ARTICLE DE RECHERCHE - Recherche en IA spécialisée au MIT

ARTICLE DE RECHERCHE

Predictive Coding and Internal Error Correction in
Speech Production

un accès ouvert

journal

Alex Teghipco1

, Kayoko Okada2, Emma Murphy2, and Gregory Hickok1

1Department of Cognitive Sciences, Université de Californie, Irvine, Californie, Etats-Unis
2Département de psychologie, Loyola Marymount University, Les anges, Californie, Etats-Unis

Mots clés: speech production, IRMf, internal error correction, predictive coding, internal models,
imagined speech, overt speech, nonwords, taboo words, tongue twisters

ABSTRAIT

Speech production involves the careful orchestration of sophisticated systems, yet overt
speech errors rarely occur under naturalistic conditions. The present functional magnetic
resonance imaging study sought neural evidence for internal error detection and correction by
leveraging a tongue twister paradigm that induces the potential for speech errors while
excluding any overt errors from analysis. Previous work using the same paradigm in the
context of silently articulated and imagined speech production tasks has demonstrated forward
predictive signals in auditory cortex during speech and presented suggestive evidence of
internal error correction in left posterior middle temporal gyrus (pMTG) on the basis that this
area tended toward showing a stronger response when potential speech errors are biased
toward nonwords compared to words (Okada et al., 2018). The present study built on this prior
work by attempting to replicate the forward prediction and lexicality effects in nearly twice as
many participants but introduced novel stimuli designed to further tax internal error correction
and detection mechanisms by biasing speech errors toward taboo words. The forward
prediction effect was replicated. While no evidence was found for a significant difference in
brain response as a function of lexical status of the potential speech error, biasing potential
errors toward taboo words elicited significantly greater response in left pMTG than biasing
errors toward (neutral) words. Other brain areas showed preferential response for taboo words
as well but responded below baseline and were less likely to reflect language processing as
indicated by a decoding analysis, implicating left pMTG in internal error correction.

INTRODUCTION

Speaking is a deceptively complex task involving several computational stages: selecting
words from a mental dictionary that can contain tens of thousands of entries, correctly acces-
sing and coding the sequence of sounds that could have many possible permutations, et
executing the motor commands to reproduce those sounds with coordinated movements of
several independent articulators within the vocal tract (Levelt, 1989). Given the system’s com-
plexity, the ample opportunities for error, and the fact that speech is articulated at a rate of
approximately five syllables per second (Jacewicz et al., 2010), it is remarkable that the vast
majority of words are accurately produced (estimates put the number at approximately 99.9%;
voir, par exemple., Garnham et al., 1982; Levelt, 1992). How is this achieved? One possibility is that the
mechanism is so exquisitely tuned that it simply makes very few coding errors. Another pos-
sibility is that coding errors occur more frequently, but are unconsciously detected and

Citation: Teghipco, UN., Okada, K.,
Murphy, E., & Hickok, G. (2023).
Predictive coding and internal error
correction in speech production.
Neurobiology of Language, 4(1),
81–119. https://est ce que je.org/10.1162
/nol_a_00088

EST CE QUE JE:
https://doi.org/10.1162/nol_a_00088

Reçu: 16 Mars 2022
Accepté: 2 Novembre 2022

Intérêts concurrents: Les auteurs ont
a déclaré qu'aucun intérêt concurrent
exister.

Auteur correspondant:
Alex Teghipco
ateghipc@uci.edu

Éditeur de manipulation:
Kate Watkins

droits d'auteur: © 2023
Massachusetts Institute of Technology
Publié sous Creative Commons
Attribution 4.0 International
(CC PAR 4.0) Licence

La presse du MIT

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

Forward predictive coding:
The idea that motor plans lead to
predictions of sensory consequences,
facilitating error detection and
correction in motor control.

Internal (speech) error correction:
The correction of a speech error that
is committed prior to speech output.

Functional magnetic resonance
imaging (IRMf):
Neuroimaging method used to detect
changes in blood flow and oxygen in
the brain.

Lexicality effect:
Stronger neural response when
recitation is biased to increase the
potential for errors to be nonwords
than words.

Posterior middle temporal gyrus
(pMTG):
Brain region crucial for lexical
traitement.

corrected internally, prior to ever being spoken (Hickok, 2012; Levelt, 1983; Nozari et al.,
2011). The latter possibility is consistent with the idea from the motor control literature that
the brain simulates the position and trajectory of the motor effector it is controlling, a so-called
forward internal model, as a mechanism for improving the speed and accuracy of movements
via predictive coding (Kawato, 1999; Shadmehr & Krakauer, 2008; Wolpert et al., 1995). Inter-
nal predictive coding may provide a mechanism to detect and correct speech errors prior to
producing them (Hickok, 2012).

Although talkers can detect and correct overtly produced speech errors, evidence for inter-
nal error correction in speech is limited to inferences based on the timing of error corrections.
Par exemple, overtly realized error corrections, such as “v-horizontal,” have been argued to
occur too quickly to be accomplished using overt feedback alone, which in turn suggests the
existence of at least an internal mechanism to detect, if not correct errors (Nooteboom, 2005;
Nozari et al., 2011). Further evidence for internal error detection comes from electrophysio-
logical measures, which have identified signals that predict speech errors prior to their vocal-
ization (Möller et al., 2007). Strong direct evidence for internal error correction is sparse,
cependant.

In one fMRI study, Okada et al. (2018) reported suggestive evidence for the existence of an
internal error correction mechanism. In their experiment, the authors presented participants
with tongue twister sequences that were designed to bias speech errors toward either words
(REEF LEECH ➔ LEAF REACH) or nonwords ( WREATH LEAGUE ➔ LEATH REEG), a paradigm
previously investigated by Oppenheim and Dell (2008). Behavioral research on slips of the
tongue has shown that nonword errors are more rare than real word errors, the “lexical bias”
effect (Baars et al., 1975; Dell, 1986; Levelt et al., 1999; Nooteboom, 2005). Several previous
behavioral studies have shown that the lexical bias effect holds even when subjects do not
phonate their speech and self-report their errors (Corley et al., 2011; Oppenheim & Dell,
2008, 2010). Although not the main focus of their study, a lexicality effect was reported by
Okada et al. (2018) in the left posterior middle temporal gyrus (pMTG), a region implicated
in lexical-level processes. C'est, Okada and colleagues found greater activation in MTG
when participants recited tongue twisters that were biased to produce nonword errors com-
pared to word-errors, even on trials in which participants responded accurately. This is par-
ticularly interesting because no speech errors were committed on these trials—the activation
differences observed in MTG reflect the potential for a word versus nonword error. Le
authors suggested that speech errors resulting in nonwords are more readily detectable and
therefore would be more likely to be internally corrected prior to speaking. The fact that they
found a clear effect of lexical status of the error bias even under conditions of accurate per-
formance demonstrates that the system detected the distinction internally, and this could be
the case only if in fact an internal error was committed and then corrected prior to accurate
output. Although this is a very interesting finding, the pMTG activation in their study did not
reach statistical significance corrected for multiple comparisons, and as the authors note, ce
work requires replication. Runnqvist et al. (2020) recently reported on a study that used a very
similar design to Okada et al. (2018), and while they reported evidence of internal error cor-
rection in the cerebellum, they did not detect an effect in the cerebral cortex.

The present research builds on this prior work by attempting a direct replication of Okada
et autres. (2018) along with an extension to potentially render the paradigm more sensitive to
detecting evidence for an internal error correction mechanism. The experiment we introduce
uses the same tongue twister stimulus list and the same tasks (silent articulation and imaging)
as previous fMRI work, but additionally includes a set of stimuli designed to increase error
salience and therefore error detection and correction, thereby increasing our chances of

Neurobiology of Language

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

observing evidence for such a process. Previous work has shown that tongue twisters with the
potential to induce taboo word slips (par exemple., FULL BUD BUCK FUSS) elicit significantly fewer
slips than neutral tongue twisters, suggesting a higher rate of internal error detection and cor-
rection (Motley et al., 1982). We hypothesized that tongue twisters biased toward taboo word
errors would elicit greater activation in pMTG lexical networks compared to non-taboo word
errors. The basis for this is that detecting and correcting word-level speech errors presumably
drives activity to areas critical for lexical processing, which implicates pMTG (par exemple., Indefrey,
2011; Lau & Namyst, 2019), and that prior work has found suggestive evidence for internal
error correction in pMTG using tongue twisters (Okada et al., 2018; a possible mechanism of
this effect is explored in the Discussion section).

In addition to mapping an error correction mechanism in the brain, we aim to replicate
Okada et al.’s (2018) finding that motor-driven forward predictive signals are present in audi-
tory cortex in a speech production task. In their study, Okada et al. (2018) had participants
silently recite a sequence of tongue twisters in an fMRI experiment. Two speech production
conditions were included, one in which speech was articulated without phonating (silent artic-
ulation) and one in which speech production was imagined without articulation (imagined).
Both conditions were matched for acoustic input (c'est à dire., no speech input). Previous behavioral
research has shown that these two tasks engage different levels of linguistic/motor planning.
Imagined speech engages lexical-level processes but not lower-level phonological processes
whereas silently articulating speech engages both levels of processing. Donc, engaging
motor-phonological processes should generate a forward prediction of the acoustic conse-
quences of the executed (silent) speech, whereas engaging lexical-level processes should
pas. As expected, a contrast of silently articulated speech compared to imagined speech
revealed activity in left inferior frontal gyrus (IFG) and premotor cortex, areas involved in
speech articulation. More interestingly, they found robust activity in bilateral auditory cortex
when motor articulators were engaged, but not when speech was imagined, and this activation
was present even though there was no external auditory stimulation. The authors suggest that
these activations reflect stronger forward predictions generated in the articulation condition
compared to the imagining condition (Levelt, 1989).

The main goal of the present research is to examine a speech production mechanism that
has been elusive thus far: neural evidence of internal error correction. To that end, we leverage
a tongue twister paradigm that has been previously used to generate suggestive evidence of
internal error correction during speech in the pMTG. Using the same stimulus set and design
as this prior work, we attempt to replicate evidence for internal error correction in a much
larger sample of participants. Critique, we also improve our odds of finding evidence of inter-
nal error correction by introducing additional stimuli designed to tax the error correction
mechanism: tongue twisters that elicit taboo word errors. We also attempt to replicate the for-
ward predictive signal effect reported by this previous work. Replication is particularly impor-
tant since the aforementioned study was the first fMRI experiment to show evidence of forward
predictive signals involving auditory cortex, a sensory region that plays an important role in
speech production.

MATERIALS AND METHODS

Participants

Forty participants (25 females) entre 18 et 40 years of age were recruited from the Uni-
versity of California, Irvine (UCI) community. Participants received monetary compensation
for their participation. The volunteers were right-handed, native English speakers with normal

Neurobiology of Language

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

Taboo effect:
Stronger neural response when
recitation is biased to increase the
potential for errors to be taboo words
than non-taboo words.

or corrected-to-normal vision, no known history of neurological disease, and no contraindica-
tions for MRI. Informed consent was obtained from each participant prior to participation in the
study in accordance with guidelines from the local ethics committee at UCI that approved this
étude. A handful of participants were excluded from analysis for excessive head motion (N = 2)
and unanticipated scanning termination for a variety of reasons (N = 4; c'est à dire., claustrophobia,
equipment malfunction, excessive tardiness to the point of being unable to collect more than
a single fMRI session), leaving a total of 34 participants to contribute to the results. Four par-
ticipants from the group were not able to complete all nine scanning sessions but completed
89% (N = 2; 8 sessions), 77% (N = 1; 7), et 66% (N = 1; 6) of the fMRI sessions.

The number of participants recruited for this study was anticipated based on a power anal-
ysis of pilot data collected for four participants. The power analysis was carried out over two
regions of interest (ROI) using the fMRIpower toolbox (Mumford & Nichols, 2008). Anatom-
ical ROIs for this analysis were selected from the Harvard–Oxford cortical atlas (Desikan et al.,
2006), and adjustment for Type 1 error was made by applying a Bonferroni correction based
on the number of ROIs being compared ( p < 0.005). The power analysis indicated that approximately 40 participants were necessary to achieve 80% power for detecting a forward prediction effect within Heschl’s gyrus (effect size of 1.0623; effect sizes expressed in standard deviation units, which is analogous to Cohen’s d ) and internal error correction effects in tem- porooccipital middle temporal gyrus (toMTG; effect size of 0.3815 for a lexicality effect and 0.3985 for taboo effect). We note that a smaller sample size of 34 was indicated to achieve 72% power for detecting these effects. See the Data Analysis section for more information about how these effects were measured. Temporooccipital middle temporal gyrus was used for the power analysis instead of pMTG because this anatomical area aligned more closely with the foci reported by the study we sought to replicate (Okada et al., 2018). Stimuli and Task Scanning took place at the Facility for Imaging and Brain Research at UCI. Participants were scanned while they recited a set of four words (e.g., lean reed reef leach) in sync with a visual metronome. Thirty-two sets of tongue twisters used in previous experiments were employed in the current study (Oppenheim & Dell, 2010). These tongue twisters are known to behaviorally elicit a lexical bias effect. Lexical bias refers to the tendency for word errors to create a real word instead of a nonword (e.g., target word is “reef” but slips to “leaf,” is more likely than if target word is “wreath” and slips to “leath” because leath is a nonword). These stimuli were designed so that if an error occurred on the third or fourth word of each sequence, the out- come would yield either a real word error (e.g., “leaf”) or a nonword error (e.g., “leath”). In addition to these tongue twisters, we included 32 taboo tongue twisters (see Table 1 for exam- ple of stimuli). In a behavioral pilot involving 28 participants (15 females), we found that these taboo tongue twisters were effective in eliciting speech errors (non-taboo error rate = 23%, taboo word error rate = 20%). The present study followed the experimental procedure outlined in Okada et al. (2018). On each trial, a tongue twister phrase was visually presented on screen for 3 s, and then subjects were cued to silently articulate the sequence or imagine saying the sequence without mouth movements (see Figure 1). The presented cue was a cartoon face that remained on screen for 500 ms and contained a red arrow pointing either to the head or to the lips. An arrow pointing to the head cued the participants to imagine saying the word, and an arrow pointing to the lips cued the participants to silently articulate the words. A red fixation appeared on screen 500 ms after cue offset and served as the visual metronome, flashing at a rate of 2/s. Participants Neurobiology of Language 84 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u n o / l / l a r t i c e - p d f / / / / / 4 1 8 1 2 0 7 4 4 9 4 n o _ a _ 0 0 0 8 8 p d . l f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Predictive coding and error correction Table 1. Stimulus set Non-taboo tongue twister nod mod mock knob mine bikes bit mice bike wild wise bile bane gave gan bait name make mail nag wing bib bit whip six finch fill sin jail cheek cheap jean lean reed reef leech yore wan wok yawn gun bulb buck gull singe fib fish sip sing hitch his sick zinc niece need zest jog mod mock job job rob rock jot van match mat verve lull nudge buck love zing bib bit zip hinge fib fit hip sing that them zed daft gab gas dam than bunk nuzz there chicks fich fizz chin king hitch his kcik nab match mat nerve gun bulb but gull pen bunk nus pair zing that then zed gore wan watt gone goon nab nap gar nun bulb but null Neurobiology of Language Taboo tongue twister cod mod mock cob dine bite bike dice dial bile bike dies shave bane bit shank fake name nag fail shine bib bit ship ding sin sick dill dale chip chick dean queen reed reef queer core mud mum caught fun bulb buck full nib singe sip knit ditch sing sick diss pink niece need pest cot mod mock cob call rob rock cot shack match mat shave full bud buck fun ting bib bit tip shin fib fit ship shing them that shed dab laugh lamb dan jan bid bizz jar jicks finch fizz gin ping hitch hiss pick shab volt vat shot con grub grunt cup fend bus buck fair shang then that shed whole gone gore hat food tab tuck far pun mull miss pull 85 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u n o / l / l a r t i c e - p d f / / / / / 4 1 8 1 2 0 7 4 4 9 4 n o _ a _ 0 0 0 8 8 p d . l f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Predictive coding and error correction Example of a single trial. Participants were presented with a tongue twister sequence, Figure 1. which remained on screen for 3 s, followed by a cue to either articulate the sequence or imagine the sequence. They recited each word in sync with the visual metronome. recited one word per fixation in sync with the metronome. The interstimulus interval was 500 ms. After recitation, participants indicated with a button press if they were correct or incor- rect on the sequence using their left hand. Recall failure was treated as an incorrect response to ensure that only trials where rehearsal was successful (i.e., error free) would be analyzed. Participants were also instructed to indicate an incorrect response if overt production occurred accidentally. Prior to scanning, participants spent roughly 5 min (more when necessary) prac- ticing silently reciting words in a way that minimized but did not eliminate articulatory move- ments. Continuous feedback was provided by lab staff during this practice period. A single trial in the experiment was 8 s in length and there were approximately 42 trials in each session. Each session consisted of an equal number of tongue twister phrases biased to produce word errors or nonword errors and taboo errors. There were eight experimental ses- sions and each session consisted of approximately 14 trials of each type, which were ran- domly presented along with six rest trials (fixation). The study started with a high-resolution structural scan. This was followed by a short practice session of the experiment using approx- imately 10 trials to further familiarize subjects with the task. Scanning was conducted during the practice session to acclimatize subjects to the fMRI environment, as well as to monitor head movement and provide feedback prior to the start of the experiment. Participants also received feedback about head movement in between sessions based on qualitative assessment of the images that had been collected. The study lasted approximately 1.5 hr. Stimulus pre- sentation and timing was controlled using PsychToolbox (Borgo et al., 2012) implemented in MATLAB (Mathworks, 2022). Imaging Imaging data were collected on a 3T Siemens Prisma scanner (Siemens Medical Solutions, 2022) equipped with a 32-channel RF receiver head coil. A single T1-weighted MPRAGE sequence was acquired (matrix = 256 × 256 mm, TR = 2.3 s, TE = 2.32 ms, flip angle = 8°, size = 0.937 × 0.937 × 0.9 mm). An echo-planar imaging (EPI) pulse sequence was collected for each of the eight experimental sessions and the practice session (matrix = 100 × 100 mm, TR = 2 s, TE = 35 ms, flip angle = 90°, size = 2.4 × 2.4 × 2.4 mm, 56 slices). Data Analysis Data were preprocessed and analyzed using the FMRIB’s Software Library (FSL; Jenkinson et al., 2012). First, rigid-body motion correction was performed with FSL’s intramodal motion Neurobiology of Language 86 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u n o / l / l a r t i c e - p d f / / / / / 4 1 8 1 2 0 7 4 4 9 4 n o _ a _ 0 0 0 8 8 p d . l f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Predictive coding and error correction correction tool (MCFLIRT) using the normalized correlation cost function and the middle vol- ume as the initial template (Grabner et al., 2006). Participants with excessive head movement (>0.3 mean framewise displacement, as defined by Power et al., 2012, across sessions) étaient
eliminated from further analysis. Mean framewise displacement among the remaining partic-
ipants (N = 34) was relatively low (M = 0.15, SD = 0.07). Echo planar images were high pass
filtered by calculating the minimal period that retains 90% of the variance in the design matrix
regressors (this amounts to a roughly 0.01 hz cutoff ). These images were then spatially
smoothed using an isotropic 8-mm full width half maximum (FWHM) gaussian, and the ana-
tomical image for each subject was coregistered to their middle EPI volume. Data analysis was
performed with FSL’s fMRI Expert Analysis Tool (Jenkinson et al., 2012) and proceeded in three
steps: (je) modeling within-session parameter estimates for events of interest using fixed effects,
(ii) using these parameter estimates in a between-session analysis to model participant mean
response, et (iii) using mean participant response in a between-subjects analysis to model
group response using mixed effects (FSL’s FLAME1+2). Parameter estimates for each partici-
pant were transformed into standardized space using the MNI152 template (Jenkinson et al.,
2002). All trials on which participants indicated making speech errors were excluded from
analyse. En moyenne, participants reported errors on 14% of trials (SD = 8%). Error rates were
comparable across different types of tongue twister (biased toward word errors: M = 15.3%,
SD = 10%; biased toward nonword errors: M = 15.6%, SD = 9.1%; biased toward taboo
errors: M = 12.4%, SD = 7.8%).

Regressors for events of interest were created by convolving the predictor variables repre-
senting the time course of stimulus presentation with a gamma variate function. As in the
previous experiment, regressors modeled the following experimental trial types: “Articulation:
Nonword Errors,” “Articulation: Word Errors,” “Articulation: Taboo Errors,” “Imagining: Non-
word Errors,” “Imagining: Word Errors,” “Imagining: Taboo Errors.” All trials on which partic-
ipants reported making an incorrect response, the visual presentation of words on all trials,
and the six motion parameter estimates determined during the realignment stage of prepro-
cessing were included in the model as nuisance regressors. Regressors were used to generate
parameter estimates for each condition of speech task (c'est à dire., imagined and silently articulated)
and error type (c'est à dire., tongue twisters biased toward: nonword, word, and taboo word errors).
Parameter estimates for each speech task modeled tongue twisters of all error types and
parameter estimates for each error type modeled both silently articulated and imagined
tongue twisters. Multiple contrasts were set up using these parameter estimates. As the pri-
mary goal of the present research was to seek evidence for internal error correction, we first
tested for the lexicality effect described by Okada et al. (2018), followed by a taboo effect
based on the new stimuli introduced in this study, and finally the forward prediction effect
also described by Okada et al. (2018). Identical to prior work, the lexicality effect was based
on the contrast between parameter estimates for the nonword and word tongue twister con-
ditions. The same logic was extended to the novel taboo stimuli and the taboo effect was
based on the contrast between taboo and word tongue twister conditions. The forward pre-
diction effect was based on the contrast between the two speech task conditions (c'est à dire., silently
articulated vs. imagined). For each of these three contrasts an additional analysis was per-
formed contrasting each condition in the pair against baseline fixation. The contrasts between
each condition and baseline fixation were used to mask the contrast maps, allowing us to
distinguish brain areas that show a significant difference in BOLD response between condi-
tions but overall show below-baseline response.

Consistent with previous research, we expected to find taboo and lexicality effects in
pMTG, and a forward predictive signal effect in auditory cortex. These hypotheses target

Neurobiology of Language

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

specific regions and as such we complemented whole-brain analyses with a ROI approach
that compared mean condition-level parameter estimates used in the whole brain contrast
analyses within predefined anatomical areas. The same ROI analysis was carried out across
and within participants, the latter of which allowed us to characterize the consistency of the
effects we sought. Although the ROI analyses directly tested our hypotheses and provided
complementary information to the whole-brain contrast analysis approach, we elected to place
relatively greater focus on the whole-brain analysis as it allowed us to more comprehensively
test which brain areas are involved in the effects we were interested in (c'est à dire., including those
areas not implicated in prior work), as well as to determine more precisely where effects within
anatomical ROIs were observed. Left and right hemisphere anatomical ROIs were extracted
from the Harvard–Oxford atlas. The anatomically defined pMTG area was used to test for
the lexicality and taboo effects, and Heschl’s gyrus was used to test for the forward prediction
effect. The ROI from the power analysis was substituted based on the results of the whole brain
analyses. To foreshadow our findings, these analyses revealed a taboo effect consistent with
the location of the effect reported by Okada et al. (2018), but more clearly concentrated within
anatomical pMTG than toMTG. Group effects within ROIs were tested with paired t tests car-
ried out across participants. Tests were additionally carried out across voxels for each partic-
ipant to characterize the consistency of the effects. Minimum significance in the ROI analyses
was based on a Bonferroni corrected p-value threshold of 0.05, and minimum significance for
whole brain voxelwise testing was based on a voxelwise p-value threshold of 0.01 (c'est à dire., cluster-
forming Z-threshold of 2.6) and a GRF-based cluster size p-value threshold of 0.05.

In addition to the core analyses detailed above, we present post hoc analyses that give fur-
ther context for some of our findings. We emphasize that these secondary analyses are inci-
dental by separating them from the main results section. The aim of the post hoc analyses is to
better functionally characterize the network of regions associated with the internal error cor-
rection effects (c'est à dire., lexicality and taboo effects), particularly the wider network capturing
regions where neural response increased during the nonword or taboo conditions relative to
the word condition but overall remained below-baseline. Functions associated with this net-
work and with each of its constituent regions were probed using the Neurosynth meta-analytic
database, which associates word frequencies in the abstracts of studies with their activation
foci, enabling meta-analysis of groups of studies that frequently use particular terms (Yarkoni
et coll., 2011). Network-level decoding was performed by computing the Pearson correlation
coefficient between contrast activation maps from the current study and each of the meta-
analyses that were generated for the 3,228 terms frequently used in the neuroimaging literature
and embedded in Neurosynth. Briefly, performing a meta-analysis for each term involved sep-
arating all studies in Neurosynth into two groups: those that used a particular term frequently
(minimum rate of 1/1,000 words which has been shown to control for incidental word usage;
Yarkoni et al., 2011) and those that did not. Suivant, a search was performed for voxels where
activity was more consistently reported in the set of studies that do frequently use the term
relative to those that do not. This was accomplished by extracting the activation tables from
these two groups of studies, creating contingency tables at each voxel that described whether
activity was present and whether a phrase was used, and then performing a chi-square test.
Due to their smaller size, regions were decoded in a slightly different way—by computing the
mean posterior probability that a phrase was used within a study if activity was observed in
each of a regions’ voxels. Posterior probability estimates assumed a uniform prior (c'est à dire., tous
terms are equally likely to appear) and were generated for each term in Neurosynth. More
complete details about how posterior probability was computed can be found in prior work
(Yarkoni et al., 2011).

Neurobiology of Language

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

RÉSULTATS

Core Findings

Internal error correction effects

Our first analysis aimed to replicate the lexicality effect described in previous work (Okada
et coll., 2018). Critique, all analyses we present were restricted to error-free trials. A group-level
contrast between the nonword and word conditions of the experiment revealed no significant
differences in brain response at our minimal significance thresholds (Z > 2.6 or p < 0.01; clus- ter corrected at p < 0.05). We tested for effects on the cusp of significance by avoiding cluster correction but did not find any significant differences in brain response. Lowering the cluster- forming threshold further (Z > 2.3) did not reveal any areas where response to nonwords was
greater than response to words. En plus, we tested whether the mean parameter estimates
within the pMTG ROI differed between these two conditions but found no significant differ-
ence across participants in either the left hemisphere (parameter estimates for words: M = 9.59,
SD = 41.84; parameter estimates for nonwords: M = 8.37, SD = 39.4; t(33) = 0.56, p = 0.58) ou
the right hemisphere (parameter estimates for words: M = −4.3, SD = 32.18; parameter esti-
mates for nonwords: M = −7.5, SD = 28.26; t(33) = −1.35, p = 0.19). Because we failed to
replicate the lexicality effect reported in previous work (c'est à dire., no effect was observed at the
group-level), we did not investigate how consistently the effect appeared within participants.

We next evaluated whether the novel taboo stimuli generated evidence for internal error
correction by contrasting the whole-brain parameter estimates for the taboo and word condi-
tion. As we anticipated, the taboo condition appears to have successfully increased the load
on internal error detection and correction. Although no areas of the brain showed significantly
greater response for the word condition than the taboo condition, relatively higher response for
the taboo condition was found in a wide network that included the pMTG bilaterally (Z > 2.6
or p < 0.01, cluster corrected at p < 0.05 with a minimum significant cluster size of 803 voxels; see Figure 2A). A more comprehensive description of brain regions in this network was pro- vided by registering the contrast map to anatomical areas of the Harvard–Oxford cortical atlas (Table 3). Incidental overlap between the contrast map and anatomical areas as a result of activity spilling over an anatomical boundary in a way that is inappreciable was deempha- sized by focusing only on those areas in which more than 5% of voxels showed a significant difference between conditions. For visualization, both the atlas and the contrast map were projected onto the fsaverage inflated cortical surface using a recently developed procedure that implements registration fusion with advanced normalization tools to improve projection accuracy (Wu et al., 2018; Figure 2B). The network of regions that responded more strongly to the taboo than the word condition spanned the bilateral frontal poles (FPs), bilateral frontal medial cortex (FMC), bilateral superior frontal gyrus (SFG), right anterior cingulate gyrus (aCG), bilateral posterior cingulate gyrus (pCG), bilateral paracingulate gyrus (paraCG), right subcallosal cortex, bilateral precuneous cortex (preCC), bilateral anterior middle temporal gyrus (aMTG), bilateral pMTG, left posterior inferior temporal gyrus (pITG), bilateral angular gyrus (AG), and bilateral superior lateral occipital cortex (sLOC; Figure 2). The regions that responded more strongly to the taboo than the word condition were then evaluated based on whether they showed above-baseline response during the taboo condi- tion. Our first approach was to mask the contrast map between taboo and word conditions by the contrast map for the taboo condition (i.e., taboo > word AND taboo > baseline; les deux
contrasts set to Z > 2.6 or p < 0.01, cluster corrected at p < 0.05). The resulting map produced a single large cluster and a handful of implausibly small and less meaningful clusters. A 20- voxel cluster-extent threshold was applied to the resulting map to emphasize the largest and Neurobiology of Language 89 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u n o / l / l a r t i c e - p d f / / / / / 4 1 8 1 2 0 7 4 4 9 4 n o _ a _ 0 0 0 8 8 p d . l f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Predictive coding and error correction Figure 2. Group contrast between taboo and (neutral) word conditions. (A) select slices from the volume-based group activation map. Sagittal slices cut through peak activation observed in anterior middle temporal gyrus (x = ±64) and three separate peaks in posterior middle temporal gyrus (x = ±60, ±56). They also show the emergence of activity in superior lateral occipital cortex (x = ±56) and its splitting into more anterior and posterior foci, with the anterior activity intruding into angular gyrus (x = ±42), as well as more medial activations (x = ±10; i.e., posterior cingulate gyrus and frontal medial cortex). Axial slices show the same patterns: peak of activation in anterior middle temporal gyrus (z = −26), two peaks of activation in posterior middle temporal gyrus (z = −14), a third peak of activation in posterior middle temporal gyrus (z = −4), and two peaks of activation in superior lateral occipital cortex (z = 22, 44). (B) Volume-based results projected onto the fsaverage surface along with the Harvard–Oxford cortical atlas to better visualize overlap between activations and anatomical regions. Bilateral anatomical regions are shown as colored outlines and any region overlapping with activation is highlighted by an opaque superimposed number in one hemisphere that corresponds to the region’s index within the atlas. The labels for these indices are provided in Figure 4. l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u n o / l / l a r t i c e - p d f / / / / / 4 1 8 1 2 0 7 4 4 9 4 n o _ a _ 0 0 0 8 8 p d . l f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 most readily interpretable cluster in our visualization of this result in Figure 3. We emphasize that the individual maps submitted to the masking procedure were all cluster corrected them- selves and the exceedingly small clusters in the masked map may reflect uninteresting differ- ences between contrasts (e.g., noise caused by co-registration). In general, this analysis revealed that the left pMTG was by far the largest area to show significantly greater response for taboo than word conditions while also responding significantly above-baseline to the taboo condition (see Figure 3). The cluster of activity that centered on left pMTG crossed only super- ficially over the boundary between this area and posterior superior temporal gyrus (STG), toMTG and temporooccipital inferior temporal gyrus (ITG). We found additional clusters that peaked in pMTG below the cluster-extent threshold but note that we also found clusters below this threshold which peaked in right AG and left sLOC (Table 2). In an additional analysis, we averaged the parameter estimates for the taboo and word con- ditions (i.e., taboo > baseline, word > baseline) within the portion of each anatomical area that
exhibited a significant difference between these two conditions (c'est à dire., the contrast map from
Chiffre 2). This areal analysis was used to estimate taboo effect size and provided an additional
glimpse into which areas differentially responded to the taboo and word conditions while
simultaneously showing a stronger preference for baseline fixation than the task. Dans l'ensemble, ce
analysis highlighted the same group of areas as having both greater response to the taboo than

Neurobiology of Language

Predictive coding and error correction

Chiffre 3. Group contrast between taboo and (neutral) word conditions: masked by significant
activity during the taboo condition (c'est à dire., taboo > baseline) et 20 voxel cluster-extent threshold
is applied, revealing a large area in pMTG.

Tableau 2.

Clusters from the taboo versus word conditions contrast masked by taboo > baseline contrast

Cluster
size
(2 mm
voxels)
140

Cohen’s
d
1.78

Peak
z-
valeur
3.88

Peak x-
coordinate
−60

Peak y-
coordinate
−38

Peak z-
coordinate
−8

Anatomical atlas areas overlapping with
cluster (% of anatomical area covered)
Left middle temporal gyrus, posterior

division (7%); left superior temporal
gyrus (2%); left middle temporal gyrus,
temporooccipital part (2%); left inferior
temporal gyrus, temporooccipital part
(<1%) Right angular gyrus (>1%); right lateral

cortex occipital, superior division (>1%)

Left lateral occipital cortex, superior

division (>1%); left angular gyrus (>1%)

2.13

3.06

3.56

3.67

−46

−56

−60

Left lateral occipital cortex, superior

0.81

3.5

−34

−70

division (>1%)

Left middle temporal gyrus, posterior

division (>1%); left inferior temporal
gyrus, posterior division (>1%)

3.9

2.84

−54

−30

−17

Anatomical
region
chevauchement
with peak

Left middle
temporal
gyrus,
posterior
division

Right angular

gyrus

Left lateral
occipital
cortex,
superior
division

Left middle
temporal
gyrus,
posterior
division

Neurobiology of Language

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Chiffre 4. Group contrast between taboo and (neutral) word conditions: parameter estimates, or beta values for the individual conditions
being contrasted, presented as an average within each anatomical region. Only voxels showing a significant effect in the contrast were
included in the averages. Averages are presented separately for left hemisphere (navy and violet colored bars) and right hemisphere regions
(salmon and yellow colored bars). Error bars represent standard error of the mean. Regions are organized on the x-axis based on descending
mean difference between taboo and word parameter estimates. The magnitude of this difference is shown as a gray line that corresponds to the
secondary y-axis.

word condition and above-baseline response to the taboo condition: bilateral pMTG, left AG,
left pITG, and left aMTG (voir la figure 4). Plus loin, we report medium to large effect sizes (c'est à dire.,
taboo vs. words) in all areas, with the pITG showing the largest effect size in the left hemi-
sphère, and preCC showing the largest effect size in the right hemisphere (Tableau 3). Areas that
additionally responded above baseline to the taboo condition all showed large effect sizes. Non
hemispheric differences were found for effect size (gauche: M = 1.17, SD = 0.6; droite: M = 0.88,
SD = 0.33; t(9) = 1.04, p = 0.32) or the spatial extent of activity within areas (gauche: M = 22%,
SD = 16.18%; droite: M = 19.77%, SD = 16.13%; t(9) = 1.22, p = 0.25), but activation maxima
inside areas was typically higher in the left hemisphere (gauche: M = 3.93, SD = 0.37; droite: M =
3.6, SD = 0.26; t(9) = 3.96, p < 0.01). Finally, we report that only a single area tended to show above-baseline response to both taboo and word conditions: the left pMTG (Figure 4). This area overlapped exclusively with the largest cluster identified when masking the taboo versus words contrast by the taboo > baseline contrast (cf. Figure 2B and Figure 3).

In the preceding whole brain analyses we showed evidence for internal error correction
(c'est à dire., taboo effect) within a portion of the pMTG that our hypotheses targeted. We next carried
out a more specific test for the taboo effect across all voxels of anatomically defined pMTG.
Mean parameter estimates for the taboo and word conditions were extracted from pMTG
voxels for each participant and a paired t test between conditions was performed over

Neurobiology of Language

Predictive coding and error correction

Tableau 3.

Areal effect sizes for taboo versus word condition contrast

ROI size Cohen’s d

Atlas ROI
(1) Frontal pole

(3) Gyrus frontal supérieur

(11) Middle temporal gyrus, anterior division

6997

3347

512

(12) Middle temporal gyrus, posterior division

1400

(15) Inferior temporal gyrus, posterior division

1175

(21) Angular gyrus

(22) Lateral occipital cortex, superior division

(25) Frontal medial cortex

(27) Subcallosal cortex

(28) Paracingulate gyrus

(29) Cingulate gyrus, anterior division

(30) Cingulate gyrus, posterior division

(31) Precuneous cortex

1197

5251

561

782

1720

1385

1332

3047

Atlas ROI

(1) Frontal pole

(3) Gyrus frontal supérieur

(11) Middle temporal gyrus, anterior division

8195

3005

472

(12) Middle temporal gyrus, posterior division

1374

(15) Inferior temporal gyrus, posterior division

1056

(21) Angular gyrus

(22) Lateral occipital cortex, superior division

(25) Frontal medial cortex

(27) Subcallosal cortex

(28) Paracingulate gyrus

(29) Cingulate gyrus, anterior division

(30) Cingulate gyrus, posterior division

(31) Precuneous cortex

1658

5127

594

749

1650

1526

1393

3141

Left hemisphere

Peak z-value
4.2

Peak x
−12

Peak y
66

3.68

3.99

3.94

3.3

4.25

4.41

3.44

–

3.73

–

4.36

−10

−60

−50

−36

−6

–

−10

–

−10

−6

−26

−32

−56

−68

–

−50

−54

Peak z % ROI active

−24

−8

−16

−10

–

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

0.6

0.61

1.62

0.83

2.38

1.53

0.78

0.6

–

0.71

–

1.6

1.62

ROI size Cohen’s d

Peak z-value

Peak x

Peak y

Peak z % ROI active

Right hemisphere

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

0.53

0.61

0.79

1.49

–

1.01

0.75

0.63

0.7

0.71

0.79

1.12

1.53

3.98

3.6

3.45

3.88

–

3.69

3.91

3.36

3.15

3.59

3.26

3.55

3.74

–

−2

−28

–

−54

−68

−52

−60

−26

−4

–

−20

−4

participants (Figure 5A). The taboo condition elicited higher parameter estimates than the
word condition in the left pMTG (taboo: M = 18.53, SD = 36.77; word: M = 9.59, SD =
41.84; t(33) = 3.12, Bonferroni corrected p < 0.05). The same effect was on the cusp of significance in the right hemisphere (taboo: M = 2.84, SD = 25.66; word: M = −4.23, SD = Neurobiology of Language 93 Predictive coding and error correction Figure 5. Taboo effect in posterior middle temporal gyrus (pMTG) within-participants. Anatomical ROI-based analysis of the internal error correction effect in pMTG. (A) Parameter estimates, or beta values for taboo and word conditions are averaged across all voxels of left pMTG (left) and right pMTG (right) in each subject. The shaded portions of the line plot for each subject represents standard error of the mean. Results of a paired t test across participants is presented below, showing significantly higher response in left pMTG to the taboo condition, and a difference between the two conditions that is on the cusp of significance for right pMTG. (B) T values showing magnitude and significance ( p < 0.001; Bonferroni corrected) of parameter estimate differences between the two conditions are shown for each subject as a bar plot. Note the asterisk marks insignificant results. Each subject is assigned a color that is consistent between panels. Bars are presented in pairs such that the first and darker shaded bar of any pair represents the result of the t test performed on the left hemisphere ROI, and the second and lighter shaded bar represents the result of the t test performed over the right hemisphere ROI. Overall, 26/34 participants show a significant effect in at least one hemisphere (22 show an effect in both hemispheres). l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u n o / l / l a r t i c e - p d f / / / / / 4 1 8 1 2 0 7 4 4 9 4 n o _ a _ 0 0 0 8 8 p d . l f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 32.61; t(33) = 2.58, Bonferroni corrected p = 0.05). In addition, we ensured that the left hemisphere internal error correction effect was present independently during imagined (taboo: M = 10.24, SD = 19.49; word: M = 4.75, SD = 24.53; t(33) = 2.05, Bonferroni cor- rected p < 0.05) and silently articulated tongue twister trials (taboo: M = 2.84, SD = 25.66; word: M = −4.23, SD = 32.61; t(33) = 2.99, Bonferroni corrected p < 0.05). Finally, we characterized the consistency of the taboo effect within anatomical pMTG by applying paired t tests between the taboo and word conditions within participants (Figure 5B). This analysis yielded a significant taboo effect in roughly 76% of participants (N = 26/34; Bonferroni-corrected p < 0.001). The effect occurred most often bilaterally (N = 22/26) and showed no hemispheric preference when it occasionally occurred in a single hemisphere (left: N = 2/26; right: N = 2/26). In four participants the effect was present, but response in pMTG was higher during baseline fixation than during the taboo condition. That is, response in pMTG was greater for the taboo than word conditions but highest for baseline fixation. In all cases where the effect was not present, we saw a significant reverse effect (N = 9). We considered whether portions of pMTG still showed a taboo effect in those participants who did not exhibit a mean effect across the entire pMTG by inspecting significant voxel-level differences between taboo and word conditions. No participant who showed a nonsignificant mean taboo effect presented with significant differences between taboo and word condition within portions of pMTG. Neurobiology of Language 94 Predictive coding and error correction Forward prediction effect One aim of the current study was to replicate the forward predictive signal effect described in prior work (Okada et al., 2018) by identifying brain regions that respond more strongly during silently articulated than imagined tongue twisters. To that end, we confirmed that silently artic- ulated speech activates portions of auditory cortex more strongly than imagined speech, even though both conditions lack auditory input and do not involve overt production ( p < 0.01 clus- ter corrected at p < 0.05; see Figure 6A–B). Overall, silently articulated tongue twisters pro- duced greater activity than imagined tongue twisters in and around Heschl’s gyrus (HG), but also in a broad network of speech-related regions that span STG, planum temporale (PT), pMTG, toMTG, precentral gyrus, postcentral gyrus (postCG), insula, IFG, aCG, and the cere- bellum (Figure 6A–B). We also found that silently articulated tongue twisters yielded greater activity in other brain regions, most of which have also been observed to activate during speech processing (Figure 6A–B). These regions were found in temporal cortex (i.e., bilateral planum polare, bilateral temporal pole), inferior temporal and neighboring portions of occipital cortex (i.e., left temporooccipital ITG, bilateral lingual gyrus, bilateral temporal occipital fusiform cor- tex, bilateral occipital fusiform gyrus), parietal cortex (i.e., bilateral parietal operculum, bilateral anterior and posterior supramarginal gyrus, bilateral superior parietal lobule), and frontal cortex (i.e., bilateral supplementary motor cortex, bilateral middle frontal gyrus, bilateral frontal and central operculum, bilateral SFG, left frontal orbital cortex; Figure 6A–B). One point of difference between our results here and prior work (Okada et al., 2018) is that we also report regions that express greater activity for imagined than silently articulated word lists in several areas including bilateral anterior parahippocampal gyrus (paraHG) and l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u n o / l / l a r t i c e - p d f / / / / / 4 1 8 1 2 0 7 4 4 9 4 n o _ a _ 0 0 0 8 8 p d . l f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Figure 6. Group contrast between silently articulated and imagined speech. (A) Select slices from the volume-based group activation map. Sagittal slices cut through peak activation for the contrast in precentral and postcentral gyri (x = ±58) and follow Heschl’s gyrus (x = ±50, ±44, and ±38). Axial slices show the most ventral activations in the contrast, including the cerebellum and neighboring mesial structures (e.g., anterior paraHG and brainstem; z = −26), more dorsal activations in inferior temporal cortex (z = −14), activations along middle and superior temporal cortex (including Heschl’s gyrus; z = 8 and 14), and the most posterior activation in the contrast within posterior supramarginal gyrus (z = 40). (B) Volume-based results projected onto the fsaverage surface along with the Harvard–Oxford cortical atlas to better visualize overlap between activations and anatomical regions. Bilateral anatomical regions are shown as colored outlines and any region overlapping with activation is highlighted by an opaque superimposed number in one hemisphere that corresponds to the region’s index within the atlas. The labels for these indices are provided in Figure 7. (C) Masked by significant activity during either condition (i.e., silent articulation > base-
line OR imagined > baseline), revealing that areas which showed relatively stronger response to imagined speech were deactivated during the
imagined speech task.

Neurobiology of Language

Predictive coding and error correction

posterior paraHG, bilateral posterior temporal fusiform cortex, bilateral AG, bilateral sLOC,
bilateral FP, bilateral FMC, bilateral subcallosal cortex, bilateral paraCG, right preCC, and right
occipital pole (OccP; Figure 6A–B).

Just as for the taboo effect, the contrast between silently articulated and imagined tongue
twisters was masked by areas that showed significant above-baseline response to each of these
two conditions, revealing below-baseline response almost exclusively in areas that produced
significantly higher response to imagined than silently articulated tongue twisters (Figure 6C).
En effet, all areas that showed greater response to imagined tongue twisters also showed
below-baseline response during the imagined condition, while only a few small areas in infe-
rior temporal and occipital cortex that showed greater response to silently articulated tongue
twisters exhibited below-baseline response during the silent articulation condition (cf.
Figure 6B with 6C). The latter areas included bilateral lingual gyrus, bilateral occipital fusiform
gyrus, and right OccP (cf. Figure 6B with 6C).

Areal parameter estimates for silently articulated and imagined tongue twister conditions
confirmed that areas with stronger response to the silently articulated condition all showed
above-baseline response (Chiffre 7). Cependant, this analysis also revealed areas with stronger
response to the imagined condition that showed above-baseline response, mainly left AG, gauche
FP, and left anterior paraHG (Chiffre 7). Notably, parameter estimates for these three areas were
relatively low for the imagined condition. The contrast between silently articulated and imag-
ined conditions that was masked by above-baseline imagined response did not pick out voxels
in these areas due to the more stringent statistical significance criteria in that analysis.

Areal effect sizes for the contrast between silently articulated and imagined conditions were
large for left anterior paraHG, but small for left AG, and negligible but significant for the left FP
(Tableau 4). For other areas which showed greater response during the imagined than silently

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Chiffre 7. Group contrast between silently articulated and imagined speech: parameter estimates, or beta values for the individual conditions
being contrasted, presented as an average within each anatomical region. Only voxels showing a significant effect in the contrast were
included in the averages. Averages are presented separately for left hemisphere (blue and cyan colored bars) and right hemisphere regions
(red and yellow bars). Error bars represent standard error of the mean. Regions are organized on the x-axis based on descending mean
difference between articulated and imagined betas. The magnitude of this difference is shown as a gray line that corresponds to the sec-
ondary y-axis.

Neurobiology of Language

ROI
size

6997

1463

3347

Cohen’s
d
−0.1

0.95

0.36

Peak
z-value

4.82

5.28

5.3

3280

0.52

4.97

836

1.12

5.33

Peak
x-coord
−44

−36

−10

−48

915

0.91

6.02

−62

4981

2465

340

1.22

0.72

1.37

6.59

5.56

5.17

−62

−56

−66

−8

(10) Superior

1019

1.93

4.72

Tableau 4.

Areal effect sizes for articulated versus imagined word lists

Left hemisphere

Right hemisphere

Peak
y-coord

Peak
z-coord

Peak
z-value

Peak
x-coord

Peak
y-coord

Peak
z-coord

% ROI
active

ROI
size

8195

1465

3005

Cohen’s
d
−0.46

1.09

0.79

4.36

5.37

4.79

3141

0.65

5.65

745

0.56

4.4

−6

847

0.68

5.19

−2

4822

2490

348

1.45

0.83

1.47

7.01

4.68

5.02

1155

2.2

5.84

−6

−28

1400

1.39

4.06

−54

−40

−2

1374

2.02

4.67

−20

−4

1019

1.71

3.96

−54

−42

1287

1.93

5.28

−48

808

1.35

4.83

−52

−58

−12

907

–

(17) Postcentral gyrus

4046

1.55

(18) Superior

parietal lobule

1731

0.9

6.38

4.72

−60

−44

−4

−38

3627

1669

2.22

0.93

6.46

3.58

−6

−50

Atlas ROI

(1) Frontal pole

(2) Insular cortex

(3) Superior

frontal gyrus

(4) Middle

frontal gyrus

(5) Inférieur

frontal gyrus,
pars triangularis

(6) Inférieur

frontal gyrus,
pars opercularis

(7) Precentral gyrus

(8) Temporal Pole

(9) Superior

temporal gyrus,
anterior division

temporal gyrus,
posterior division

(12) Middle

temporal gyrus,
posterior division

(13) Middle

temporal gyrus,
temporooccipital
part

(16) Inférieur

temporal gyrus,
temporooccipital
part

N
e
toi
r
o
b
o
o
g
oui

L
un
n
g
toi
un
g
e

9
7

P.
r
e
d
je
c
t
je
v
e

c
o
d
je
n
g

un
n
d

e
r
r
o
r

c
o
r
r
e
c
t
je
o
n

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

<1 35 7 N e u r o b o o g y i l o f L a n g u a g e Atlas ROI (19) Supramarginal gyrus, anterior division (20) posterior division (21) Angular gyrus (22) Lateral occipital cortex, superior division Table 4. (continued ) Left hemisphere Right hemisphere ROI size 1128 Cohen’s d Peak z-value 0.77 4.93 Peak x-coord −66 Peak y-coord −30 Peak z-coord % ROI active ROI size Cohen’s d Peak z-value Peak x-coord 26 82 959 1.64 4.85 48 Peak y-coord −32 Peak z-coord % ROI active 54 45 1357 0.8 4.87 −52 −40 24 46 1516 1.04 4.91 58 −42 16 34 1197 5251 −0.35 −0.24 3.83 4.54 (25) Frontal medial 561 −0.93 −3.72 874 0.79 5.28 cortex (26) Juxtapositional lobule cortex (formerly s upplementary motor cortex) (27) Subcallosal cortex gyrus (29) Cingulate gyrus, anterior division (31) Precuneous cortex 782 −2.52 −4.53 −4 (28) Paracingulate 1720 0.01 4.9 1385 0.34 4.95 3047 – – – −38 −28 −2 −8 −12 −12 −50 −64 34 2 22 14 12 – 20 44 54 −14 64 −6 38 38 – −6 (33) orbital 1943 0.76 4.42 −50 cortex (34) Parahippocampal 709 −1.41 −4.19 −16 −10 −22 gyrus, anterior division (35) Parahippocampal gyrus, posterior division (36) Lingual gyrus (38) Temporal fusiform cortex, posterior division 9 8 475 −0.8 −4.51 −28 −32 −12 1934 1001 0.67 5.93 −0.69 −3.92 −18 −28 −62 −34 −12 −16 12 12 46 20 14 17 20 25 68 48 41 42 1658 5127 −0.29 −0.48 4.33 −4.24 594 −1.18 −3.93 834 0.9 4.76 749 −2.56 −4.45 1650 −0.01 4.41 1526 0.3 4.37 0 3141 1681 – – – – 42 28 2 2 −2 10 2 – – −46 −86 34 6 22 16 18 – – 52 34 −12 74 −4 40 38 – – 766 −1.59 −4.25 16 −10 −22 378 −1.07 −3.72 26 −32 −10 2037 0.43 5.35 812 −0.99 −3.53 14 34 −66 −20 −12 −18 20 27 42 57 44 38 40 <1 4 11 39 26 8 P r e d i c t i v e c o d i n g a n d e r r o r c o r r e c t i o n l D o w n o a d e d f r o m h t t p : > 4.2; Figure 8A–B). Critique, the resulting independent cluster within temporal
cortex was just posteroventral to HG, remaining in the vicinity of auditory cortex. At this
threshold, the frontal cluster peaking in left precentral gyrus extended into left supramarginal

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Chiffre 8. Group contrast between silently articulated and imagined speech. Minimal significance threshold from Figure 13 was increased by
increments of 0.1 until the frontal-temporal cluster splintered into multiple independent clusters. (UN) The results of the same contrast thresh-
olded at Z > 4.2 (cluster corr., p < 0.05) are shown in volume-space. (B) The same maps from Figure 6A projected onto the fsaverage surface. Neurobiology of Language 100 Predictive coding and error correction Forward prediction effects in Heschl’s gyrus within-participants. Atlas region of interest (ROI) based analysis of the forward pre- Figure 9. diction effect in Heschl’s gyrus is shown. (A) Beta values for silently articulated word lists and imagined word lists are averaged across all voxels of left and right Heschl’s gyri in each participant. The shaded portions of the line plot for each participant represents standard error of the mean. The bottom portion of the panel also reports the result of a t test for a group-level difference between conditions (i.e., paired t test across participants), which shows significantly greater response during silently articulated speech in Heschl’s gyri. (B) T values showing mag- nitude and significance ( p < 0.001; Bonferroni corrected) of beta differences between the two conditions are shown for each participant as a bar plot. Note the asterisk marks insignificant results. Each participant is assigned a color that is consistent between panels (as well as with Figure 5). Bars representing the results of t tests performed on both hemisphere ROIs are presented in pairs: left (darker shaded bar) and right (lighter shaded bar). Overall, 26/34 participants show a significant effect in at least one hemisphere (20 show an effect in both hemispheres). gyrus and posterior PT. The same cluster also covered left insular cortex, operculum, IFG (pars opercularis), and postCG. Activity from left frontal cortex spilled over into anterior temporal regions as well as anterior PT. Independent clusters were also found in left supplementary motor area (SMA), extending into preCC, and in left cerebellum. Anatomical overlap was remarkably similar for right hemisphere activity, but the frontal-temporal cluster remained intact, the right SMA was absent, and temporal activity was approximately closer to HG. A test for the forward prediction effect was also carried out across all voxels of HG. Mean parameter estimates for the silently articulated and imagined conditions were extracted from HG voxels for each participant and a paired t test between conditions was performed over participants (Figure 9A). A forward prediction effect for mean response in HG was present at the group level, with silently articulated tongue twisters producing higher mean parameter estimates across participants than imagined tongue twisters in both the left hemisphere (artic- ulated: M = 67.4, SD = 79.84; imagined: M = 36.39, SD = 58.1; t(33) = 3.62, Bonferroni- corrected p < 0.01) and the right hemisphere (articulated: M = 36.6, SD = 74.46; imagined: M = 4.15, SD = 59; t(33) = 4.52, Bonferroni-corrected p < 0.01). The forward prediction effect based on mean response in HG was present in roughly 76% (N = 26/34) of participants (Bonferroni-corrected p < 0.001; Figure 9B). Most of these par- ticipants (N = 20/26) presented with a bilateral effect, but in one participant the effect was present only in the left hemisphere, and in five participants it was present only in the right hemisphere. Four participants that showed the effect also showed a stronger response to base- line than silent articulation. Most participants that showed no forward prediction effect in Neurobiology of Language 101 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u n o / l / l a r t i c e - p d f / / / / / 4 1 8 1 2 0 7 4 4 9 4 n o _ a _ 0 0 0 8 8 p d . l f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Predictive coding and error correction Figure 10. Contrast maps for example participants who show no mean forward prediction effect in Heschl’s gyrus. Maximum and minimum z-scores vary among participants. Participants 2, 13, 14, 15, 22, and 32 show higher response for imagined than silently articulated speech across Heschl’s gyrus bilaterally. Participant 11 shows higher response to imagined speech in left Heschl’s gyrus and no significant difference in the right hemisphere. Note, there is still higher response to silently articulated speech within Heschl’s gyrus for all participants, except Participant 13, who still shows higher response to silently articulated speech in the posterior superior temporal sulcus. Participants 16 and 23 all show higher response during silently articulated speech than imagined speech in Heschl’s gyrus bilaterally, but also present with response during silently articulated speech that is below baseline fixation. These participants’ contrast maps are masked by activity that is above baseline during silent articulation. Each participant shows higher activity for silently articulated than imagined speech in areas bordering Heschl’s gyrus. l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u n o / l / l a r t i c e - p d f / / / / / 4 1 8 1 2 0 7 4 4 9 4 n o _ a _ 0 0 0 8 8 p d . l f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 either hemisphere presented with a significant effect in the reverse direction (i.e., imagined >
silently articulated across HG; N = 7/8), and a few participants with an established right hemi-
sphere forward prediction effect presented with a reverse effect in the contralateral hemisphere
(N = 4). Inspecting significant voxel-level differences between speech task conditions within
participants revealed that the majority of participants that failed to show a mean effect in HG
bilaterally still had voxels within that area that showed greater response to silently articulated
than imagined speech (p < 0.05, cluster corrected p < 0.05; N = 6/8). Notably, however, this activity often peaked or at least covered areas bordering HG, including the PT, the posterior STG, the anterior STG, and the planum polare (N = 4/6; see participants 2, 14, 22, and 32 in Figure 10). In at least one additional participant we found evidence of stronger response to silently articulated speech, but in the posterior superior temporal sulcus (N = 1/8; see partic- ipant 13). By masking participant-level contrast maps (i.e., silently articulated vs. imagined) by condition-level activation maps (i.e., silently articulated vs. baseline), we also found that half of the participants that presented with a mean effect in HG but showed below-baseline response during articulated speech (bilaterally) showed some voxels in HG that responded more strongly to articulated speech and produced above-baseline response (N = 2/4; see par- ticipants 16 and 23 in Figure 10). In general, we found at least one of the two effects that we sought (i.e., internal error correction and forward prediction) in 94% of participants (N = 32/34) and both effects in 59% of participants (N = 20). No significant relationship was documented between mean framewise displacement and participant-level t scores for contrasts between conditions (silently articulated vs. imagined speech in left HG: r(33) = 0.21, p = 0.21; silently articulated versus imagined speech in right HG: r(33) = 0.02, p = 0.93; taboo errors versus word errors in left pMTG: r(33) = −0.14, p = 0.43; taboo errors versus word errors in right pMTG: r(33) = −0.11, p = 0.53). Neurobiology of Language 102 Predictive coding and error correction Exploratory Findings Comparing the default mode and taboo networks The wider network that responded more strongly to taboo than word conditions while some- times showing below-baseline response (referred to as TN throughout) bears a strong resem- blance to the default mode network (DMN). Despite the stronger response to baseline fixation in some portions of this network, the fact that the class of the error word—the main difference between conditions—appears to be driving changes in activity entirely across the DMN is intriguing. Areas of the DMN commonly activate during mentalization and theory of mind tasks (Mars et al., 2012), suggesting that taboo words may be tapping into social cognitive systems. Here, we sought to describe the functional relationship between areas of the TN and the DMN. First, we demonstrated that areas of the TN are in fact recruited by the DMN. To define the boundaries of the DMN, we performed a meta-analysis for the phrase “default mode” in Neu- rosynth (FDR-corrected p < 0.01). The resulting meta-analytic map was found to cover canon- ical areas of the DMN (Mars et al., 2012): the anterior medial prefrontal cortex, pCG, AG, MTG, and paraHG. We refer to this map as the DMN throughout the results. The boundaries of the DMN were superimposed over the TN in Figure 11A, qualitatively showing that roughly the same regions activated across the two networks, but that activity was more focal in the TN and often formed a subset of each of the regions that was active in the DMN (Figure 11A). The major exception to this pattern was the paraHG, which appeared in the DMN but not in the l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u n o / l / l a r t i c e - p d f / / / / / 4 1 8 1 2 0 7 4 4 9 4 n o _ a _ 0 0 0 8 8 p d . l f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Figure 11. Relationship between taboo and default mode networks. Decoding the relationship between taboo and default mode networks. (A) Spatial overlap between cortical areas significantly associated with the taboo network (TN) from Figure 2 and Figure 3 and cortical areas significantly associated with the default mode network (DMN). The TN is represented by purple to yellow colors on the cortical surface and the DMN is represented by a maroon outline. The DMN is defined using the Neurosynth meta-analytic database, by analyzing activity associated with studies frequently using the term “default mode.” The resulting map represents areas significantly likely to activate across the set of studies frequently using this term (FDR corrected p < 0.01). (B) Percentage of TN voxels that overlap with the DMN, plotted as a function of network threshold. Overlap in the left hemisphere is plotted as a peach line and overlap in the right hemisphere is plotted as a teal line. Neurobiology of Language 103 Predictive coding and error correction TN. Although we saw particularly strong overlap between the two networks in the MTG, the TN contained large portions of the pMTG that did not appear in the DMN. We calculated that 45% of the voxels in the left TN and 36% of the voxels in the right TN overlapped with the DMN (Figure 11B). The finding of greater overlap in the left TN than the right was invariant to network significance thresholds (Figure 11B). To better understand the relationship between the TN and the DMN, we used Neurosynth to probe for the main functional associations of the TN and then compared the extent to which the same functions were associated with the DMN. That is, we executed the same meta-analysis procedure that we used to define the DMN over the entire set of 3,228 terms frequently used in the neuroimaging literature that are embedded in Neurosynth. We then estimated the Pearson correlation coefficient between each of these meta-analytic networks and the TN. Networks were not thresholded prior to correlation because our goal was to deter- mine purely whether the likelihood of seeing activity in a group of neuroimaging studies related to the likelihood of observing activity in the TN. However, the resulting correlation coefficients were corrected for multiple comparisons, and we focused only on the top 100 out of 2,952 significant correlations (FDR-corrected p < 0.001). Finally, each of the 100 net- works most strongly correlated to the TN were in turn correlated to the DMN (Figure 12A). The DMN (i.e., the term “default mode”) marked the strongest relationship to the TN, explaining roughly 20% of activity in the TN and validating our early assessment of the areas that form the TN. The set of unique terms that explained the greatest amount of variance in TN activity (r >
0.35, r2 > 0.12) additionally included “theory of mind,” “mentalizing,” “social” [traitement],
“social cognition,” “autobiographical memory,” “self-referential,” and “beliefs.” In this set,
associations were higher to the DMN only for the following terms: “default mode,” “self-
referential.” and “autobiographical memory.” For the other terms in the set, association to
the TN was on average modestly higher (mean difference in r = 0.07, standard error of the
mean = 0.01), except for the terms “social” [traitement] (DMN r = 0.27, TN r = 0.42) et
“social cognition” (DMN r = 0.23, TN r = 0.36), which showed a difference that was about
twice that of the mean.

A host of other terms were overall significantly associated with the TN as well. Broadly,
these covered other memory systems (par exemple., “episodic,” “semantic memory”), generic memory
processes (par exemple., “retrieval,” “recall,” “remembering,” “recollection”), emotional processes and
stimuli (par exemple., “emotions,” “affective,” “valence,” “emotional regulation”), references to resting
state (par exemple., “intrinsic,” “resting”), other social cognitive processes (par exemple., “social interaction,»
“inference,” “intentions,” “morals,” “personality,” “future”), decision-making (par exemple., “judg-
ments,” “value”), langue (par exemple., “semantic,” “sentence, “comprehension”; note semantic
memory does not show a strong difference in association strength to the TN and the DMN),
and disorders (par exemple., “depression,” “spectrum”).

Across all terms that more strongly associated with the TN than the DMN, the language-
related category of terms (par exemple., semantics and sentence comprehension) showed the largest
difference in association strength (c'est à dire., between TN and DMN), followed by emotion pro-
cesses and social cognition (see the bilateral plots in Figure 12B, which show the terms
from Figure 12A organized by association difference). Of the terms with stronger associa-
tion to the DMN, resting state terms and self-reference showed some of the largest differ-
ences with the TN. When we decomposed differences in association to the DMN and TN
by hemisphere, we found that the high association difference for language-related terms
was being entirely driven by activity in the left hemisphere, whereas the association difference
for resting state terms and self-reference was moderately higher in the right hemisphere
(Figure 12B).

Neurobiology of Language

104

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Chiffre 12. Groups of studies associated with the TN are found by correlating TN z-scores across the whole brain with the likelihood of seeing
activity for studies frequently using each of the other terms contained in the Neurosynth database. (UN) The top 100 significant correlations (FDR
corrected p < 0.001; r > 0.18) are presented as purple dots and organized by descending association strength. For each of these terms, le
correlation between the likelihood of seeing activity for that term and the likelihood of seeing activity for the “default mode” term (c'est à dire., DMN)
is also shown by red dots. A line is drawn between the purple and red dots to emphasize the magnitude of difference between a term and its
association to the TN and DMN. The color of the lines shows which of the two networks showed the strongest association to the term. (B)
Terms associated with the TN are sorted by descending difference in their association strength to the TN minus their association strength to the
DMN. Associations between each term and the two networks and the differences between them are computed separately for the left hemi-
sphère (pale purple dots), right hemisphere (deep violet dots), and bilaterally (c'est à dire., matching the data shown in panel (UN); medium violet dots).
Hemisphere-dependent variability is highlighted by connecting each set of dots by a line. The color of the line shows whether the term was
overall more strongly associated with the TN (violet line) or DMN (maroon line).

Neurobiology of Language

105

Predictive coding and error correction

Decoding regions of the taboo network

The finding that activity in the left TN more closely resembled language-related terms than
activity in the left DMN was remarkable because spatial overlap between the TN and DMN
was higher in the left hemisphere. This suggested that what small differences existed in the
distribution of activity between the left TN and DMN were functionally relevant for language.
We tested whether the largest of such differences, the segment of the left pMTG that does not
show overlap with the DMN was driving the taboo network-level association with language.
Pour faire ça, we separated the left MTG region in the TN into two segments based on its overlap
with the DMN (see top panel of Figure 13B) and then decoded the functions of both segments
using Neurosynth. We note that the portion of the pMTG in the TN that did not overlap with
the DMN was nearly identical to the largest region from the contrast between taboo and word
conditions that also showed above-baseline response (cf. Figure 9B with Figure 3). In the inter-
est of better characterizing the functional contributions of each region in the TN, we decoded
the other regions of the TN as well, including the unsegmented left MTG (see top panel of
Figure 13A). We decoded regions by generating a map for each term in Neurosynth that cap-
tures the posterior probability that a study uses a term if activity was observed in a particular
voxel, assuming a uniform prior (c'est à dire., all terms are equally likely to appear). Mean probabilities
were calculated across voxels of each region in the TN, and we presented all terms that
showed at least a 65% mean likelihood of being used when activity was observed in one
of the regions in the TN.

Decoding each region of the TN (Figure 13A) broadly showed that different functional cat-
egories were relatively more likely to be referenced in the context of different portions of the
TN. The majority of language-related terms, including sentence comprehension and lexical-
semantics were more likely to be used when there was activity in the left MTG compared to
other regions of the TN (par exemple., see “language network,” “reading,” “syntactic,” “phonological,»
“lexical,” “word,” “words,” “sentence,” “sentences,” “concepts,” “semantic,” “meaning,»
“aphasia” in Figure 13A). Auditory comprehension terms were most likely to be used when
the right MTG was active (par exemple., see “voice,” “speaker,” “speech,” “acoustic,” “comprehen-
sion,” “listening,” “linguistic” in Figure 9A). The DMN and social cognition, including theory
of mind and self-reference were most likely to be referenced when activity was in bilateral
pCG (par exemple., see “thoughts,” “moral,” “personal,” “theory of mind,” “episodic memory,” “auto-
biographical memory,” “self-referential,” “social interactions,” “default mode” in Figure 9A).
Beliefs were most likely to be referenced when either of parietal regions in the TN were active,
but the right parietal region was more strongly associated with multiple beliefs. The term
“mentalization,” but curiously also “read,” “readers,” and “written” were most likely to be
used when the right parietal region was active as well. “Word form,” “retrieval,” “recollec-
tion,” “reappraisal,” “semantics,” “judgment tasks” and “native English” are other terms that
were more likely to be used when activity was observed in the left parietal region. No terms
were more likely to be used when activity was found in the bilateral frontal cortex region
instead of anywhere else in the TN.

Despite the relative differences in term use probabilities between regions of the TN in dif-
ferent functional domains, it was overwhelmingly social cognitive and memory terms that
were most likely to be used when activity was found in any region of the TN. Par exemple,
the three terms that were most likely to appear when activity was found in each of the regions
of the TN (Figure 9A) were largely the same: “autobiographical” (left MTG, right MTG, gauche
pariétal, bilateral pCG, bilateral frontal cortex), “autobiographical memory” (right MTG, bilat-
eral pCG), “theory of mind” (right MTG, left parietal, bilateral frontal cortex), “mind” (gauche
MTG, bilateral frontal cortex), “mental states” (left MTG), “mentalizing” (right parietal),

Neurobiology of Language

106

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Chiffre 13. Decoding clusters of the TN. (UN) The association between clusters in the TN from Figure 2A and groups of different studies in
Neurosynth is found by computing the probability of studies in the database using a specific term when activity is reported in each of the
voxels belonging to a cluster, then taking the mean of these probabilities across voxels of the cluster. (Top) Clusters in the TN by color; clusters
are defined in volume space but projected onto the cortical surface to facilitate visualization. (Bottom) Mean probabilities for each cluster in a
set of terms that has a greater than 65% mean probability of activating voxels in at least one of the six TN clusters from this panel, or two
additional TN clusters from panel (B). The terms are presented in groups (by background color) based on which cluster was associated with the
highest mean probability of using a term. Terms are further organized within-group by the difference in probability for the most likely cluster
associated with a term, and the mean probability of the remaining clusters. Bars around each point represent standard error of the mean across
voxels. Clusters with ambiguous mean probabilities are not shown (<50%). (B) Two segments of the left MTG cluster from panel (A). (Top) Left MTG cluster from panel A segmented into a posterior area in left anatomical MTG that has no spatial overlap with the DMN, and an anterior area that has some spatial overlap with the DMN. (Bottom) Mean probabilities for the same set of terms in panel (A). Neurobiology of Language 107 Predictive coding and error correction “beliefs” (right parietal), and “default network” (right parietal, bilateral pCG), and “default mode” (left parietal). However, in segmenting the MTG region we found that language-related terms were more likely to be used than social cognitive terms when the posterior segment of the left MTG region was active (Figure 13B; e.g., “sentences,” “sentence,” “language net- work,” “language comprehension,” “comprehension,” “linguistic,” “speaker,” “voice”). In the context of activating the posterior segment of the left MTG, the highest likelihood of using a term from the social cognitive category was 67.6% for the term “autobiographical.” This likelihood was 2.4% lower than the likelihood of using the term “semantic” (70%), 3.7% lower than the likelihood of using the term “lexical” (71.3%), 7.5% lower than using the term “com- prehension” (75.1%), and 10.1% lower than using the term “sentences” (77.7%). These par- ticular language-related terms were highlighted here on the basis that they aligned with the same terms associated with activity across the entire TN (i.e., sentence comprehension and semantics), and that they additionally implicated the posterior segment of the left MTG region in lexical processes. Language-related terms were also more likely to be used when activity was observed in the posterior segment of the left MTG than the anterior segment of the left MTG. When the pos- terior segment of the left MTG was active, the likelihood of using the term “semantic” was 1.9% higher (posterior left MTG: 70%; anterior left MTG: 68.1%), the likelihood of using the term “lexical” was 8.5% higher (posterior left MTG: 71.3%; anterior left MTG: 62.8%), the likelihood of using the term “comprehension” was 13.2% higher (posterior left MTG: 75.1%; anterior left MTG: 61.9%), and the likelihood of using the term “sentences” was 19% higher (posterior left MTG: 77.7%; anterior left MTG: 58.7%). In general, the likelihood that language-related terms were used when the posterior segment of the MTG was active was strikingly high. The top percentile (>75.4%) of all probabilities that a term was used when
activity was found in any region of the TN was comprised of four terms for the pCG (c'est à dire., “auto-
biographical,” “autobiographical memory,” “default network”) and five terms for the posterior
segment of the MTG (“linguistic,” “speaker,” “sentence,” “sentences,” “language network”).
Entre-temps, social cognitive and memory terms were still most likely to be used when the
anterior segment of the MTG that overlapped with the DMN was active (par exemple., see “autobio-
graphical,” “mind,” “mental states,” “theory of mind,” “default mode” in Figure 13B). Splitting
the left MTG region into anterior and posterior portions typically increased the likelihood that
social cognitive terms were used in the anterior portion and decreases the likelihood that they
were used in the posterior portion (par exemple., see “autobiographical,” “mind,” “mental states,” “the-
ory of mind,” “default mode,” “social cognitive” in Figure 13B).

Enfin, the range of language-related terms associated with the posterior segment of the left
MTG was much broader than the range of terms associated with the whole left MTG. Ce
included terms related to auditory comprehension that were more strongly associated with
the right MTG than the left MTG region, but not more strongly associated with the right
MTG than the posterior segment of the left MTG. Other language-related terms that followed
this pattern included “visual word form,” “English,” “semantics,” and judgment task” (previ-
ously more strongly associated with the left parietal region than other regions), and “read,»
“readers,” and “written” (previously more strongly associated with the right parietal region
than other regions). En outre, several language-related terms were only likely to be used
when activity was found in the posterior segment of the left MTG, including the terms “ortho-
graphic,” “visual word,” “audiovisual,” and “speech perception.”

Despite the strong association between language-related terms and activity in the posterior
segment of the left MTG, we did document a couple of language-related terms that were more
likely to be used when activity was found in the anterior segment of the left MTG than the

Neurobiology of Language

108

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

posterior segment. These were “concepts” (posterior left MTG: 59.6%; anterior left MTG: 67%)
and “semantic memory” (posterior left MTG: 64.4%; anterior left MTG: 66.3%). Notably, other
regions were likely to activate these terms as well. The term “semantic memory” was about as
likely to be used when the bilateral pCG was active (66.2%), and only slightly less likely to be
used when the right MTG (62.7%) or the left parietal region of the TN (62.6%) were active. Le
term “concepts” also had a relatively moderate likelihood of being used when the left parietal
region of the TN was active (61%).

DISCUSSION

Main Effect of Internal Error Correction in Posterior Middle Temporal Gyrus

The main goal of the present research was to evaluate neural evidence for internal error detec-
tion and correction in speech production. To that end, we have leveraged a tongue twister
paradigm that has generated suggestive evidence of internal error correction during speech
production in the pMTG. C'est, prior research (Okada et al., 2018) had suggested that tongue
twisters designed such that slips would result in nonword errors generate greater activity in
pMTG than tongue twisters designed such that slips would result in word errors, even though
no overt errors occurred. This is intriguing because both conditions involved successfully recit-
ing the same real words, and the only difference between them was an increased potential to
produce a nonword speech error. The one caveat to this finding was that the lexicality effect
ultimately failed to survive cluster correction. Ici, we have tested for the same lexicality
effect, using the same stimuli, but in a much larger group of participants. En outre, we have
introduced a new class of stimuli designed to put greater demands on internal error correction
mechanisms by biasing potential speech errors toward taboo words.

The present study found no evidence of a lexicality effect. This could be because it is a
subtle effect not easily detectible with fMRI. Alternativement, it could be that the lexicality effect
is caused not by a word versus nonword difference in salience to the error detection system,
but by the nature of the lexical activation system itself, which has been argued to be biased
toward activating real words over nonwords in the first place (Nozari & Dell, 2009). On this
voir, we observed no lexicality effect of internal error processing because lexicality is not a
feature that this system naturally monitors. We submit that our null result is most consistent
with this latter possibility.

In contrast, we observed strong taboo word effect on pMTG activation (our a priori ROI).
Biasing potential speech errors toward taboo words during word list recitation generated a
significantly stronger response in the pMTG relative to biasing potential speech errors toward
(neutral) words. Surtout, this effect was observed on correctly produced trials where no
errors occurred, indicating that at some level the system detected taboo words even though
they were not present either in the stimulus or the ultimately expressed utterance. Based on the
psycholinguistic literature on taboo word slips, which occur at a lower frequency than non-
taboo word slips, we reason that during recitation of the taboo-biased word lists, slips occurred
yielding the lexical activation of a taboo word, lequel, due to its salience, had a higher chance
of being corrected than a non-taboo word slip (see Gabitov et al., 2020, for evidence of an
important relation between error detection and the brain’s salience network). This implies that
among the correct trials, internal error correction occurred more frequently for taboo-biased
lists than non-taboo biased lists leading to greater activation of the pMTG, an area implicated
in lexical access (par exemple., Gow, 2012; Indefrey, 2011; Lau & Namyst, 2019) and lexical-semantics
(par exemple., Baldo et al., 2013; Devereux et al., 2013).

Neurobiology of Language

109

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

In a whole brain contrast between taboo and non-taboo word conditions, we demon-
strated that portions of pMTG exhibited the taboo effect bilaterally. A more targeted ROI anal-
ysis using anatomically defined pMTG showed a stronger effect in the left hemisphere. Le
taboo effect held across all voxels of left pMTG but was on the cusp of significance in right
pMTG ( p = 0.05). Within-participant testing for the effect showed that it was quite consistent,
presenting in at least one hemisphere in 76% (N = 26/34) of participants. In the majority of
cases the effect was bilateral (N = 22/34), suggesting that a significant group-level effect may
be difficult to map in right pMTG due to individual variability in functional response.

Why should error correction result in greater activation of the pMTG? The hierarchical state
feedback control model proposed by Hickok (2012) at the phonological level of speech pro-
duction can be extended to the lexical level and provides a mechanistic explanation. Hickok
(2012) proposes that phonological encoding is implemented in a sensorimotor-like architec-
ture with a sensory-related system in the posterior temporal lobe that codes the phonological
targets of speech acts and a motor-related system in the inferior frontal lobe that codes
phonological-level motor planning codes (a “syllabary” in psycholinguistic terms) that aim
to reproduce the target. Error detection is achieved by comparing the activated “sensory-
phonological” target with the “motor-phonological” plan: Spécifiquement, the activated motor-
phonological plan generates a forward prediction signal (ultimately an inhibitory input) à
the sensory-phonological system where a match or mismatch is detected (via a cancellation
mechanism for a match or failure of cancellation for a mismatch). Note that this does not imply
that axons projection from motor to sensory systems are inhibitory signals. Long-range excit-
atory signals are likely to activate local inhibitory interneurons that generate the inhibitory
effect (Isaacson & Scanziani, 2011). In the case of a mismatch, the sensory-phonological target
remains active (it is not cancelled) and sends a corrective activation to its corresponding
motor-phonological plan. Ainsi, error correction results in greater action of the target
sensory-related network. A simple computational simulation presented by Hickok (2012) con-
firmed the feasibility of mechanism. Although Hickok (2012) did not propose that the lexical
level was organized with a similar sensorimotor-like architecture, it has recently been
extended to the lexical-syntactic level by Matchin and Hickok (2020; see also Isaacson &
Scanziani, 2011).

Main Effect of Forward Predictive Signals in Auditory Cortex

A secondary aim of the current research was to replicate evidence for forward predictive sig-
nals in auditory cortex (Okada et al., 2018). We successfully replicated this effect, montrant que
auditory cortex does seem to receive predictive signals during speech production. C'est, nous
found that engaging motor-phonological and lexical-level processes together during silently
articulated speech generated significantly greater activity in and around auditory cortex than
engaging only lexical-level processes during imagined speech. Critique, the activity we
observed in auditory cortex cannot reflect auditory stimulation because no speech output
was generated in either of the speech task conditions. This forward predictive signal effect
was apparent at the group-level based on a whole brain contrast between speech tasks, et
a ROI analysis that showed that the effect held across all voxels of anatomically defined HG in
both hemispheres. Within-participant testing revealed that the forward predictive signal effect
was about as consistent as the taboo effect and presented in the same proportion of participants
(N = 26/34). Cependant, we also found that the majority of participants who did not exhibit a
mean forward predictive signal effect in HG still presented higher brain response to silently
articulated speech either in portions of HG or in bordering auditory association areas, showing
that there can be substantial individual variability in the precise locus of the effect.

Neurobiology of Language

110

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

Response around auditory cortex can be explained in the context of a mismatch error
between the auditory consequence of a planned articulatory sequence that is expected by
the system, and the absence of any such signal as a result of silent articulation. Mismatch
error induced by altered auditory feedback has produced a similar pattern of response in
auditory cortex (Tourville et al., 2008). Speech induced suppression of auditory response
as documented by electrocorticographic, electroencephalographic, and magnetoencephalo-
graphic recordings has also been interpreted in the context of a similar kind of mismatch error
between a forward prediction of auditory consequences of speech and actual auditory input
(Ford et al., 2010; Greenlee et al., 2011; Guenther & Hickok, 2015; Ventura et al., 2009). Dans
models of motor control, mismatch is often extended not only to the content of speech, mais
also to its fundamental acoustic properties (par exemple., pitch; Guenther & Hickok, 2015). Ce
explains why FMRI studies have demonstrated a similar decrease in auditory response during
silently articulated speech when it is coupled with hearing another person produce the same
speech stimuli (Agnew et al., 2013).

Network Associated With the Taboo Effect

Comparison between the taboo and word condition revealed preferential response to the
taboo condition in areas outside pMTG and commonly associated with the DMN, y compris
in AG and adjacent portions of sLOC, aMTG, pCG, paraCG, SFG, and several medial frontal
domaines. An additional comparison between the taboo condition and baseline fixation demon-
strated that most areas that showed a taboo effect were deactivated during the taboo condition.
The strongest evidence for above-baseline response was found in left pMTG, although we note
that an areal analysis we performed also implied above-baseline response could be found
within left AG and left pITG at significance thresholds lower than those used in the present
study for direct contrast analyses.

The finding that left pMTG showed a taboo effect and had the clearest above-baseline
response during the taboo condition implicated this area in internal error correction above
other areas that exhibited a taboo effect. It also implied that internal error correction processes
are likely left lateralized. This is further supported by the aforementioned ROI analysis of
pMTG that demonstrated a stronger effect in the left hemisphere. The broader network of areas
that exhibited a taboo effect also showed evidence for left lateralization—maximum activation
within areas was significantly higher in the left hemisphere.

Recent research has also implicated left pMTG in word–picture interference using taboo
distractor words (Hansen et al., 2019). This work has also demonstrated increased activity
for taboo distractors in a thalamocortical network of areas that do not appear in the taboo
effect we map using tongue twisters. The functional interpretation of this thalamocortical net-
work is consistent with our results here—left pMTG is interpreted to reflect lexical processes,
whereas other areas of the thalamocortical network (par exemple., aCG, thalamus, IFG) support pro-
cessing the arousing properties of taboo words and deploying attention mechanisms to sup-
press them. That our results did not reveal areas strongly associated with attention suggests that
internal correction of speech errors, including taboo speech errors, may not require interven-
tion from attentional mechanisms.

Although we predicted a robust response in left pMTG for taboo tongue twisters, we did not
expect them to generate increased, albeit below-baseline, activity across areas of the DMN.
We suggest that the taboo effect tapped into a broader lexical-semantic network and that activ-
ity in the DMN was driven by the social features of taboo words. Areas of the DMN appear to
show a preference for social concepts embedded in narrative stories (Huth et al., 2016) et

Neurobiology of Language

111

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

respond more strongly to words that connect to interpersonal interactions compared to those
that do not (Lin et al., 2015, 2020). Observations like these have been used to argue that the
brain areas responsible for theory of mind encode intentionality in semantic space (Binder
et coll., 2016). In this context, that taboo words generally elicited stronger response across
the DMN is less surprising. Taboo words tend to be highly versatile and convey very different
intentions (positive, negative, and otherwise) depending on the context in which they are used
(Vingerhoets et al., 2013). Par exemple, work in natural language processing has demonstrated
that detection of hate speech is considerably less successful, suffering from high false positive
rates, when the pragmatic functions of taboo words are not explicitly modeled (Davidson
et coll., 2017).

Networks Associated With Speech Tasks

Previous research (Okada et al., 2018) has demonstrated that the contrast between silently
articulated and imagined speech produces activity that is right lateralized in certain regions
(par exemple., sensorimotor cortex). In the present study we broke this contrast apart into anatomical
domaines, which showed that although generally the same areas were recruited bilaterally and
maximum response within areas did not differ between hemispheres, larger portions of areas
in the left hemisphere were sensitive to the difference between articulated and imagined
speech, while the smaller portions of areas in the right hemisphere that were sensitive to this
difference presented larger effect sizes. Dans l'ensemble, this describes a left hemisphere network that
was more sensitive to differences between silently articulated and imagined speech and a right
hemisphere network with higher specificity. En fait, the largest two effect sizes we reported
were in the right PT and HG, underlining the robustness of auditory response during silently
articulated speech. It is worth noting too that HG was one of the few areas that broke away
from the trend we described across the contrast—response in HG was substantially higher in
the right hemisphere, a much larger portion of the right HG was sensitive to the contrast, et
right HG also showed a larger effect size. This pattern echoed our findings at the participant
level, where a forward prediction effect was typically present in HG for both hemispheres (N =
20/26), but when the effect was occasionally unilateral it was most often found in the right
hemisphere (N = 5). Ainsi, it is possible that right auditory cortex plays a larger role in process-
ing forward predictive models in speech.

At the group level, greater response during silently articulated than imagined speech was
found in a network that closely followed recent pooled analysis of speech production exper-
iments (Tourville et al., 2019). The core of this network was highly consistent with the prior
study we sought to replicate and primarily covered areas connected to phonological and artic-
ulatory processes, including the primary sensorimotor and somatosensory cortices (c'est à dire., pre-
central and postcentral gyri), aCG, SMA, the cerebellum, the IFG, the PT, the supramarginal
gyrus, the insula, and the adjacent central, frontal, and parietal operculum (par exemple., Hickok, 2012,
2014; Kearney & Guenther, 2019; Prix, 2012; Tourville et al., 2019). As alluded to earlier, le
silent articulation network also contained areas involved in auditory processing, which were
HG and the planum polare, but also associative auditory areas in anterior and posterior seg-
ments of the STG (Binder et al., 1996; Kearney & Guenther, 2019). We note that the anatom-
ical STG we used for reference included portions of the posterior superior temporal sulcus
putatively involved in speech perception (Venezia et al., 2017). Stronger activity in these areas
is consistent with the behavioral research this study builds on, which has demonstrated that
increasing the amount of articulation in speech imagery induces higher rates of phonological
speech errors that involve similarly articulated phonemes (Oppenheim & Dell, 2010). Le
silent articulation network also recruited areas not as closely related to speech production.

Neurobiology of Language

112

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

Par exemple, some of these areas are commonly associated with executive function (c'est à dire., le
intraparietal sulcus, which crosses anatomically defined superior parietal lobule, posterior
supramarginal gyrus, and sLOC; Velenosi et al., 2020), orthographic processing (c'est à dire., tempor-
ooccipital portion of ITG and neighboring fusiform areas; Prix, 2012), and lexical and/or
semantic processing (c'est à dire., MTG, temporal pole, middle frontal gyrus, SFG, and an area of ana-
tomically defined orbitofrontal cortex that maps onto the pars orbitalis portion of the IFG in
atlases that make the distinction between pars orbitalis and pars triangularis; par exemple., Binder et al.,
2009; Prix, 2012). Nevertheless, all of these areas play a role during speech production
(Prix, 2012; Tourville et al., 2019), and it is possible that the phonological-level processes
that distinguished silently articulated speech from imagined speech had the effect of producing
greater activity in areas that interface with phonological features.

Curiously, some areas responded more strongly to imagined than silently articulated speech,
including paraHG, AG, sLOC, temporal fusiform cortex, OccP, and several medial frontal areas.
These are many of the same areas that exhibited stronger response to taboo tongue twisters rel-
ative to word tongue twisters, and just as in the case of the taboo effect, they were all deactivated
during the imagined speech condition. Cependant, areal effect analysis hinted that left AG, FP, et
anterior paraHG may contain above-baseline response at lower significance thresholds than
those used in the current work for contrast analyses. Although prior research did not observe
any areas associated with imagined speech (Okada et al., 2018), a similar network has been
documented in auditory and visual imagery tasks (par exemple., Daselaar et al., 2010; Pearson, 2019;
Tian et al., 2016), and functional connectivity between the AG and the other areas of this net-
work has been shown to increase during imagined musical performance (instrumental or vocal;
Tanaka & Kirino, 2019). The AG itself has been routinely implicated in semantic, episodic, et
autobiographical memory, and response in this area has been related to subjective vividness
during episodic memory retrieval and encoding (Tibon et al., 2019). Similarity between areas
that show up in the DMN and areas involved in various mental imagery tasks (including auditory
imagery) has implied a close relationship between mental imagery and the kinds of internally
directed processes that have been connected to DMN areas (Daselaar et al., 2010; Pearson,
2019). Ainsi, the pattern of activity that we found could simply reflect the fact that imagined
speech is relatively more internally oriented than articulated speech, or perhaps that imagined
articulation taps into a domain general imagery network.

Given the association between many of the areas in the imagined speech network and
mémoire, another possibility is that imagined speech recruits memory systems to facilitate
imaging the auditory consequences of simulated speech under conditions where precise pre-
diction is more difficult. Although we have shown here that forward predictions were stronger
when articulatory features were salient in speech imagery (c'est à dire., silently articulated speech),
they still appear to be generated in a weaker form when such features are impoverished or
absent (c'est à dire., imagined speech; Tian et al., 2016). Plus loin, if the quality of forward predictions
produced in this memory-driven process (Tian et al., 2016) can be substantially different, it
could also explain our observation that different portions of auditory cortex can simulta-
neously show greater activity for silently articulated and imagined speech in a minority of par-
ticipants. At least one other explanation is that the areas associated with imagined speech
reflect a different process for maintaining speech images in memory. Par exemple, quelques
research has related activity in areas of the imagined speech network to the detail of ongoing
thoughts during working memory maintenance (Sormaz et al., 2018), and other studies have
implicated the paraHG in the maintenance of novel information during working memory
(Schon et al., 2016). Because verbal memory taps into areas involved in speech production
(Buchsbaum & D'Esposito, 2019), it may be the case that weaker access to articulatory features

Neurobiology of Language

113

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

in imagined speech requires additional engagement in nonverbal memory systems for main-
taining speech imagery. More research is necessary to adjudicate between these possibilities
and we emphasize that no areas that generated greater response to imagined than articulated
speech showed clear above-baseline response during the imagined speech condition.

Decoding the Taboo Network

We have suggested that the richness of social features in taboo words—particularly
intentionality—has driven activity to areas of the DMN in the taboo contrast. We have also
suggested that activity in the left pMTG does not reflect these social features, in part because
this area alone shows clear above-baseline response during the recitation of taboo tongue
twisters. A functional distinction between pMTG and other areas of the taboo network was
further supported by a meta-analysis of studies that frequently mentioned the DMN, lequel
showed strong overlap in most areas, but not pMTG.

Our explanation for the response pattern in the TN predicted that pMTG should be asso-
ciated with language-specific processes, especially at the word level, while other areas of the
TN should be associated with DMN functions relevant to processing intentionality. We tested
these predictions by decoding the taboo network and its constituent regions in Neurosynth
(Yarkoni et al., 2011). Decoding the TN involved correlating activity that we mapped as part
of the taboo effect (c'est à dire., taboo > words) with the likelihood of finding activity in different
groups of neuroimaging studies that frequently mentioned different neuroimaging terms
(c'est à dire., meta-analyses for different terms in Neurosynth). This unsurprisingly revealed that
DMN studies were most strongly associated with TN activity, followed by studies on func-
tions commonly ascribed to areas of the DMN (par exemple., “theory of mind,” “mentalization,” “self-
reference,” “autobiographical memory,” “social cognition”). Cependant, activity in the TN was
also related to comprehension (par exemple., “comprehension,” “semantic,” “sentence”). When we
correlated the same term-based meta-analyses with the likelihood of observing activity in the
DMN (c'est à dire., the earlier DMN meta-analysis), we found that comprehension terms showed by
far the largest difference between association strength to the TN and the DMN. De plus,
this difference in association was primarily driven by activity in the left hemisphere, lequel
also overlapped more extensively with the DMN. This pattern suggested that left hemisphere
areas of the TN that fall outside the DMN play a more substantial role in comprehension, et
therefore that left pMTG may be driving the network-level association between the TN and
comprehension.

We directly tested whether the posterior portions of the left MTG cluster in the TN was more
closely connected to word-level processing by decoding the TN at a regional level, decoding
both the entire MTG cluster as well as the posterior and anterior portions of it that sit outside
and inside of the DMN, respectivement. Decoding was made to be more sensitive to the spatial
distribution of activity within smaller regions by capturing the mean likelihood that a study
uses a particular term (that is frequently used in neuroimaging studies) if there is activity in
each voxel of a TN cluster. We found that although virtually all language-related terms were
more likely to be used when there was activity in the MTG compared to other regions of the
TN, seeing activity in any cluster of the TN, including the MTG, was most likely to signal
engagement in “autobiographical memory,” “theory of mind,” “mentalization,” or “default
mode network.” This is exactly what might have been expected based on the results of the
network-level decoding. Cependant, separating the left MTG cluster based on overlap with
the DMN produced a pattern that supported word processing in the posterior segment of this
region. D'abord, we found that activity in the anterior segment of MTG was still most likely to

Neurobiology of Language

114

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

signal DMN-related functions, whereas activity in the posterior segment was most likely to
signal engagement in language processes. Plus loin, relative to the whole MTG cluster, the like-
lihood of using language-related terms generally increased and the likelihood of using
DMN-related terms generally decreased when the posterior segment of MTG was active.
The reverse of this trend was observed in the anterior segment of MTG. Closer inspection of
the terms that were most likely to be used when the posterior segment of the MTG was active
revealed an association with comprehension that was broad, covering both sentence and word
stimuli. Terms that were likely to be used such as “speaker,” “acoustic,” “lexical,” “meaning,»
“comprehension,” “word,” “linguistic,” and “phonological” were all consistent with a role for
the pMTG in lexical processing. Other terms, such as “semantic,” “sentences,” “language
comprehension,” “syntactic,” and “reading” may reflect the activity of lexical processes
engaged during semantic tasks. In any case, these patterns clearly showed that one specific
portion of the TN—the left pMTG—was much more strongly associated with language, and at
least played a relatively more significant role in lexical processing.

Regional decoding of the TN also indicated that the posterior segment of MTG was
involved in semantic processes, but not necessarily semantic memory. Activity in the anterior
segment of MTG was more strongly associated with concepts and semantic memory than
activity in the posterior segment of MTG. Plus loin, semantic memory was just as strongly asso-
ciated with the pCG and had a remarkably weak association with the posterior segment of
MTG (c'est à dire., less than 50% likelihood of term use). The stronger association to both social cog-
nitive terms and semantic memory in areas of the TN that happened to show stronger overlap
with the DMN was consistent with our interpretation that taboo words drive response in areas
of the DMN as a result of their semantic features, mainly intentionality. This interpretation
aligns with a recent proposal for a componential model of semantic representation that sug-
gests areas involved in theory of mind and mentalization are used to encode and understand
words that place stronger emphasis on intentionality (Binder et al., 2016). It also aligns with a
growing body of work that shows words rich in social semantic features tap into areas that
overlap with the DMN (par exemple., Huth et al., 2016; Lin et al., 2015, 2018, 2020; Vingerhoets
et coll., 2013).

Conclusions and Limitations

En résumé, the present study has provided evidence that forward predictive signals are pres-
ent in auditory cortex during speech production, and that error detection and correction in
speech involves left pMTG. We have shown that silently articulated speech produces greater
activity than imagined speech within early auditory areas, even though no auditory input is
present during silently articulated speech. We have also shown that biasing potential word
errors toward taboo words rather than neutral words during word list recitation generates
greater activity in left pMTG, even though no overt speech errors are produced in either con-
dition. Although we found similar response in areas that demonstrably overlapped with the
DMN as well as right pMTG, these areas appeared to respond more robustly to baseline fix-
ation than word list recitation. Plus loin, we have provided evidence from meta-analysis that
activity in left pMTG is most likely to reflect word-level processes and by extension word-level
error correction. We have hypothesized, with some support from meta-analysis, that the other
areas of the taboo network are involved in processing the social semantic features of taboo
words, particularly intentionality.

Nevertheless, there are several limitations to the present research. D'abord, overt speech typ-
ically increases the amount of motion observed during scanning. Although here we used

Neurobiology of Language

115

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

mouthed speech, it is difficult to assess the full impact motion had on the results. En général,
we reported what can be considered as relatively low motion in our group of participants
(par exemple., Eichenbaum et al., 2021; Yang et al., 2019). We also tried to mitigate the possibility that
motion had a substantial impact on our results by showing that our primary effects of interest
(c'est à dire., the ROI analyses for forward prediction and internal error correction) are not correlated
with mean framewise displacement. One other concern is that overt speech has been shown
to produce susceptibility artifacts. It is uncertain the extent to which such artifacts may have
impinged on our results, but we note that past studies have found these artifacts to dispro-
portionately affect insular and opercular areas (Kemeny et al., 2005), whereas the current
study focuses on lateral temporal cortex. Enfin, our failure to find a lexicality effect in
pMTG for tongue twisters suggests that evidence for such an effect trending toward signifi-
cance in prior research may reflect a false positive result. Cependant, we suspect this is not the
case and that a larger amount of data is necessary to achieve significance for such a small
effect size, especially considering that our manipulation of tongue twisters to elicit a greater
potential for taboo word errors did successfully produce stronger activity in pMTG, as we
had hypothesized.

REMERCIEMENTS

The authors wish to thank the individuals that participated in this research. Discretionary
research funding was provided by University of California, Irvine.

INFORMATIONS SUR LE FINANCEMENT

Gregory Hickok, National Institute on Deafness and Other Communication Disorders (https://
dx.doi.org/10.13039/100000055), Award ID: R01 DC009659. Alex Teghipco, National Sci-
ence Foundation (https://dx.doi.org/10.13039/100000001), Award ID: DGE-1321846.

CONTRIBUTIONS DES AUTEURS

Alex Teghipco: Conservation des données: Lead; Analyse formelle: Lead; Enquête: Lead; Méthodologie:
Supporting; Logiciel: Lead; Visualisation: Lead; Rédaction – ébauche originale: Lead; Rédaction – révision
& édition: Equal. Kayoko Okada: Conceptualisation: Equal; Analyse formelle: Supporting; Fund-
ing acquisition: Supporting; Enquête: Supporting; Méthodologie: Equal; Ressources: Equal;
Surveillance: Equal; Rédaction – ébauche originale: Equal; Rédaction – révision & édition: Equal. Emma
Murphy: Conservation des données: Supporting; Analyse formelle: Supporting; Enquête: Supporting;
Ressources: Supporting. Gregory Hickok: Conceptualisation: Lead; Acquisition de financement: Lead;
Méthodologie: Equal; Surveillance: Equal; Rédaction – ébauche originale: Supporting; Rédaction – révision
& édition: Equal.

DATA AND CODE AVAILABILITY STATEMENT

Group-level contrast maps for this study can be found on neurovault (https://identifiers.org
/neurovault.collection:12982). Raw data is available to download on data dryad (https://est ce que je
.org/10.7280/D1BH77). Example code for replicating meta-analyses can be found on our
GitHub repository (https://github.com/alexteghipco/MetaAnalysisResources). FMRI analyses
can be replicated with FMRIB’s Software Library ( Version 6.0.5.1; Jenkinson et al., 2012),
using parameters provided in the methods (see FSL course preparatory material for guides
on how to replicate purely using GUIs). Additional parameter information that may be useful
is provided along with the neurovault collection linked above as metadata. Surface visualiza-
tions can be replicated using the brainSurfer toolbox for MATLAB (https://doi.org/10.5281

Neurobiology of Language

116

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

/zenodo.7271544). A saved GUI state that can be loaded into brainSurfer (Teghipco, 2022) à
show interactive and manipulable surface renderings identical to those shown in our figures
(c'est à dire., including all thresholds, colormaps, and other visualization settings) can be found on fig-
share (10.6084/m9.figshare.21357525).

RÉFÉRENCES

Agnew, Z. K., McGettigan, C., Banks, B., & Scott, S. K. (2013).
Articulatory movements modulate auditory responses to speech.
NeuroImage, 73, 191–199. https://doi.org/10.1016/j.neuroimage
.2012.08.020, PubMed: 22982103

Baars, B. J., Motley, M.. T., & MacKay, D. G. (1975). Output editing
for lexical status in artificially elicited slips of the tongue. Journal
of Verbal Learning and Verbal Behavior, 14(4), 382–391. https://
doi.org/10.1016/S0022-5371(75)80017-X

Baldo, J.. V., Arévalo, UN., Patterson, J.. P., & Dronkers, N. F. (2013).
Grey and white matter correlates of picture naming: Evidence
from a voxel-based lesion analysis of the Boston Naming Test.
Cortex, 49(3), 658–667. https://doi.org/10.1016/j.cortex.2012
.03.001, PubMed: 22482693

Binder, J.. R., Conant, L. L., Humphries, C. J., Fernandino, L.,
Simons, S. B., Aguilar, M., & Desai, R.. H. (2016). Toward a
brain-based componential semantic representation. Cognitive
Neuropsychology, 33(3–4), 130–174. https://est ce que je.org/10.1080
/02643294.2016.1147426, PubMed: 27310469

Binder, J.. R., Desai, R.. H., Graves, W. W., & Conant, L. L. (2009).
Where is the semantic system? A critical review and
meta-analysis of 120 functional neuroimaging studies. Cérébral
Cortex, 19(12), 2767–2796. https://doi.org/10.1093/cercor
/bhp055, PubMed: 19329570

Binder, J.. R., Frost, J.. UN., Hammeke, T. UN., Rao, S. M., & Cox, R.. W.
(1996). Function of the left planum temporale in auditory and
linguistic processing. Cerveau, 119(4), 1239–1247. https://doi.org
/10.1093/brain/119.4.1239, PubMed: 8813286

Borgo, M., Soranzo, UN., & Grassi, M.. (2012). Psychtoolbox: Sound,
keyboard and mouse. In MATLAB for psychologists (pp. 249–273).
Springer. https://doi.org/10.1007/978-1-4614-2197-9_10

Buchsbaum, B. R., & D'Esposito, M.. (2019). A sensorimotor view of
verbal working memory. Cortex, 112, 134–148. https://est ce que je.org/10
.1016/j.cortex.2018.11.010, PubMed: 30639088

Corley, M., Brocklehurst, P.. H., & Moat, H. S. (2011). Error biases
in inner and overt speech: Evidence from tongue twisters. Journal
of Experimental Psychology: Apprentissage, Mémoire, et cognitif,
37(1), 162–175. https://doi.org/10.1037/a0021321, PubMed:
21244112

Daselaar, S. M., Porat, Y., Huijbers, W., & Pennartz, C. M.. (2010).
Modality-specific and modality-independent components of the
human imagery system. NeuroImage, 52(2), 677–685. https://est ce que je
.org/10.1016/j.neuroimage.2010.04.239, PubMed: 20420931
Davidson, T., Warmsley, D., Macy, M., & Weber, je. (2017). Auto-
mated hate speech detection and the problem of offensive lan-
guage. Proceedings of the International AAAI Conference on
Web and Social Media, 11(1), 512–515. https://doi.org/10.1609
/icwsm.v11i1.14955

Dell, G. S. (1986). A spreading-activation theory of retrieval in sen-
tence production. Psychological Review, 93(3), 283–321. https://
doi.org/10.1037/0033-295X.93.3.283, PubMed: 3749399

Desikan, R.. S., Ségonne, F., Fischl, B., Quinn, B. T., Dickerson,
B. C., Blacker, D., Buckner, R.. L., Dale, UN. M., Maguire, R.. P.,

Hyman, B. R., Albert, M.. S., & Killiany, R.. J.. (2006). An auto-
mated labeling system for subdividing the human cerebral cortex
on MRI scans into gyral based regions of interest. NeuroImage,
31(3), 968–980. https://doi.org/10.1016/j.neuroimage.2006.01
.021, PubMed: 16530430

Devereux, B. J., Clarke, UN., Marouchos, UN., & Tyler, L. K. (2013).
Representational similarity analysis reveals commonalities and
differences in the semantic processing of words and objects. Jour-
nal of Neuroscience, 33(48), 18906–18916. https://est ce que je.org/10
.1523/JNEUROSCI.3809-13.2013, PubMed: 24285896

Eichenbaum, UN., Pappas, JE., Lurie, D., Cohen, J.. R., & D'Esposito,
M.. (2021). Differential contributions of static and time-varying
functional connectivity to human behavior. Network Neurosci-
ence, 5(1), 145–165. https://doi.org/10.1162/netn_a_00172,
PubMed: 33688610

Ford, J.. M., Roach, B. J., & Mathalon, D. H. (2010). Assessing cor-
ollary discharge in humans using noninvasive neurophysiologi-
cal methods. Nature Protocols, 5(6), 1160–1168. https://doi.org
/10.1038/nprot.2010.67, PubMed: 20539291

Gabitov, E., Lungu, O., Albouy, G., & Doyon, J.. (2020). Movement
errors during skilled motor performance engage distinct prediction
error mechanisms. Communications Biology, 3(1), Article 763.
https://doi.org/10.1038/s42003-020-01465-4, PubMed: 33311566
Garnham, UN., Shillcock, R.. C., Brun, G. D., Mill, UN. je. D., & Cutler,
UN. (1982). Slips of the tongue in the London-Lund corpus of spon-
taneous conversation. In A. Cutler (Ed.), Slips of the tongue and
language production (pp. 251–263). Walter de Gruyter. https://
doi.org/10.1515/9783110828306.251

Gow, D. W., Jr.. (2012). The cortical organization of lexical knowl-
bord: A dual lexicon model of spoken language processing.
Brain and Language, 121(3), 273–288. https://est ce que je.org/10.1016/j
.bandl.2012.03.005, PubMed: 22498237

Grabner, G., Janke, UN. L., Budge, M.. M., Forgeron, D., Pruessner, J., &
Collins, D. L. (2006). Symmetric atlasing and model based seg-
mentation: An application to the hippocampus in older adults. Dans
International Conference on Medical Image Computing and
Computer-Assisted Intervention (pp. 58–66). Springer. https://
doi.org/10.1007/11866763_8, PubMed: 17354756

Greenlee, J.. D. W., Jackson, UN. W., Chen, F., Larson, C. R., Oya, H.,
Kawasaki, H., Chen, H., & Howard, M.. UN., III. (2011). Human
auditory cortical activation during self-vocalization. PLOS One,
6(3), Article e14744. https://doi.org/10.1371/journal.pone
.0014744, PubMed: 21390228

Guenther, F. H., & Hickok, G. (2015). Role of the auditory system in
speech production. Handbook of Clinical Neurology, 129, 161–
175. https://doi.org/10.1016/ B978-0-444-62630-1.00009-3,
PubMed: 25726268

Hansen, S. J., McMahon, K. L., & de Zubicaray, G. je. (2019). Le
neurobiology of taboo language processing: fMRI evidence dur-
ing spoken word production. Social Cognitive and Affective Neu-
roscience, 14(3), 271–279. https://doi.org/10.1093/scan/nsz009,
PubMed: 30715549

Neurobiology of Language

117

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

Hickok, G. (2012). Computational neuroanatomy of speech pro-
duction. Nature Revues Neurosciences, 13(2), 135–145. https://
doi.org/10.1038/nrn3158, PubMed: 22218206

Hickok, G. (2014). The architecture of speech production and the
role of the phoneme in speech processing. Language and Cogni-
tive Processes, 29(1), 2–20. https://doi.org/10.1080/01690965
.2013.834370, PubMed: 24489420

Huth, UN. G., de Heer, W. UN., Griffiths, T. L., Theunissen, F. E., &
Galant, J.. L. (2016). Natural speech reveals the semantic maps
that tile human cerebral cortex. Nature, 532(7600), 453–458.
https://doi.org/10.1038/nature17637, PubMed: 27121839

Indefrey, P.. (2011). The spatial and temporal signatures of word pro-
duction components: A critical update. Frontiers in Psychology,
2, 255. https://doi.org/10.3389/fpsyg.2011.00255, PubMed:
22016740

Isaacson, J.. S., & Scanziani, M.. (2011). How inhibition shapes cor-
tical activity. Neurone, 72(2), 231–243. https://est ce que je.org/10.1016/j
.neuron.2011.09.027, PubMed: 22017986

Jacewicz, E., Fox, R.. UN., & Wei, L. (2010). Between-speaker and
within-speaker variation in speech tempo of American English. Le
Journal of the Acoustical Society of America, 128(2), 839–850.
https://doi.org/10.1121/1.3459842, PubMed: 20707453

Jenkinson, M., Bannister, P., Brady, M., & Forgeron, S. (2002).
Improved optimization for the robust and accurate linear registra-
tion and motion correction of brain images. NeuroImage, 17(2),
825–841. https://doi.org/10.1006/nimg.2002.1132, PubMed:
12377157

Jenkinson, M., Beckmann, C. F., Behrens, T. E. J., Woolrich, M.. W.,
& Forgeron, S. M.. (2012). FSL. NeuroImage, 62(2), 782–790. https://
doi.org/10.1016/j.neuroimage.2011.09.015, PubMed: 21979382
Kawato, M.. (1999). Internal models for motor control and trajectory
planning. Opinion actuelle en neurobiologie, 9(6), 718–727.
https://doi.org/10.1016/S0959-4388(99)00028-8, PubMed:
10607637

Kearney, E., & Guenther, F. H. (2019). Articulating: The neural
mechanisms of speech production. Language, Cognition and
Neurosciences, 34(9), 1214–1229. https://est ce que je.org/10.1080
/23273798.2019.1589541, PubMed: 31777753

Kemeny, S., Ye, F. Q., Birn, R., & Brun, UN. R.. (2005). Comparaison
of continuous overt speech fMRI using BOLD and arterial spin
labeling. Human Brain Mapping, 24(3), 173–183. https://est ce que je
.org/10.1002/hbm.20078, PubMed: 15486986

Lau, E. F., & Namyst, UN. (2019). fMRI evidence that left posterior
temporal cortex contributes to N400 effects of predictability
independent of congruity. Brain and Language, 199, 104697.
https://doi.org/10.1016/j.bandl.2019.104697, PubMed:
31585316

Levelt, W. J.. (1983). Monitoring and self-repair in speech. Cogni-
tion, 14(1), 41–104. https://doi.org/10.1016/0010-0277(83)
90026-4, PubMed: 6685011

Levelt, W. J.. (1989). Speaking: From intention to articulation. AVEC

Presse.

Levelt, W. J.. (1992). Accessing words in speech production: Stages,
processes and representations. Cognition, 42(1–3), 1–22. https://
doi.org/10.1016/0010-0277(92)90038-J., PubMed: 1582153
Levelt, W. J., Roelofs, UN., & Meyer, UN. S. (1999). A theory of lexical
access in speech production. Behavioral and Brain Sciences, 22(1),
1–38. https://doi.org/10.1017/S0140525X99001776, PubMed:
11301520

Lin, N., Bi, Y., Zhao, Y., Luo, C., & Li, X. (2015). The theory-of-mind
network in support of action verb comprehension: Evidence from
an fMRI study. Brain and Language, 141, 1-dix. https://est ce que je.org/10
.1016/j.bandl.2014.11.004, PubMed: 25498409

Lin, N., Wang, X., Xu, Y., Wang, X., Hua, H., Zhao, Y., & Li, X.
(2018). Fine subdivisions of the semantic network supporting
social and sensory–motor semantic processing. Cortex cérébral,
28(8), 2699–2710. https://doi.org/10.1093/cercor/ bhx148,
PubMed: 28633369

Lin, N., Xu, Y., Lequel, H., Zhang, G., Zhang, M., Wang, S., Hua, H.,
& Li, X. (2020). Dissociating the neural correlates of the sociality
and plausibility effects in simple conceptual combination. Cerveau
Structure and Function, 225(3), 995–1008. https://est ce que je.org/10
.1007/s00429-020-02052-3, PubMed: 32140848

Mars, R.. B., Neubert, F. X., Noonan, M.. P., Sallet, J., Toni, JE., &
Rushworth, M.. F. (2012). On the relationship between the
“default mode network” and the “social brain.” Frontiers in
Human Neuroscience, 6, 189. https://doi.org/10.3389/fnhum
.2012.00189, PubMed: 22737119

Matchin, W., & Hickok, G. (2020). The cortical organization of syn-
tax. Cortex cérébral, 30(3), 1481–1498. https://est ce que je.org/10.1093
/cercor/bhz180, PubMed: 31670779

MathWorks. (2022). MATLAB. https://www.mathworks.com

/products/matlab.html

Möller, J., Jansma, B. M., Rodriguez-Fornells, UN., & Münte, T. F.
(2007). What the brain does before the tongue slips. Cérébral
Cortex, 17(5), 1173–1178. https://doi.org/10.1093/cercor
/bhl028, PubMed: 16831855

Motley, M.. T., Camden, C. T., & Baars, B. J.. (1982). Covert formu-
lation and editing of anomalies in speech production: Evidence
from experimentally elicited slips of the tongue. Journal of Verbal
Learning and Verbal Behavior, 21(5), 578–594. https://est ce que je.org/10
.1016/S0022-5371(82)90791-5

Mumford, J.. UN., & Nichols, T. E. (2008). Power calculation for group
fMRI studies accounting for arbitrary design and temporal auto-
correlation. NeuroImage, 39(1), 261–268. https://est ce que je.org/10
.1016/j.neuroimage.2007.07.061, PubMed: 17919925

Nooteboom, S. G. (2005). Lexical bias revisited: Detecting,
rejecting and repairing speech errors in inner speech. Speech
Communication, 47(1–2), 43–58. https://est ce que je.org/10.1016/j
.specom.2005.02.003

Nozari, N., & Dell, G. S. (2009). More on lexical bias: How effi-
cient can a “lexical editor” be? Journal of Memory and Language,
60(2), 291–307. https://doi.org/10.1016/j.jml.2008.09.006,
PubMed: 20126302

Nozari, N., Dell, G. S., & Schwartz, M.. F. (2011). Is comprehension
necessary for error detection? A conflict-based account of monitor-
ing in speech production. Psychologie Cognitive, 63(1), 1–33. https://
doi.org/10.1016/j.cogpsych.2011.05.001, PubMed: 21652015
Okada, K., Matchin, W., & Hickok, G. (2018). Neural evidence for
predictive coding in auditory cortex during speech production.
Psychonomic Bulletin & Review, 25(1), 423–430. https://doi.org
/10.3758/s13423-017-1284-x, PubMed: 28397076

Oppenheim, G. M., & Dell, G. S. (2008). Inner speech slips exhibit
lexical bias, but not the phonemic similarity effect. Cognition,
106(1), 528–537. https://doi.org/10.1016/j.cognition.2007.02
.006, PubMed: 17407776

Oppenheim, G. M., & Dell, G. S. (2010). Motor movement matters:
The flexible abstractness of inner speech. Mémoire & Cognition,
38(8), 1147–1160. https://doi.org/10.3758/ MC.38.8.1147,
PubMed: 21156877

Pearson, J.. (2019). The human imagination: The cognitive neuro-
science of visual mental imagery. Nature Revues Neurosciences,
20(10), 624–634. https://doi.org/10.1038/s41583-019-0202-9,
PubMed: 31384033

Power, J.. D., Barnes, K. UN., Snyder, UN. Z., Schlaggar, B. L., &
Petersen, S. E. (2012). Spurious but systematic correlations in

Neurobiology of Language

118

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
n
o

je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
8
1
2
0
7
4
4
9
4
n
o
_
un
_
0
0
0
8
8
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Predictive coding and error correction

functional connectivity MRI networks arise from subject motion.
NeuroImage, 59(3), 2142–2154. https://est ce que je.org/10.1016/j
.neuroimage.2011.10.018, PubMed: 22019881

Prix, C. J.. (2012). A review and synthesis of the first 20 years of PET
and fMRI studies of heard speech, spoken language and reading.
NeuroImage, 62(2), 816–847. https://est ce que je.org/10.1016/j
.neuroimage.2012.04.062, PubMed: 22584224

Runnqvist, E., Chanoine, V., Strijkers, K., Patamadilok, C., Bonnard,
M., Nazarian, B., Sein, J., Anton, J.-L., Belin, P., & Alario, F.-X.
(2020). The cerebellum is involved in internal and external
speech error monitoring. bioRxiv. https://doi.org/10.1101/2020
.11.26.400317

Schon, K., Newmark, R.. E., Ross, R.. S., & Arrière, C. E. (2016). A work-
ing memory buffer in parahippocampal regions: Evidence from a
load effect during the delay period. Cortex cérébral, 26(5),
1965–1974. https://doi.org/10.1093/cercor/ bhv013, PubMed:
25662713

Shadmehr, R., & Krakauer, J.. W. (2008). A computational neuroanatomy
for motor control. Experimental Brain Research, 185(3), 359–381.
https://doi.org/10.1007/s00221-008-1280-5, PubMed: 18251019
Siemens Medical Solutions. (2022). 3T Prisma scanner [Apparatus].
https://www.siemens-healthineers.com/en-us/magnetic
-resonance-imaging/3t-mri-scanner/magnetom-prisma

Sormaz, M., Murphy, C., Wang, H. T., Hymers, M., Karapanagiotidis,
T., Poerio, G., Margulies, D. S., Jeffereis, E., Smallwood, J.. (2018).
Default mode network can support the level of detail in experi-
ence during active task states. Proceedings of the National Acad-
emy of Sciences, 115(37), 9318–9323. https://doi.org/10.1073
/pnas.1721259115, PubMed: 30150393

Tanaka, S., & Kirino, E. (2019). Increased functional connectivity of
the angular gyrus during imagined music performance. Frontiers
in Human Neuroscience, 13, 92. https://doi.org/10.3389/fnhum
.2019.00092, PubMed: 30936827

Teghipco, UN. (2022). brainSurfer ( Version 2.0.2) [Computer soft-

ware]. https://doi.org/10.5281/zenodo.7271544

Tian, X., Zarate, J.. M., & Poeppel, D. (2016). Mental imagery of
speech implicates two mechanisms of perceptual reactivation.
Cortex, 77, 1–12. https://doi.org/10.1016/j.cortex.2016.01.002,
PubMed: 26889603

Tibon, R., Fuhrmann, D., Levy, D. UN., Simons, J.. S., & Henson, R.. N.
(2019). Multimodal integration and vividness in the angular gyrus
during episodic encoding and retrieval. Journal des neurosciences,
39(22), 4365–4374. https://doi.org/10.1523/ JNEUROSCI.2102
-18.2018, PubMed: 30902869

Tourville, J.. UN., Nieto-Castañón, UN., Heyne, M., & Guenther, F. H.
(2019). Functional parcellation of the speech production cortex.
Journal of Speech, Language, and Hearing Research, 62(8S),
3055–3070. https://doi.org/10.1044/2019_JSLHR-S-CSMC7-18
-0442, PubMed: 31465713

Tourville, J.. UN., Reilly, K. J., & Guenther, F. H. (2008). Neural mech-
anisms underlying auditory feedback control of speech. Neuro-
Image, 39(3), 1429–1443. https://doi.org/10.1016/j.neuroimage
.2007.09.054, PubMed: 18035557

Velenosi, L. UN., Wu, Oui. H., Schmidt, T. T., & Blankenburg, F. (2020).
Intraparietal sulcus maintains working memory representations of
somatosensory categories in an adaptive, context-dependent
manière. NeuroImage, 221, Article 117146. https://est ce que je.org/10
.1016/j.neuroimage.2020.117146, PubMed: 32659356

Venezia, J.. H., Vaden, K. JE.,

Jr., Rong, F., Maddox, D., Saberi, K., &
Hickok, G. (2017). Auditory, visual and audiovisual speech pro-
cessing streams in superior temporal sulcus. Frontiers in Human
Neurosciences, 11, 174. https://doi.org/10.3389/fnhum.2017
.00174, PubMed: 28439236

Ventura, M.. JE., Nagarajan, S. S., & Houde, J.. F. (2009). Speech target
modulates speaking induced suppression in auditory cortex.
BMC Neuroscience, 10(1), Article 58. https://doi.org/10.1186
/1471-2202-10-58, PubMed: 19523234

Vingerhoets, UN. J., Bylsma, L. M., & de Vlam, C. (2013). Swearing:
A biopsychosocial perspective. Psihologijske Teme, 22(2),
287–304.

Wolpert, D. M., Ghahramani, Z., & Jordan, M.. je. (1995). An inter-
nal model for sensorimotor integration. Science, 269(5232),
1880–1882. https://doi.org/10.1126/science.7569931, PubMed:
7569931

Wu, J., Ngo, G. H., Greve, D., Li, J., Il, T., Fischl, B., Eickoff, S. B.,
& Yeo, B. T. T. (2018). Accurate nonlinear mapping between MNI
volumetric and FreeSurfer surface coordinate systems. Human
Brain Mapping, 39(9), 3793–3808. https://doi.org/10.1002/hbm
.24213, PubMed: 29770530

Lequel, Z., Zhuang, X., Sreenivasan, K., Mishra, V., Cordes, D., &
Alzheimer’s Disease Neuroimaging Initiative. (2019). Robust
motion regression of resting-state data using a convolutional neu-
ral network model. Frontiers in Neuroscience, 13, 169. https://est ce que je
.org/10.3389/fnins.2019.00169, PubMed: 31057348

Yarkoni, T., Poldrack, R.. UN., Nichols, T. E., Van Essen, D. C., &
Wager, T. D. (2011). Large-scale automated synthesis of human
functional neuroimaging data. Nature Methods, 8(8), 665–670.
https://doi.org/10.1038/nmeth.1635, PubMed: 21706013