Introduction

Auditory phoneme identification, segmentation and discrimination during speech processing requires explicit access to certain sublexical speech segments, in contrast to speech processing in the context of more holistic speech recognition and comprehension, and is supported by a dorsal processing stream running from posterior superior temporal gyrus (STG) via inferior parietal regions towards frontal regions [1–3]. Initial, primary acoustic–phonetic analysis is supported by the middle and posterior part of the STG and superior temporal sulcus (STS) bilaterally, although hemispheric asymmetries do occur [1, 2, 4–8]. Left superior temporal areas ought to be specialized for processing temporal changes, whereas the right homologue areas should focus on spectral analyses [9]. Continuing with the time domain, the left hemisphere processes shorter timescales (25–50 ms; e.g. extraction of segment boundaries) and the right hemisphere longer timescales (200–300 ms; e.g. sentential prosody) [10]. From then on, activation spreads from the posterior STG via the inferior parietal cortex to the motor cortex and inferior frontal gyrus (BA 44/45) along the superior longitudinal and arcuate fascicle [2, 3, 7, 11, 12]. Within the dorsal pathway, the inferior parietal cortex serves as an auditory-motor interface between auditory features and articulatory-based representations in achieving successful detection of phonological changes [13–15]. Several studies have shown that the premotor and primary motor regions, usually active during speech production, are differentially activated in a somatotopic manner, based on articulatory characteristics of the phonemic contrasts, during phoneme perception and discrimination processes (e.g. activation of lip area during perception of bilabial place of articulation) [16–20]. Moreover, inferior frontal regions have been implicated in speech recognition as well [21, 22], when a new, non-familiar word or a pseudoword is encountered. Such word stimuli require more intensive segmentation processes and articulatory-based representations of the different segments (i.e. phonemes), hence sublexical processing, which involves more participation of frontal and motor areas within the dorsal circuitry [11, 23–25].

Importantly, by measuring event-related potentials (ERPs) and estimating the source of this event-related electrical activity through non-invasive source reconstruction, some interesting findings have been revealed regarding the spatiotemporal neural activity during phoneme perception. Already 100 ms after phoneme presentation, auditory association cortices (STG) are supposed to be spatially mapped along an anterior-posterior dimension determined by mutually exclusive place of articulation (PoA) features, as derived from the N100m (magnetic counterpart of the N100) and Mismatch Field (magnetic counterpart of the Mismatch Negativity) [26–29]. Moreover, even earlier, around 50 ms after phoneme presentation, neuronal generators of the P50 m are topographically aligned in more primary auditory regions, such as Heschl’s gyrus, along a medial–lateral dimension according to PoA features [30]. Generally, ERPs like the P50 and N100 potential represent earlier, sensory-perceptual processes of phoneme detection and identification [30, 31], while ERPs like the (pre-attentive) Mismatch Negativity (MMN; elicited during passive oddball paradigm) and (attentive) P300 potential (elicited during active oddball paradigm) are more associated with phoneme discrimination processes [32, 33]. However, by comparing the neuronal generators of the MMN in the conventional design (distraction by video) and more attention-demanding condition (distraction by video, yet response to target requested), increased attention to the auditory stimuli was linked to greater participation of motor areas and inferior frontal regions during later stages (200–400 ms) of the discrimination processes [34].

Clearly, phonemic contrasts play a significant role in terms of neural activity in auditory and motor regions during phoneme perception and discrimination. Especially for PoA an important interaction has been demonstrated between the spatial mapping and temporal organization of neural activity associated with PoA features [27, 30]. For instance, source reconstruction of the N100m showed that stop-consonants with a labial PoA are processed more anterior on the superior temporal horizontal axis, whereas stop-consonants with a dorsal PoA are processed more posterior on the superior temporal horizontal axis [27]. However, it is remarkable that phonemes have only been compared at opposite sides of one phonemic contrast continuum (e.g. /d/ and /g/ for PoA) with respect to their somatotopic organization in auditory and motor cortices, whereby the focus mainly was on the PoA features [16, 30]. This can be related to the more unambiguous, invariable articulatory gestures inherent to PoA features for which potentially more clear-cut spatial representations exist in the brain [35]. Such motor properties are less well distinguishable for phonemic contrasts like voicing (e.g. /d/ and /t/) and manner of articulation (MoA) (e.g. /d/ and /z/), for which perception and discrimination processes rely more on well-defined acoustic cues [36]. On the contrary, acoustic cues of the PoA contrast are highly variable depending on the phonetic context, whereby invariant acoustic features are missing.

The present study aims to investigate on the one hand whether phoneme discrimination based on PoA elicits more motor cortical activity, considering its limited invariant acoustic cues and clear-cut motor representations, and on the other hand whether phoneme discrimination based on voicing or MoA evokes more auditory cortical activity, considering their well-defined acoustic features and unclear motor representations. Because of the important temporal aspect of neural activity during speech processing [30, 37, 38], the current study implements neurophysiological measures (MMN, P300) in response to an auditory oddball paradigm on which source reconstruction is performed. As such, it is examined whether a spatial cortical pattern associated with a particular phonemic contrast occurs with a certain temporal differentiation compared to the other phonemic contrasts. A second part of the study was to inspect whether increased attention can be related to higher activation of frontal areas [34], when the passively elicited MMN (conventional design) and actively elicited P300 are compared. These hypotheses are set to be evaluated (in a later stage of this research project) in a group of persons with acute aphasia (within 3 months post-stroke), detect potential aberrant activation patterns (hence, inspect the integrity of the dorsal processing stream in acute stroke) and appraise possible clinical implementation of the proposed source reconstruction method.

Methods

Subjects

Forty-seven subjects (mean age 47.64 years ± 13.81), who were mainly recruited from hospital staff and senior club houses, participated in the study (24 female). All persons investigated were right-handed, as verified with the Dutch Handedness Inventory (DHI [39]). All the participants had Dutch as native language and reported to have normal hearing. None of them had neurological, psychiatric or speech and language developmental disorders, as verified by a questionnaire. At time of testing, none of the participants was on medication. The study was approved by the Ethics Committee of the University Hospital Ghent (Belgium) and an informed consent was obtained from all the subjects.

Paradigm and stimuli

An auditory phoneme discrimination (APD) task was used and consisted of three different auditory oddball paradigms both in a pre-attentive, passive and attentive, active condition. During the passive APD tasks, the subjects were instructed to ignore the stimuli and to focus their attention to a silent movie. To ensure that participants did not specifically pay attention towards the stimuli during the passive paradigm, they were asked questions about the movie afterwards, which they were informed about before the experiment started. During the active APD tasks, the subjects had to push a button every time they heard the infrequent stimulus. The sequence of the individual tasks was counterbalanced across subjects, meaning that every subject started with another task. Within the passive and active APD tasks, each block consisted of a total of 250 stimuli and 150 stimuli, respectively. The standard phoneme was [bə] and the deviant phonemes were [gə] (covering a PoA difference), [pə] (covering a voicing difference) and [mə] (covering a MoA difference) (Fig. 1). A summary of the acoustic–phonetic composition of every phoneme (analyzed with PRAAT [40]) is provided in Table 1. The APD tasks were created in such a way that the standard and deviant stimulus only differed in one phonemic contrast. Stimuli were generated using the Dutch website NeXTeNS (http://nextens.uvt.nl/demo.html) where text was converted to speech [41]. In all stimulus blocks, the standard and deviant phoneme appeared with a probability of 0.80 and 0.20, respectively. The stimuli were presented in a random order in which two deviants could not follow each other without having a standard in between. All the spoken phonemes had a duration of 150 ms. The interstimulus interval (ISI) was set at 500 ms in the passive APD tasks and 2000 ms in the active APD tasks.

Fig. 1
figure 1

Phoneme stimuli and their spectrograms. The left-side column represents the speech waveforms of every phoneme in the time domain (150 ms duration). The right-side column represents their spectrograms in the frequency domain, with time on the x axis and frequency (kHz) on the y axis

Full size image
Table 1 Detailed acoustic–phonetic composition of the phoneme stimuli
Full size table

Electroencephalogram (EEG) recording and analysis

The EEG (0.5–100 Hz band-pass, notch filter disabled) was recorded through 23 Ag/AgCl-electrodes using a linked earlobes reference and an electrode placed on the forehead as ground. Electrodes were placed on the scalp according to the international 10–20 system. The impedance of the electrodes was kept below 5 kΩ. Data was collected using a SynAmp (Neuroscan) amplifier and was continuously digitized at a 500 Hz sampling rate.

EEG analysis was performed using BrainVision Analyzer 2 (Brain Products, Munich, Germany). According to our previous studies (e.g., reference [42]), the following procedure for preprocessing of EEG data was used: the EEG signal was filtered with a 0.5–30 Hz band-pass filter. Because no electrooculogram was applied, independent component analysis (ICA) had to be used to remove artifacts caused by eye movements by excluding two components based on inspection of the components’ spatial distribution (prefrontal locations). These components comprised blinking and horizontal and vertical eye movements (saccades). For the three active APD tasks the EEG was segmented into 1100 ms long epochs from 100 ms pre-stimulus to 1000 ms post-stimulus. For the three passive APD tasks the EEG was segmented into 500 ms long epochs from 100 ms pre-stimulus to 400 ms post-stimulus. The epochs were baseline corrected using a pre-stimulus window of 100 ms. All epochs containing artifacts exceeding 100 µV were rejected from further analysis, as were false responses in the active APD tasks. The data was converted to an average reference for further source reconstruction analysis. Standard and deviant trials were averaged separately. Difference waves were calculated only for the passive APD tasks to elicit the MMN by subtracting the average ERP of standard trials from the average ERP of deviant trials. Peak detection was carried out semi-automatically to measure peak latencies of the MMN in the passive APD tasks and peak latencies of the N100 and P300 in the active APD tasks. For the active APD tasks, reaction times and percentage of correct responses were calculated as well.

Source reconstruction and statistical analyses

All analyses were performed in the Statistical Parametric Mapping 8 software package (SPM 8: Welcome Department of Cognitive Neurology, University College, London, United Kingdom) implemented in MATLAB (the MathWorks, Inc., Massachusetts, USA).

First, a sensor-space analysis was executed to search for time points with significant differences in activation between phonemic contrasts PoA, voicing and MoA and the passive and active APD tasks. In the sensor-space analysis, the ERP data for each trial and each condition were converted into 3D images. These were generated by constructing 2D, 64 × 64 voxels resolution, scalp maps for each time point (using interpolation to estimate the activation between the electrodes) and by stacking the scalp maps over peristimulus time, resulting in [64 × 64 × number of time points]-images [43]. Using these images, statistical analyses were performed for each condition. Paired t tests were used to calculate two-tailed, two-sided (both activations AND deactivations) F-contrasts for the main effect of phonemic contrast (PoA vs voicing; voicing vs MoA; PoA vs MoA). The resulting statistical parametric map (SPM) was family-wise error (FWE) corrected for multiple comparisons using Random Field Theory [43–45], which takes into account the spatial correlation across voxels (i.e., that the tests are not independent). The sensor analysis was performed to search for maximal differences between conditions, so time windows were not chosen based on the peaks of the evoked waveforms for statistical analysis at the source level [46].

In a next step, source reconstruction was executed at the significant time points from the sensor analysis. The default 3-layered scalp-skull-brain template head model (the MNI brain) was used based on the Colin27 template [47], implemented in the SPM software. The default electrode positions were transformed to match the template head model, in which 8196 dipoles are assumed on a template cortical surface mesh. During the reconstruction, the coordinate system in which the electrode positions were originally represented was coregistered with the coordinate system of the template head model (i.e., MNI coordinates), again using the default electrode positions. In a next step, the forward model was calculated for each dipole on the cortical mesh, based on assumptions about the physical properties of the head, using the “bemcp” method (BEM) implemented in FieldTrip [48]. Finally, the actual inverse reconstruction was performed at group level for every condition over the whole ERP time window and was based on an empirical approach using the multiple sparse priors (MSP) algorithm [49]. Based on these reconstructions, 3D images were generated containing the evoked energy of the reconstructed activity in time windows corresponding with the significant time points from the sensor analysis. Using these images, additional second level analyses were performed to identify the most significant areas over subjects. Paired t tests were used to calculate one-tailed, one sided (activations OR deactivations) T-contrasts for the main effect of phonemic contrast (PoA vs voicing; voicing vs MoA; PoA vs MoA). Having corrected for multiple comparisons at the sensor-level, an uncorrected p value was set at 0.05 at the source level [46]. The resulting MNI coordinates holding significant activation differences between conditions, were explored by means of the SPM Anatomy toolbox developed by Eickhoff et al. [50] and were grouped to the following regions of interest: inferior frontal cortex (IFC), sensorimotor cortex (SMC), inferior parietal cortex (IPC) and superior temporal cortex (STC). It has been established in previous research that low-density recordings provide an accurate estimate of the underlying ERP generators, are sufficient to fully describe the variance of a typical ERP data set collected during an auditory oddball paradigm, when compared to high-density recordings, [51] and has been applied in neurolinguistic research before [52].

Additional statistical analyses were performed using IBM SPSS Statistics 22 on peak latencies in the passive and active APD tasks (MMN and N100/P300, respectively). For this, six clusters with the average latency of two electrodes were created, keeping midline electrodes separate: Anterior Left (F3, F7), Anterior Midline (Fz), Anterior Right (F4, F8), Central Left (T3, C3), Central Midline (Cz), Central Right (T4, C4), Posterior Left, (T5, P3) Posterior Midline (Pz), Posterior Right (T6, P4). Repeated measures ANOVA was carried out with factor contrasts (PoA vs. voicing vs. MoA), region (anterior vs. central vs. posterior) and laterality (left vs. midline vs. right). Reaction time was statistically analyzed as well, using repeated measures ANOVA with the factor contrasts (PoA vs. voicing vs. MoA). Greenhouse-Geisser correction (GG) was applied when the assumption of sphericity was violated. Significance level was set at ≤0.05. Post hoc pairwise comparisons were computed using Bonferroni correction.

Results

ERP waveforms and behavioral results

A MMN was elicited in the passive APD task for all three phonemic contrasts, but with differences in morphology (Fig. 2). The MMN was significantly different from zero for all three phonemic contrasts, on left and right anterior, central and posterior locations [T(46) = −9.64/−17.96, p < 0.001]. In the active APD task an N100 and P300 was elicited with all three phonemic contrasts, but without clear differences in morphology (Fig. 3). With PoA as phonemic contrast, 97.65% of the deviant stimuli were identified, with voicing 96.8% and with MoA 98.34%. Subjects showed a mean response reaction time of 565.72 ms (±80.32) with PoA, 561.35 ms (±94.52) with voicing and 535.44 ms (±85.63) with MoA. There was a significant main effect for the phonemic contrasts [F(2,90) = 5.78, p = 0.004]. Faster response times occurred for the phonemic contrast MoA compared to PoA (p = 0.001) and voicing (p = 0.047).

Fig. 2
figure 2

Grand average ERPs for the passive APD task. In the upper panel, grand averaged ERPs are depicted for C3, Cz and C4 showing the response to the standard and deviant stimulus and the difference waveform (black standard; red deviant; blue difference waveform). The MMN is indicated in translucent red. In the lower panel, a detailed transcription is displayed at the Cz electrode with an overlay of the difference waves in response to the three phonemic contrasts (black voicing; red PoA; blue MoA). The time window (50–100 ms) which contains significant differences at source level during source reconstruction is indicated in translucent blue. Negative is plotted upwards

Full size image
Fig. 3
figure 3

Grand average ERPs for the active APD task. Grand averaged ERPs are depicted for P3, Pz and P4 with an overlay of the three phonemic contrasts (black voicing; red PoA; blue MoA) showing the response to the deviant stimulus. Three time windows (65–110 ms; 130–175 ms; 225–250 ms) containing significant differences at source level during source reconstruction are indicated in translucent blue. The last time window (310–340 ms) did not contain significantly different clusters at source level and is indicated in translucent red. Negative is plotted upwards

Full size image

Source reconstruction and ERP analyses

Passive APD task

Between 50 and 100 ms after stimulus presentation, the sensor-space analysis found significant differences between PoA and voicing [F(1,46) = 28.43, p < 0.001, FWE-corrected], PoA and MoA [F(1,46) = 85.76, p < 0.001, FWE-corrected] and MoA and voicing [F(1,46) = 54.43, p < 0.001, FWE-corrected].

Second-level source analyses revealed the following pattern in the same time window: higher activation for MoA in sensorimotor regions compared to voicing [T(46) = 4.28, p < 0.001] and PoA [T(46) = 4.05, p < 0.001], higher activation for MoA in inferior parietal regions compared to voicing [T(46) = 2.79, p = 0.004] and PoA [T(46) = 3.05, p = 0.002] and higher activation for MoA in superior temporal regions compared to voicing [T(46) = 3.64, p < 0.001] and PoA [T(46) = 3.66, p < 0.001]. No differences in activation were detected in any of the regions of interest when voicing and PoA were compared.

The time window with these activation differences (50–100 ms) corresponded to an epoch before the actual MMN onset and to a clear peak between 50 and 100 ms with MoA as phonemic contrast (Fig. 2). No significant timing differences between the phonemic contrasts were found at the level of MMN peak latency. Results of the source reconstruction can be found in Figs. 4 and 5 and Table 2.

Fig. 4
figure 4

Maximum intensity projections (MIP) of differences between phonemic contrasts during the passive APD task. The maximum intensity is projected on a glass brain in three orthogonal planes; axial (bottom plot), sagittal (top left plot) and coronal (top right plot). In the top plots of each box the time course of the region(s) with statistical differences in maximal activity is represented, with time going from the bottom of the plot (−100 ms) to the top (400 ms). The bottom plot shows the MIP at the time of the maximal activation. The MIP can be seen as an image in which all relevant activity above the α = 0.05 (uncorrected) threshold is superimposed in a certain direction (axial, sagittal, and coronal). The red arrow indicates the global maximum activity. P posterior, A anterior, L left, R right

Full size image
Fig. 5
figure 5

Differences in activity levels between phonemic contrasts during the passive APD task. The y axis represents the amount of reconstructed activity per cluster (x axis), which is a number without unit. On the x axis the clusters of sensorimotor (SMC), inferior parietal (IPC) and superior temporal cortex (STC) are displayed. Between 50 and 100 ms pre-attentive phoneme discrimination based on MoA elicited more activation of sensorimotor, inferior parietal and superior temporal regions compared to voicing and PoA

Full size image
Table 2 Results of the source reconstruction for the comparison between the phonemic contrasts PoA, voicing and MoA during passive, pre-attentive phoneme discrimination (MMN)
Full size table

Active APD task

The sensor-space analysis determined a first time window, between 65 and 110 ms, holding significant differences between PoA and voicing [F(1,46) = 85.39, p < 0.001, FWE-corrected], PoA and MoA [F(1,46) = 112.10, p < 0.001, FWE-corrected] and MoA and voicing [F(1,46) = 41.02, p < 0.001, FWE-corrected]. Source reconstruction showed that MoA elicited higher activation compared to PoA in inferior frontal [T(46) = 2.25, p = 0.015], sensorimotor [T(46) = 4.85, p < 0.001], inferior parietal [T(46) = 3.94, p < 0.001] and superior temporal regions [T(46) = 4.27, p < 0.001] and higher activation compared to voicing in sensorimotor areas [T(46) = 2.84, p = 0.003]. In turn, voicing elicited higher activation in inferior frontal [T(46) = 3.46, p = 0.001], sensorimotor [T(46) = 4.67, p < 0.001], inferior parietal [T(46) = 3.57, p < 0.001] and superior temporal regions [T(46) = 3.67, p < 0.001] compared to PoA.

Further on, in a second time window, between 130 and 175 ms, the sensor-space analysis showed significant differences only between PoA and voicing [F(1,46) = 90.26, p < 0.001, FWE-corrected] and PoA and MoA [F(1,46) = 91.11, p < 0.001, FWE-corrected]. The difference between MoA and voicing disappeared. Moreover, during source reconstruction a reversed pattern emerged, compared to the 65 and 110 ms time-window, showing higher activation for PoA in inferior frontal regions compared to both voicing [T(46) = 3.19, p = 0.001] and MoA [T(46) = 4.23, p < 0.001], sensorimotor regions compared to both voicing [T(46) = 3.79, p < 0.001] and MoA [T(46) = 3.90, p < 0.001] and inferior parietal regions compared to both voicing [T(46) = 3.12, p = 0.002] and MoA [T(46) = 2.94, p = 0.003] with an additional higher activation in superior temporal regions for PoA compared to MoA [T(46) = 2.52, p = 0.008].

The established differences in this and the previous time window occurred around the N100 potential (Fig. 3) and were related to latency differences between phonemic contrasts. A region × contrasts interaction occurred in the statistical analysis on the N100 peak latency [F(4,180) = 15.27, p < 0.001, GG ɛ = 0.68], showing a longer latency for PoA compared to voicing (p = 0.001) and MoA (p < 0.001) and voicing compared to MoA (p < 0.001) in anterior regions, a longer latency for PoA (p < 0.001) and voicing (p < 0.001) compared to MoA in central regions and a longer latency for voicing compared to MoA (p = 0.003) in posterior regions. Because of these latency differences, possible differences in activation in the regions of interest between phonemic contrasts are searched for in different segments of the N100 potential (i.e. onset and peak).

Next, the sensor-space analysis defined a third (225–250 ms) time window which only contained significant differences between PoA and MoA [F(1,46) = 36.18, p < 0.001, FWE-corrected]. Higher activation of sensorimotor [T(46) = 3.34, p = 0.001] and superior temporal regions [T(46) = 2.84, p = 0.003] was elicited for PoA compared to MoA. A last time window (310–340 ms), corresponding to the onset of the P300 wave (Fig. 3), revealed significant differences between PoA and MoA during sensor-space analysis [F(1,46) = 35.23, p < 0.001, FWE-corrected], though statistical analysis on source level could not confirm this. However, from the statistical analysis on the P300 peak latency a significant main effect for contrasts emerged, showing a longer latency for PoA compared to MoA [F(2,90) = 12.15, p < 0.001]. Results of the source reconstruction can be found in Figs. 6 and 7 and Table 3.

Fig. 6
figure 6

Maximum intensity projections (MIP) of differences between phonemic contrasts during the active APD task. The maximum intensity is projected on a glass brain in three orthogonal planes; axial (bottom plot), sagittal (top left plot) and coronal (top right plot). In the top plots of each box the time course of the region(s) with statistical differences in maximal activity is represented, with time going from the bottom of the plot (−100 ms) to the top (900 ms). The bottom plot shows the MIP at the time of the maximal activation. The MIP can be seen as an image in which all relevant activity above the α = 0.05 (uncorrected) threshold is superimposed in a certain direction (axial, sagittal, and coronal). The red arrow indicates the global maximum activity. P posterior, A anterior, L left, R right

Full size image
Fig. 7
figure 7

Differences in activity levels between phonemic contrasts during the active APD task represented in the first two time windows. The y axis represents the amount of reconstructed activity per cluster (x axis), which is a number without unit. On the x axis the inferior frontal (IFC), sensorimotor (SMC), inferior parietal (IPC) and superior temporal cortex (STC) are shown. Between 65 and 110 ms attentive phoneme discrimination based on MoA and voicing requires more activation of inferior frontal, sensorimotor, inferior parietal and superior temporal regions. Between 130 and 175 ms a reversed pattern emerged showing more involvement of the same motor and auditory areas during phoneme discrimination based on PoA

Full size image
Table 3 Results of the source reconstruction for the comparison between the phonemic contrasts PoA, voicing and MoA during active, attentive phoneme discrimination (P300)
Full size table

Discussion

We aimed to determine whether phoneme discrimination based on PoA elicits more motor cortical activity and whether phoneme discrimination based on voicing or MoA evokes more auditory cortical activity. Moreover, we investigated whether potential different cortical activation patterns related to these phonemic contrasts are organized in different temporal integration windows and whether increased attentional load would primarily be related to increased frontal activity. The current data show several time windows with significant differences in activation of auditory and motor regions associated with phoneme discrimination based on the three different phonemic contrasts and attentional load.

Already between 50 and 100 ms, pre-attentive phoneme discrimination with MoA as phonemic contrast elicits more activation in sensorimotor, inferior parietal and auditory superior temporal regions. Further on in time no dissimilarities arose between the different phonemic contrasts. Within the MoA-related difference waveform this time window corresponds to a pronounced peak before the definite MMN onset (Fig. 2). It seems that this pronounced pre-MMN peak represents the need for higher activation of auditory, temporal regions in conjunction with sensorimotor regions during phoneme discrimination based on MoA, which seems to correspond to a dorsal processing stream typically involved in sublexical speech processing [53, 54]. Moreover, considering that this effect occurs before the actual MMN, it is possible that a preparatory phase before the actual phoneme discrimination process is needed, specifically when discrimination is dependent on MoA. So probably, during this preparatory phase a higher reliance on integration of both auditory features and motor based representations of the articulatory movements related to the phonemic contrast MoA is needed [7]. MoA is a phonemic contrast that has been rarely investigated, whereas PoA has been investigated extensively with respect to its representation in both auditory and motor regions [17, 27, 28, 55, 56]. Most of the studies implemented PoA as phonemic contrast most likely because of its clear-cut somatotopic correlates of the articulatory movements in the motor cortex (i.e. the lip/tongue area) [18, 55, 57]. This is less straightforward for MoA, as it determines phonemes from two perspectives: (1) nasality (height of the velum), which differentiates between oral ([bə]) and nasal ([mə]) phonemes; (2) duration or degree of constriction, which differentiates between plosives ([bə]) and fricatives ([və]). This possibly makes it more difficult to define proper subregions to investigate in the motor cortex. Nevertheless, the present study demonstrates that MoA compared to PoA and voicing reflects more auditory processing, probably related to the large formant transition of F1 and F2 (Table 1) due to deep resonances inherent to nasal sounds, and auditory-to-motor mapping in the run-up to phoneme discrimination, which partially corresponds to our first hypothesis (questioning whether phoneme discrimination based on PoA elicits more motor cortical activity and whether phoneme discrimination based on voicing or MoA evokes more auditory cortical activity). Another interesting finding, not completely fulfilling our first hypothesis, is the fact that voicing did not show any difference with PoA and relied less on auditory cortical activity than MoA. Compared to MoA, voicing has more delineated motor features (related to the laryngeal motor function), which already proved to be supportive for discrimination of vocal pitch [55]. Perhaps this explains the lack of difference between voicing and PoA in motor regions (equal reliance on motor features) and auditory regions (lack of increased auditory processing for voicing) and the existing difference between MoA and voicing in auditory regions (less reliance on auditory features for voicing, when compared to MoA, than hypothesized).

During attentive phoneme discrimination, a spatial differentiation starts already between 65 and 110 ms, just as during the pre-attentive condition, where motor areas, inferior parietal regions and auditory superior temporal regions are more activated during phoneme discrimination based on MoA and voicing. Later on, between 130 and 175 ms a reversed pattern emerged showing more activation of exactly the same motor and auditory regions during phoneme discrimination based on PoA, instead of MoA and voicing. Finally, between 225 and 250 ms only the difference between PoA and MoA remained, showing more sensorimotor and superior temporal activation during phoneme discrimination based on PoA. Based on the underlying ERP waveforms, it appears that in the early time window (65–110 ms) an initiation of an N100 is not yet present with PoA as phonemic contrast while this is the case for MoA and voicing. Moreover, the following time frame (130–175 ms) corresponds to the N100 onset of PoA while the N100 in response to MoA and voicing is already in its offset phase. These timing differences around the N100 potential are supported by statistically significant differences between phonemic contrasts in N100 peak latency (showing a longer latency for PoA compared to voicing and MoA). The N100 potential is generally related to primary analyses, preparatory processes like detection and identification of phonemes based on their different constituent features [27, 31, 58]. So, the preparation for phoneme discrimination based on PoA occurs later than the preparation for phoneme discrimination based on MoA and voicing, which clearly illustrates the importance of taking temporal processing into account. Nonetheless, the same structures are engaged, showing auditory-motor integration during these preparatory processes, seemingly rejecting the hypothesis of more auditory cortical activity for MoA and/or voicing. However, in the comparison between PoA and voicing the latter relies on both auditory temporal and motor regions between 65 and 110 ms, whereas the former does not and depends only on motor areas in the later time window between 130 and 175 ms. Hence, PoA does appear to be more motor imprinted, most likely due to its distinctive articulatory gestures, while voicing is more acoustically imprinted in the brain, probably related to its distinct rapid temporal cues (voice-onset time), at least during active, attentive phoneme discrimination (as there is more motor involvement during passive phoneme discrimination, see above).

A last finding is the absence of regional activation differences between any of the phonemic contrasts during pre-attentive or attentive phoneme discrimination in the time window associated with the MMN peak or onset of the P300 peak (between 310 and 340 ms). However, in the attentive condition there were significant longer latencies for PoA compared to MoA, showing a similar pattern as the timing differences around the N100 potential as mentioned above. So, attentive phoneme discrimination with PoA remains to be prolonged compared to MoA. Apparently, in contrast with the phonological preparatory processes, the actual phoneme discrimination process does not evoke an increased activation of auditory-motor regions depending on the phonemic contrast, despite the timing differences. In addition, studies which found differences between articulatory representations in the motor cortex mainly used passive listening to phonemes, merely probing phoneme perception and recognition of one single phonemic contrast [16–18, 20]. Perhaps purely discriminating between phonemes does not lead to altered activity in auditory and motor regions with respect to different phonemic contrasts whereas the preparatory phases, such as phoneme detection and identification, do so. The fact that the auditory-motor integration during the preparatory phases for PoA phoneme discrimination occur later than the auditory-acoustic processes as preparation for MoA/voicing phoneme discrimination, can perhaps be explained, although still speculative, from a developmental perspective. Already in utero, the foetus is capable of hearing environmental sounds, considering that the auditory organ reaches its adult size by the twentieth week of gestation [59]. This is supported by the presence of auditory evoked responses in the foetal brain as measured with foetal magnetoencephalography (fMEG) [60]. Such auditory evoked magnetic fields have proven that foetuses, mostly between 27 and 36 weeks gestational age, are capable of detecting tone-frequency changes [61–64] and that neonates can process rapid temporal cues within 3 weeks after birth [60]. Evidently, processing of sounds during foetal development and early infancy highly relies on acoustic-auditory features. As already mentioned in the introduction, the phonemic contrasts voicing and MoA are highly acoustically founded and well-defined (temporal cue of voice onset time and F2 transition related to deep resonances, respectively) whereas PoA has a more elusive acoustic nature (lack of specific, invariant acoustic cues and presence of multiple acoustic variables; [36]) with a marked audio-visual, motor-phonetic component. The motor features are learned and related to auditory signals well after birth, when the infant starts babbling and receives tactile, proprioceptive and auditory feedback when producing sounds and sees and hears sounds being produced [65]. Because of the vague acoustic composition of PoA, an integrative perception strategy must be engaged based on auditory and motor features, whereas perception of voicing and MoA can be completed successfully based on auditory features only. It is possible that the early auditory-acoustic imprinting of voicing and MoA and the later auditory-motor integration for PoA during phoneme perception found in the last study might be a reflection of a very early, prenatal development of the central auditory system and a later, postnatal formation of the auditory-motor integration network. The fact that this (spatio)temporal dissociation of the preparatory processes for PoA and MoA/voicing only surfaces during the attentive phoneme discrimination task, can potentially be related to the explicit conscious cognitive process underlying the nature of this active oddball paradigm, opening a ‘gate’ for uncovering long-established developmental patterns (which is less possible with automatic, pre-attentive tasks).

In any way, it is demonstrated that starting from 50 ms after phoneme presentation, activity of inferior frontal, sensorimotor, inferior parietal and auditory regions (corresponding to a dorsal processing stream) temporally differentiates between the three phonemic contrasts during phoneme discrimination.

An effect of attention was observed when the different phonemic contrasts were compared in the pre-attentive and attentive condition. Activity differences between phonemic contrasts were found in the inferior frontal region during attentive phoneme discrimination, whereas no such dissimilarities were found during pre-attentive phoneme discrimination. Frontal regions have repeatedly been associated with selective and directed attention processes [66–68]. Moreover, a study where the MMN was compared in a conventional, pre-attentive condition and a more attention-demanding design attributed greater inferior frontal involvement during the attention-demanding design to some form of “attention-switching towards auditory linguistic input” [34]. With respect to the present results in the attentive condition, this might indicate that a voluntary switch of attention to the heard phonemes occurs in an early time window for MoA and voicing and later in time, between 130 and 175 ms, for PoA. This might also explain why the PoA-related preparatory analyses start later in time than the MoA-related and voicing-related preparatory processes.

To conclude, pre-attentive as well as attentive auditory phoneme discrimination demonstrated a clear differentiation between all three phonemic contrasts. Firstly, during the early, preparatory phases (50–100 ms) more intense auditory processing and auditory-to-motor mapping was required for pre-attentive phoneme discrimination based on MoA. Secondly, there was a delayed initiation of the preparatory phases with the phonemic contrast PoA, though the same auditory and motor areas were engaged as the other two phonemic contrasts (65–110 ms; 130–175 ms), during attentive phoneme discrimination. A larger fronto-parietal attention-related network was active during attentive (voluntary) as well as pre-attentive (involuntary) phoneme discrimination, but showed greater variation in time and degree of influence for the phonemic contrast PoA. Above that, with the phonemic contrasts MoA and voicing there was a higher emphasis on auditory regions during involuntary attention-allocation, whereas for PoA attention-allocation relied solely on motor areas. Finally, it must be taken into account that the present study is limited in its spatial sampling because of the less dense electrode coverage. As such, conclusions regarding location of activation differences must be interpreted with due caution. However, we do believe this source reconstruction method is worth evaluating in a population of persons with acute aphasia to assess sublexical processing on a neuronal level.