, Oostenveld R, Jensen O. 2008. PrestimulusOscillatory Activity in the Alpha Band Predicts Visual Discrimin-ation Ability. J Neurosci. 28:1816–1823.

Van Veen BD, Van Drongelen W, Yuchtman M, Suzuki A. 1997. Local-ization of brain electrical activity via linearly constrained minimumvariance spatial filtering. Biomedical Engineering. 44:867–880.

Weisz N, Muller N, Jatzev S, Bertrand O. 2014. Oscillatory alpha modu-lations in right auditory regions reflect the validity of acoustic cuesin an auditory spatial attention task. Cereb Cortex. 24:2579–2590.

Worden MS, Foxe JJ, Wang N, Simpson GV. 2000. Anticipatory biasing ofvisuospatial attention indexed by retinotopically specific alpha-bandelectroencephalography increases over occipital cortex. J Neurosci.20:RC63.

Cerebral Cortex November 2015, V 25 N 11 4037

at Erlangen N

uernberg University on A

ugust 15, 2016http://cercor.oxfordjournals.org/

Dow

nloaded from