Abstract
Although much effort has been directed toward understanding the neural basis of speech processing, the neural processes involved in the categorical perception of speech have been relatively less studied, and many questions remain open. In this functional magnetic resonance imaging (fMRI) study, we probed the cortical regions mediating categorical speech perception using an advanced brain-mapping technique, whole-brain multivariate pattern-based analysis (MVPA). Normal healthy human subjects (native English speakers) were scanned while they listened to 10 consonant–vowel syllables along the /ba/–/da/ continuum. Outside of the scanner, individuals' own category boundaries were measured to divide the fMRI data into /ba/ and /da/ conditions per subject. The whole-brain MVPA revealed that Broca's area and the left pre-supplementary motor area evoked distinct neural activity patterns between the two perceptual categories (/ba/ vs /da/). Broca's area was also found when the same analysis was applied to another dataset (Raizada and Poldrack, 2007), which previously yielded the supramarginal gyrus using a univariate adaptation–fMRI paradigm. The consistent MVPA findings from two independent datasets strongly indicate that Broca's area participates in categorical speech perception, with a possible role of translating speech signals into articulatory codes. The difference in results between univariate and multivariate pattern-based analyses of the same data suggest that processes in different cortical areas along the dorsal speech perception stream are distributed on different spatial scales.
Introduction
One of the core aspects of speech perception is mapping complex time-varying acoustic signals into discrete speech units. The “categorical speech perception” phenomenon was experimentally demonstrated in the 1950s, whereby synthetic syllables along the continuum between prototypes (e.g., /ba/ vs /da/) were perceived categorically despite their linear acoustic variations (Liberman et al., 1957). According to one of the most influential speech theories (Liberman and Mattingly, 1985), this perceptual categorization of incoming auditory speech occurs because articulatory gestures serve as the brain's representations of speech sounds, and speech is perceived by mapping continuous auditory signals onto discrete articulatory gestures. More recent theories based on converging evidence from lesion, neuroimaging, and brain stimulation studies (for review, see Devlin and Aydelott, 2009) discount the role of articulatory gestures as perceptual representations, postulating instead that speech percepts and motor codes interact via feedforward and feedback connections (Hickok and Poeppel, 2007; Rauschecker and Scott, 2009).
This auditory–motor integration is thought to be achieved along a dorsal stream speech network, running from primary auditory cortex via posterior superior temporal gyrus (STG) and the inferior parietal lobule to the posterior frontal lobe (Hickok and Poeppel, 2007). The left posterior STG has been implicated in categorical phoneme perception by functional imaging (Husain et al., 2006; Desai et al., 2008; Hutchison et al., 2008) and intracranial human electrophysiology studies (Chang et al., 2010). The supramarginal gyrus (SMG) has also been implicated in categorical speech perception. Raizada and Poldrack (2007) found that the SMG evoked amplified neural activity in response to between-category phoneme pairs along the syllabic continuum of /ba/ and /da/. Furthermore, a recent meta-analysis revealed that the SMG and angular gyrus were consistently activated in fMRI studies of categorical speech perception (Turkeltaub and Coslett, 2010). Thus, posterior nodes of the dorsal speech pathway involved in spectrotemporal analysis of auditory signals, phonological processing, and sensorimotor interface have been clearly implicated in categorical perception of speech. In contrast, findings in frontal articulatory coding areas (posterior inferior frontal gyrus, premotor cortex, anterior insula) have occasionally been reported in studies of categorical speech perception (Myers et al., 2009), but are far less consistent.
We hypothesized that frontal articulation areas are involved in categorical speech perception, but that they may be invisible to subtraction-based fMRI analysis if complex articulatory gestures are represented not by different levels of activity within single voxels, but by differential neural activity patterns within a region of cortex. As such, the present fMRI study used a relatively new mapping strategy, the searchlight analysis (Kriegeskorte et al., 2006), to explore the neuroanatomical basis of categorical speech perception. This technique, based on multivariate pattern-analysis (MVPA) (Haxby et al., 2001; Pereira et al., 2009), probes local information of neural pattern differences across different conditions by moving a searchlight unit on a voxel-by-voxel basis. Some fMRI studies have used MVPA to examine speech perception (Formisano et al., 2008; Okada et al., 2010; Raizada et al., 2010; Kilian-Hütten et al., 2011). Most, however, examined the neural activity patterns exclusively within superior temporal regions. The searchlight procedure allowed us to examine the entire brain.
Materials and Methods
Subjects
Fifteen right-handed native English speakers without reported hearing difficulties or neurological disorders (10 males/5 females; ages, 19–34 years) were recruited from the Dartmouth College community. Two were excluded (one fell asleep during scanning, and one did not have a clear categorical boundary in the behavioral experiment), leaving 13 subjects total. Consent was obtained from all subjects as approved by the Committee for the Protection of Human Subjects at Dartmouth College.
Stimuli
Ten synthesized phonemes (duration of each: 300 ms) along the /ba/–/da/ continuum were created by varying the second and third formant using a SenSyn Klatt Synthesizer (Sensimetrics), as described by Raizada and Poldrack (2007). Each phoneme along the continuum was referred to by number, with 1 being the stimulus of clear /ba/ and 10 being /da/ (Fig. 1).
The average identification curve from the 13 participants. The x-axis shows the 10-step syllabic continuum in numbers between 1 (/ba/ end point) and 10 (/da/ end point). The acoustic distance between successive steps is identical. The y-axis shows the /da/ response frequency percentage at each point; error bars indicate SEs.
Experimental procedures
fMRI experiment.
During scanning, subjects listened to the stimuli while performing a nonphonetic task designed to maintain alertness, in which they were required to indicate a quieter stimulus that was presented in the catch trial block by pressing a button with their right hand.
Behavioral experiment.
After the fMRI experiment, subjects were behaviorally tested to measure a categorical boundary on the 10 syllabic tokens that were previously presented in the scanner. Each syllabic sound was presented eight times, and subjects were required to indicate whether they perceived /ba/ or /da/ by button-press. The boundary was defined as the 50% crossover point between perception of /ba/ and /da/ on each subject's psychometric curve and was used to define binary classes (/ba/ vs /da/) to label neural data per subject.
fMRI scanning
A Philips Intera Achieva 3T scanner was used with a standard EPI BOLD pulse sequence and a clustered volume acquisition. The parameters were as follows: repetition time (TR), 3 s; acquisition time, 1.88 s; silent gap, 1.12 s; 560 ms interval between stimuli and scanner-noise onset/offset; 32 slices; 3 × 3 mm within plane resolution, 4 mm thick slices with a 0.5 mm skip; interleaved slice-ordering. Each stimulus was presented in the middle of the silent gap before the next onset of EPI acquisition. In the scanner, sounds were played via high-fidelity MR compatible headphones (MR Confon). The fMRI scans were subdivided into five runs with 185 volume acquisitions per run. A block design was used, with one of the 10 phonemes repeatedly presented five times in each block. Between the blocks were rest periods lasting 15 s (5 TRs). The order of the blocks was pseudorandom and counterbalanced across subjects. There were 18 stimulus blocks per run, making 90 blocks total across the five runs. Of these, 10 contained quieter catch-trials that were excluded from subsequent analysis. The remaining 80 blocks consisted of eight blocks for each of the 10 stimuli along the /ba/–/da/ continuum.
fMRI data (searchlight procedure) analysis methods
fMRI data were preprocessed using SPM5 and MATLAB 2009b. All images were realigned to the first EPI and spatially normalized into Montreal Neurological Institute (MNI) standard stereotactic space (ICBM152 EPI template) with preserved original voxel size. After preprocessing, fMRI time courses of all voxels were extracted from unsmoothed images. Subsequently, these raw signals were high-pass filtered with a 300 s cutoff to remove slow drifts caused by the scanner, and standardized across entire runs using the z-score to normalize intensity differences among runs. When the data were subsequently entered into a classifier in pattern-based analysis, the TRs corresponding to each syllabic stimulus were labeled as /ba/ or /da/ by using the stimulus blocks' presentation times. For each stimulus-type, the time courses of its stimulus-presentation blocks were convolved with a canonical hemodynamic response function (HRF), thereby producing a regressor for that class of stimuli. The mean value across all time points of the regressor was then calculated. A time point was assigned as belonging to the given stimulus class if the value of the regressor at that time point was greater than this overall mean value.
For the primary searchlight analysis, local searchlight spheres consisting of a center voxel and its neighborhood within a three-voxel radius were constructed throughout the entire brain (∼30,000 U). Then, in each searchlight sphere, a binary classification was performed using the Gaussian Naive Bayes classifier (Pereira et al., 2009). The classifier was initially trained by a subset of datasets (training set) to build a model that appropriately set the boundary between the neural activities that were associated with a perceptual category of /ba/ and /da/. Then, this model was applied to the remaining datasets (testing set) for validation, wherein the accuracy of classification test was gained by calculating how many times the classifier correctly predicted time points corresponding either to /ba/ or to /da/ in the unseen dataset (chance level, 50%). To avoid a bias caused by a particular training/testing set, this procedure was repeated five times, such that signals from four scanning runs were served as a training set and one remaining run served as a testing set (i.e., fivefold cross-validation). The classification accuracy for each searchlight sphere was averaged across the testing set results of the five training/testing combinations and stored in each voxel of an output image for each subject (for more details, see Kriegeskorte et al., 2006).
After output images were acquired by searchlight analysis, baseline corrections in voxel values were made via two steps. First, the chance-level accuracy, i.e., 0.5, was subtracted from an accuracy stored in each voxel of the output images. Next, the mean of these accuracy values in the output image was subtracted from each voxel's value, such that the resulting output image for each subject had a mean of zero. This baseline-correction procedure has two advantages. First, it guards against false discovery (i.e., more conservative than the standard procedure with no baseline correction). Second, it helps to normalize different accuracy ranges across subjects (i.e., reducing the variance in the signal at a group level, which in turn increases the t-statistical power at the group level). The adjusted accuracy map was submitted to the random-effects analysis and threshold of p (cluster-wise corrected) <0.05 in combination with p (voxelwise uncorrected) <0.001 was applied throughout the searchlight analyses. For the random effects analyses on both current and old datasets, SPM8 was used instead to take advantage of its newly added feature, cluster-level correction based on false discovery rate in the updated version of SPM.
A new analysis, using a searchlight approach, of the Raizada and Poldrack 2007 dataset
Procedures were identical to those described above (fMRI data (searchlight procedure) analysis methods), except that the smoothed data of 12 subjects were used because unsmoothed data were not available (for details, see Raizada and Poldrack, 2007). In this 2007 study, stimuli were presented in pairs of two types using event-related design: identical pairs, where one token on the 10-step /ba/–/da/ continuum was presented twice in succession (e.g., 4-then-4); and three-step pairs, in which the two stimuli were three steps apart along the continuum (e.g., 4-then-7). In the present analysis, only the identical pairs were analyzed, as these are directly comparable to the uniform blocks of stimuli that were used in the newly acquired data.
Conventional univariate GLM analysis
For comparison purposes, we also performed standard univariate GLM analyses of the data, investigating the following four contrasts: /ba/ − /da/, /da/ − /ba/, and, as simple controls, all-speech − baseline and baseline − all-speech. These GLM analyses were applied to the smoothed images of the current dataset for each subject using SPM5. At the first-level analysis, the fMRI time courses corresponding to each block of stimulus were convolved with canonical HRF using the box-car design. Additionally, the six motion parameters, as well as the onset of button responses, were modeled; they were later regressed out as effects of noninterest. Contrast images from each subject's first-level analysis were submitted to the second-level random effects for group analysis.
Results
Behavioral results
As expected, nearly all subjects demonstrated a sharp category boundary near the middle of the syllabic continuum between /ba/ and /da/ (Fig. 1). However, one subject did not perceive the syllabic continua in a categorical manner (i.e., linear trend) and was later discarded from further analysis. This subject was a native English speaker, but spent several years during childhood in Hong Kong, which may have affected the behavioral results. The category boundary varied slightly among the subjects (Fig. 1, error bars).
fMRI
MVPA
The whole-brain searchlight analysis equipped with three-voxel radius spheres revealed two left frontal regions, namely, the pre-SMA and the pars opercularis within Broca's area (Figs. 2, 3a; Table 1). No cortical regions within the temporal or parietal lobes were significant. To further validate these results, we reanalyzed previously published adaptation fMRI data (Raizada and Poldrack, 2007) using the same MVPA procedures. Despite several differences between these two datasets (design, number of available trials, degree of smoothness, participating subjects), the left pars opercularis again was significant at a matched threshold (Fig. 3c, Table 1); no other significant areas were identified. Note that the previous dataset implicated the left SMG in categorical phoneme perception using the adaptation fMRI paradigm (Fig. 3d, Table 1). However, this area did not emerge when the MVPA procedure was used on either dataset (t12 = 1.3, t11 = 0, p = n.s; for new and old datasets, respectively).
Broca's area (MNI x, y, z = −54, 12, 22) and pre-SMA (MNI x, y, z = −3, 18, 50) clusters were overlaid on the surface-inflated rendering brain using SPM surfrend toolbox (http://spmsurfrend.sourceforge.net/). R, right; L, left.
Top, Results from new dataset. a, b, Broca's area (a, MNI x, y, z = −54, 12, 22) and left STG (b) identified by the searchlight analyses equipped with three- and eight-voxel radii of local spheres, respectively. c, d, Results from Raizada and Poldrack (2007). c, Broca's area (MNI x, y, z = −45, 9, 27) identified by the searchlight analysis with a three-voxel radius of local sphere. d, The left SMG (MNI x, y, z = −60, −24,18) previously identified by adaptation analysis (Raizada and Poldrack, 2007). The color bars indicate t-vales of the group random-effects analysis. P, posterior; A, anterior; L, left; R, right.
Cortical loci involved in categorical speech perception
To confirm that the activity identified in the MVPA analyses related to individual subjects' percepts rather than acoustic differences between items at each end of the continuum, we performed another searchlight analysis with categories binarized based on the absolute midpoint along the continuum, disregarding each subject's category boundary. As a result, no significant areas were found even at a relaxed threshold [p (voxelwise uncorrected) < 0.01].
The difference between the searchlight analysis results and the previous univariate results could be interpreted as a difference in the spatial scale of processing in different areas of the network (intravoxel in the SMG vs intervoxel in the pars opercularis). This theory raised the possibility that processing in other areas of the speech perception network might be distributed over regions of cortex too large to be detected using a three-voxel searchlight. To explore this possibility, we performed additional searchlights analyses with a systematic manipulation of the searchlight size parameter. At a radius of eight-voxels, a significant cluster emerged within the left STG in the new dataset at a matched threshold (Fig. 3d, Table 1). Interestingly, the pars opercularis and pre-SMA were no longer significant in that larger searchlight analysis. Nevertheless, the old dataset did not yield significant temporal clusters in any of the additional analyses.
Conventional univariate analysis
The comparison of all sounds versus resting blocks yielded significant clusters within the expected cortical regions along the speech network, including bilateral temporal lobes, left parietal region, and frontal regions (data not shown). In contrast, a complementary univariate comparison of items perceived by each subject as /ba/ versus /da/ revealed no significant areas even at a relaxed threshold [p (voxelwise uncorrected) < 0.01)], confirming that MVPA clusters did not represent a simple difference in voxelwise activation intensity.
Discussion
Although the neural basis of speech perception has been extensively studied in the auditory neuroimaging field, relatively little evidence has been garnered as to how speech sounds are sorted into discrete categories (Turkeltaub and Coslett, 2010). Previous evidence, based largely on standard univariate fMRI analysis, has mainly suggested roles for posterior temporal and inferior parietal regions in categorical speech perception (Celsis et al., 1999; Zevin and McCandliss, 2005; Husain et al., 2006; Raizada and Poldrack, 2007; Desai et al., 2008; Hutchison et al., 2008; Chang et al., 2010; Zevin et al., 2010; Kilian-Hütten et al., 2011). Nevertheless, theories implicating sensorimotor mapping in speech sound categorization (Devlin and Aydelott, 2009) suggest frontal regions should be involved as well. In line with this hypothesis, the present MVPA study revealed that two frontal regions, namely, Broca's area and left pre-SMA, evoked neural activities related to perceptual category of syllabic stimuli. The same Broca's area result was observed when the identical MVPA procedure was applied to another dataset that had previously implicated the SMG as an important hub of categorical speech processing (Raizada and Poldrack, 2007). The frontal network has been typically viewed as a region involved in speech production and so our findings naturally invite consideration of auditory–motor integration in speech perception.
Possible roles of Broca's area in speech perception
Converging evidence from multiple sources has demonstrated that there exists an anterior–posterior distinction within Broca's area, where anterior areas play a role in semantic processes and posterior areas are involved in phonetic and phonological processes (Poldrack et al., 1999; Gold et al., 2005; Gough et al., 2005). Lesion data as well as functional imaging results have suggested a role for pars opercularis (the posterior-most region of Broca's area) in sublexical phonetic processing during speech production (Carreiras et al., 2006; Riecker et al., 2008; Papoutsi et al., 2009). The pars opercularis was the significant region of Broca's area in the present study as well. Thus, it is possible that the same phonetic processes or representations that operate in Broca's area during speech production or decision tasks also participate in sublexical speech perception.
Another possible role for posterior Broca's area in speech perception relates to its role in motor aspects of speech, specifically articulatory planning. Recently, evidence from neuroimaging (Wilson et al., 2004) and transcranial magnetic stimulation (Watkins et al., 2003) studies has suggested that motor systems play a role in speech perception (Pulvermüller et al., 2006; Devlin and Aydelott, 2009). These findings have been also viewed as evidence supporting putative mirror neuron systems in the human brain (D'Ausilio et al., 2009). Relatedly, Skipper et al. (2007) presented evidence that visual contexts (e.g., viewing mouth movement) aid speech perception by recruiting motor systems. Consistent with Skipper et al. (2007), the present study found that a frontal motor–speech network consisting of Broca's area and the left pre-SMA evoked distinct neural patterns corresponding to the perceived syllables. One plausible scenario is that Broca's area generated articulatory representations to guide categorization of the phonetic continua into discrete syllabic units via feedback mechanisms, as proposed in various current speech models (Callan et al., 2004; Davis and Johnsrude, 2007; Rauschecker and Scott, 2009). This articulatory guidance was presumably subconscious in these experiments, as the fMRI task was orthogonal to syllabic identity and subjects were not aware of the behavioral categorization task until afterward.
We also identified pre-SMA activity related to syllable identity. The pre-SMA, which rarely if ever activates in univariate imaging studies of speech perception, has been viewed as a high-level action-control region mediating sensory–motor transition (Nachev et al., 2008), and has recently received attention in speech production (Alario et al., 2006). Lesions in pre-SMA cause deficits in production of planned speech (Ziegler et al., 1997), suggesting that it plays a role in the preparation of sequential movements, especially in conditions involving a short-term buffering of motor–speech acts. Thus, the pre-SMA finding here further supports involvement of articulatory motor processes in speech perception (Pulvermüller et al., 2006). We surmise that articulatory codes and sequences accessed in Broca's area and nearby premotor cortex may be transmitted to the pre-SMA for buffering during speech production or modeling to guide speech perception. In particular, the frontal motor system may play a role when identifying ambiguous or degraded speech signals. For example, a neuroimaging study revealed that, during an /r/ versus /l/ consonant identification task, Broca's area was more activated in Japanese speakers whose mother tongue did not differentiate the two consonant sounds than in native English speakers (Callan et al., 2004). Motor systems may also contribute to perception in adverse listening circumstances, as evidenced by the specific perceptual difficulty patients with Broca's aphasia experience when speech signals are degraded (Utman et al., 2001; Moineau et al., 2005). The current findings lend further support to the theory that auditory and motor systems interact during speech perception.
Different spatial scales of processing along the dorsal speech perception stream
Intriguingly, the MVPA and univariate adaptation analysis of the same dataset (Raizada and Poldrack, 2007) yielded completely different results, the former identifying Broca's area, and the latter identifying SMG. For this new analysis, only the within-category pairs were chosen for binary classification, whereas in the previous study, the neural activities of between-category pairs were contrasted with those of within-category pairs. Thus, it is important to underscore that the analyses were not performed for identical comparisons, although both were designed to isolate the same perceptual processes. Recently Zevin and colleagues (2010) demonstrated that the SMG is involved in domain-general change detection, and is thus commonly active during adaptation paradigms regardless of speech content. Hence, the discrepancy between the univariate and multivariate results may reflect different roles for SMG (change detection or discrimination) and Broca's area (phoneme identification) in speech perception.
However, in the visual domain, simultaneous application of a univariate adaptation paradigm (Grill-Spector and Malach, 2001) and a multivariate analysis on the identical data (Drucker and Aguirre, 2009) has also revealed differences in localization of activity, as in our study. This difference may have occurred because local networks in different areas of the brain operate at different spatial scales, reflecting different underlying computational properties. Different types of fMRI analysis reveal different spatial scales of neural processing. Adaptation-fMRI distinguishes between differently tuned populations of neurons that are intermingled within the same voxel, and so highlights processing at a subvoxel spatial scale. In contrast, MVPA reveals areas where information is represented by distributed patterns of activity, thereby highlighting processing at a multivoxel spatial scale. Thus, the findings here may indicate that Broca's area and the SMG use different spatial scales (intervoxel vs intravoxel, respectively) to process speech inputs. Recently, MVPA studies have shown distributed neural patterns for speech perception within large expanses of the temporal lobes (Formisano et al., 2008; Kilian-Hütten et al., 2011). One possible reason for the discrepancy between our main findings and those of previous studies is that auditory representations of phonemes are coded by patterns of activity distributed across large expanses of temporal cortex, while frontal representations are coded in patterns across smaller expanses. Thus, the relatively small three-voxel searchlight might be spatially tuned to detect frontal representations, but too small to detect temporal representations. To test this theory, we performed additional searchlight analyses with larger sizes of local spheres. At a radius of eight voxels, the pars opercularis cluster was no longer significant, but the searchlight revealed a significant cluster in the left posterior STG in the new dataset. This finding supports the theory that frontal and temporal speech representations are distributed at different spatial scales. Because of the large size of the searchlights in this analysis (∼21 mm radius), searchlights contributing to this STG cluster included voxels in the mid-to-posterior STG and middle temporal gyrus, and portions of the SMG. This could be taken as a caveat to our claim that SMG speech processing occurs at a small spatial scale. However, much of the STG cluster was too far from the SMG to have included it in the searchlights. The failure to replicate this finding in the old dataset should also be taken as a caveat, but was likely because unsmoothed images were no longer available and only subsets of the trials in the old dataset were used. However, future studies should also consider alternative explanations for differential spatial scale along the frontotemporal speech network, such as differing hemodynamic response profiles across different regions (Handwerker et al., 2004). Because MVPA are sensitive to coherent patterns of information-carrying activity rather than to univariate intensity differences, it is unlikely that differing hemodynamic response profiles across different regions could account for the current findings. Nonetheless, it would be interesting for future studies to investigate these questions further.
In summary, our results clearly demonstrate the presence of phoneme category-specific information in pars operculars within Broca's area. Further research is needed to clarify the specific role of this area in speech perception. Our findings also raise interesting questions regarding differences in the spatial scale of cortical representations or processes along speech-perception pathways. This principle likely applies to other perceptual domains as well. The possibility that different regions of the cortex encode information at different scales of spatial distribution, or even that a single region may encode different information at different spatial scales, should be considered and examined further in future studies.
Footnotes
This work was supported by grants from the Office of Naval Research. We thank Yale Cohen, Tilbe Goksun, David Kraemer, and Kate Onsgard for their helpful comments on the earlier version of the manuscript. Special thanks go to the Penn Postdoc Editors Association for proofreading.
The authors declare no financial conflicts of interest.
- Correspondence should be addressed to Dr. Yune-Sang Lee, Center for Cognitive Neuroscience, 3710 Goddard Laboratory, Hamilton Walk, University of Pennsylvania, Philadelphia, PA 19104. yslee{at}mail.med.upenn.edu
References
- Alario et al., 2006.↵
- Callan et al., 2004.↵
- Carreiras et al., 2006.↵
- Celsis et al., 1999.↵
- Chang et al., 2010.↵
- D'Ausilio et al., 2009.↵
- Davis and Johnsrude, 2007.↵
- Desai et al., 2008.↵
- Devlin and Aydelott, 2009.↵
- Drucker and Aguirre, 2009.↵
- Formisano et al., 2008.↵
- Gold et al., 2005.↵
- Gough et al., 2005.↵
- Grill-Spector and Malach, 2001.↵
- Handwerker et al., 2004.↵
- Haxby et al., 2001.↵
- Hickok and Poeppel, 2007.↵
- Husain et al., 2006.↵
- Hutchison et al., 2008.↵
- Kilian-Hütten et al., 2011.↵
- Kriegeskorte et al., 2006.↵
- Liberman and Mattingly, 1985.↵
- Liberman et al., 1957.↵
- Moineau et al., 2005.↵
- Myers et al., 2009.↵
- Nachev et al., 2008.↵
- Okada et al., 2010.↵
- Papoutsi et al., 2009.↵
- Pereira et al., 2009.↵
- Poldrack et al., 1999.↵
- Pulvermüller et al., 2006.↵
- Raizada and Poldrack, 2007.↵
- Raizada et al., 2010.↵
- Rauschecker and Scott, 2009.↵
- Riecker et al., 2008.↵
- Skipper et al., 2007.↵
- Turkeltaub and Coslett, 2010.↵
- Utman et al., 2001.↵
- Watkins et al., 2003.↵
- Wilson et al., 2004.↵
- Zevin and McCandliss, 2005.↵
- Zevin et al., 2010.↵
- Ziegler et al., 1997.↵
