Speech production is one of the most complex human motor acts as it bases upon linguistic processing but further requires precise sensorimotor integration. While linguistic processing involves multiple steps of abstract cognition (Indefrey and Levelt, 2004), the sensorimotor component of speech production relies on feedforward motor plans that are updated by integration of auditory and somatosensory feedback (Hickok, 2012). It is thus not surprising that a large network including prefrontal, motor, somatosensory, auditory, and associative regions has been linked with overt articulation (Price, 2010; Kell et al., 2011; Llorens et al., 2011). When studying speech, sensorimotor processing is more easily accessible than linguistic processing given the measurable physical auditory and motor signal properties. It is thus ideally suited to study spectro-temporal dynamics of large scale brain networks. Evidence is accumulating that specific spectro-temporal properties of brain regions and networks contribute to the generation of complex motor behavior (Roelfsema et al., 1997; Engel and Fries, 2010; Wang, 2010). However, little is known about the spectro-temporal dynamics and interactions underlying efficient sensorimotor mapping. During speaking, the parameters of interest are not easily accessible with standard imaging methods with high temporal resolution like EEG or MEG because of the artifacts induced by ocular and articulator movements during ongoing speech production. One approach to circumvent this problem is to study brain dynamics of preparation for speech production. As for other motor acts, motor preparation for speech production has been related to beta suppression in motor cortex (Salmelin et al., 2000; Saarinen et al., 2006). These processes could reflect covert feedforward processing prior to execution induced by a start signal and thus could serve as a model for the feedforward component of speech production.

In contrast, feedback is necessarily only produced after speech onset and thus much more difficult to study. During monkey vocalizations, suppression of activity in auditory cortices has been linked to feedback integration (Eliades and Wang, 2008). Ongoing speech production has been studied with electrocorticography, a method that is not affected by movement artifacts. By studying high gamma power and its temporal evolution during task epochs, these studies confirmed suppression of auditory cortex activity during speech production (Towle et al., 2008; Edwards et al., 2010; Flinker et al., 2010) associated with a very complex pattern of frontotemporal functional connectivity (Korzeniewska et al., 2011). Instead of studying the speech production network during active processing, another way of approaching this large scale brain network is to investigate how this network is set up. This offers the opportunity to investigate the effects of top-down implementation of task rules (Dosenbach et al., 2006; Sakai and Passingham, 2006). We hypothesize that processes termed “task set” induce recruitment of all necessary brain regions and thus increase functional connectivity within the task network. Here, we refer to these processes as cognitive planning and study them using cue-target paradigms that allow their separation from linguistic processing, motor preparation, and execution. In a cue-target reading paradigm involving covert and overt reading during fMRI, we demonstrated previously that the entire speech production network pre-activates in anticipation of linguistic stimulus material for articulation (Kell et al., 2011). Importantly, this intention to speak left-lateralizes auditory and somatosensory cortices for subsequent feedback processing. This lateralized anticipatory preactivation of auditory cortex could reflect integration of the auditory cortex into the speech network prior to a relative deactivation during ongoing feedback integration. On the one hand this supports the idea that the entire network is set up and ready for input while on the other hand these results emphasize the important contribution of sensory cortices to left-lateralization of speech production (Morillon et al., 2010; Kell et al., 2011). Taken together, these data suggest that important properties of speech production like left-lateralization can be studied by focusing on the setup of this network.

Both local oscillations and network properties of the setup of the speech production network have not been studied yet. So far, only the introspective evaluation of the intention to speak was studied with MEG (Carota et al., 2010). We expected to find task-specific changes in local oscillations in previously defined regions of the speech production network (Hickok and Poeppel, 2007; Price, 2010; Kell et al., 2011; Llorens et al., 2011). The setup of large scale brain networks has previously been related to synchronization in the beta band as demonstrated for fronto-parietal visuomotor networks (Roelfsema et al., 1997). Thus, the aim of our MEG study was twofold: first, to define the role of local oscillations contributing to the intention to speak. Second, to study the speech network’s phase dynamics in the beta band during network setup.

Preparation for overt or covert speaking was not associated with mouth movements. When analyzing the classical frequency bands, anterior sensors exhibited a condition-independent power increase ranging from delta to beta frequencies while posterior sensors showed suppression of alpha and beta power. No gamma effect was observed (Figure 3).

The power spectrum revealed an anterior broadband effect (2–15 Hz) that was only masked by the anterior alpha peak (Figure 4). In posterior sensors, alpha and beta oscillations were suppressed for both conditions. Preparation-related effects were sustained in time (Figure 5). However, differences between conditions built up only around 350 ms after cue presentation. Preparation for overt reading was associated with a pronounced alpha and beta suppression. Interestingly, this effect was particularly strong at posterior sensors while anterior sensors did not discriminate as well between conditions (Figure 5).

The broadband power increase from 2 to 15 Hz observed for both conditions at anterior sensors localized to the bilateral prefrontal cortex (not shown). Even when source-localizing the broadband effect outside the alpha peak, a large overlap of prefrontal delta/theta (Table 1) and 11–15 Hz (Table 2) sources was observed (Figure 6).

Condition-independent decreases in alpha and beta power localized to posterior brain regions in the parietal, occipital, and temporal lobe (not shown).

Preparation for overt speech production compared to preparation for covert speech production revealed pronounced alpha suppression (Figure 7; Table 3) of the left auditory cortex (planum temporale), left superior and middle temporal cortex, the SMA, and the middle occipital gyrus.

Lateralization of alpha suppression was most robust in the temporal lobe (LI = 0.4) followed by frontocentral cortices (LI = 0.2). The posterior beta decrease associated with the intention to speak localized (Figure 7; Table 4) to the left articulatory motor cortex, left area SPT, left superior and inferior temporal gyrus, and bilateral superior parietal cortex. The strongest left-lateralization of beta suppression was found in temporal lobe (LI = 0.5) followed by frontocentral (LI = 0.4) and parietal cortices (LI = 0.25). It is interesting to note that the decrease in alpha and beta power regionally overlapped in some, but not all regions of the speech production network. While both left mid- and inferior temporal and parietal regions showed parallel changes in alpha and beta power, the auditory cortex and the SMA showed only alpha suppression. Isolated beta suppression was observed in the articulatory motor cortex and area SPT.

Comparison of master-slave models against models with bidirectional coupling revealed that the speech production network is bidirectionally coupled in beta phase within hemispheres (Figure 8A). Assuming bidirectional coupling between hemispheres, the model with coupling between the SMA and bilateral dPMC, but no other interhemispheric coupling, was selected by Bayesian model comparison (Figure 8B). In all network connections, beta phase coupling was modulated by preparation for overt reading (Figure 8C).

The setup of the speech production network was associated with a sustained reduction in oscillatory alpha and beta power that occurred about 350 ms after cue presentation. These changes were primarily left-lateralized, particularly in temporal cortex. While auditory regions showed stronger suppression and lateralization of alpha power, the auditory-motor region area SPT, and the articulatory motor cortex showed primarily suppressed beta power when speaking was anticipated. Note that alpha suppression extended beyond sensory cortices: we observed this sustained power decrease also in mid-temporal regions related to phonology like the superior temporal sulcus (Hickok and Poeppel, 2007) and also in the SMA. Changes in both alpha and beta oscillations have previously been shown to be behaviorally relevant (Hammond et al., 2007; Dugue et al., 2011). Our data indicate that cognitive planning prior to actual sensory or motor processing could contribute to such changes in behavior. Those findings suggest that preparation of the speech production network extends beyond parietal regions that have previously been related to the conscious introspection of the intention to speak (Carota et al., 2010).

Alpha oscillations are thought to rhythmically inhibit sensory cortices (Jensen and Mazaheri, 2010; Klimesch et al., 2011). Auditory cortex activation has been related to alpha suppression (Hartmann et al., 2012). The observed alpha suppression in sensory, associative, and premotor regions in this study could tentatively be interpreted as an active disinhibition of task-relevant regions. This parallels findings of increased preparatory BOLD activity in auditory cortex during the intention to speak (Kell et al., 2011) given that alpha oscillation amplitude and BOLD fMRI activity are typically anti-correlated (Brookes et al., 2005).

Sustained beta power increase has been related to maintenance of a network in its status quo (Engel and Fries, 2010). The observation of beta suppression in motor-related regions could indicate a preparation for a change in the motor system (Saleh et al., 2010). Importantly, this already occurs before movement or even specific motor preparation is possible. Thus, beta suppression very likely plays an additional role to the one previously suggested in the framework of motor preparation. Whether beta oscillations play an equivalent inhibitory role in the motor system (Hammond et al., 2007; Engel and Fries, 2010) compared to alpha oscillations in more sensory regions, remains open to further research.

An important limitation of our study is the use of an explicit (overt) and an implicit (covert) condition. We thus cannot exclude that condition-specific effects are partly related to different levels of difficulty and attention. A study investigating the interaction of attentional and task factors could not reveal a modulation of speech network activity by task difficulty or attentional demand (Keller and Kell, in preparation).

Although beta power is suppressed in several nodes of the speech production network, these brain regions mutually increase their beta phase coupling in anticipation of overt reading. Beta band synchronicity has previously been related to coupling in large scale networks like the frontoparieto-occipital visuomotor network (Roelfsema et al., 1997). We extend this finding to the speech production network. Interestingly, we did not detect hierarchical asymmetry in coupling strength between top-down and bottom-up connections, suggesting that beta synchronicity may indeed play an important role for the setup of the network prior to stimulus related processing. It is conceivable that this pattern changes when ongoing top-down and bottom-up processes take place during execution. Note that our DCMs explain data only after the visual cue was decoded. This is a trial phase after bottom-up processing of the cue has taken place. In our study, the premotor cortex including the dPMC and the SMA connects the left with the right sensorimotor speech system during network setup. The premotor cortex is thus ideally suited to integrate information from both hemispheres. This region has previously been associated with auditory-motor integration (Neef et al., 2011), suggesting that the bilateral premotor cortex integrates sensory feedback into a motor program (Hickok et al., 2011).

The effects specifically related to speech preparation were preceded by condition-independent preparatory effects. An early posterior alpha and beta suppression which likely originated from parietal sources was accompanied by a prefrontal increase in low frequencies ranging from 2 to 15 Hz. Importantly, this low frequency effect was not modulated by specific task content and was thus observed during task preparation both for overt and for covert reading. Given that there is considerable debate on what broad band effects signify, the mechanisms behind such an effect have to be studied in more detail. Such fronto-parietal activity – although unlikely oscillatory in nature – could reflect the executive control of implementing a “task set” (Dosenbach et al., 2006; Sakai and Passingham, 2006). Given that this activity did not differ between conditions, the tasks seem to have required similar amounts of executive control. The observed time courses of this activity suggest that the first 350 ms after cue presentation are used to decode the instruction and to generate the appropriate rules for behavioral control. Changes specific to the set up of the speech production network (sensorimotor alpha and beta suppression) were only detected after 350 ms. This indicates that it takes at least this time to set up this large scale network. To test this claim, we performed an additional behavioral study under the assumption that short instruction delays of less than 350 ms do not allow a proper setup of the speech production network resulting in increased reaction times measured as speech onset after target presentation. Indeed, shortening instruction delays to 330 ms increased reaction time in overt reading tasks (Figure 9). This suggests that this duration does not allow for a proper setup of the speech production network, extending network setup well into the execution phase which in turn delays speech onset. We thus believe that sensorimotor alpha and beta suppression is essential to set up the speech production network.

Taken together, our findings suggest that the brain sets up the speech production network by means of beta synchronization and prepares the network for left-lateralized information processing by suppression of frontotemporal alpha and beta oscillations.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.