The present study used a novel experimental protocol to demonstrate that the production of an inner phoneme resulted in sensory attenuation of the auditory-evoked potential elicited by a simultaneously-presented audible phoneme, in the absence of any overt motor action. Crucially, the production of inner speech did not result in equal sensory attenuation to all sounds; sensory attenuation was dependent on the content of the inner phoneme matching the content of the audible phoneme. These results suggest that inner speech production is associated with a time-locked and content-specific internal forward model, similar to the one believed to operate in the production of overt speech (Hickok, 2012; Tourville and Guenther, 2011; Hickok and Poeppel, 2004; Houde et al., 2013). In short, the results of this study suggest that inner speech alone is able to elicit an efference copy and cause sensory attenuation of audible sounds.

The key finding of the present study was that the production of inner speech, by itself, led to N1-suppression to an audible sound. N1-suppression has been reported many times previously in response to overt speech (Ford et al., 2007a; Curio et al., 2000; Heinks-Maldonado et al., 2005; Oestreich et al., 2015; Houde et al., 2002). There is strong evidence that the mechanistic basis of N1-suppression to overt speech involves an efference-copy mediated IFM (Eliades and Wang, 2003; Crapse and Sommer, 2008; Rauschecker and Scott, 2009). It thus seems reasonable to assume that a similar mechanism underlies sensory attenuation in the case of inner speech. The idea that inner speech shares mechanistic features with overt speech is consistent with the conceptualization of inner speech ‘as a kind of action’ (Jones and Fernyhough, 2007, p.396) and, more generally, with Hughlings Jackson’s belief that thinking is merely our most complex motor act: ‘sensori-motor processes… form the anatomical strata of mental states’ (Hughlings Jackson, 1958, p.49). In the words of Oppenheim and Dell (2010, p.1158), ‘inner speech cannot be independent of the movements that a person would use to express it’.

It was notable that inner speech production resulted in N1-suppression if – and only if – the content of the inner phoneme matched the content of the audible phoneme; that is, only in the Match condition. That said, we note that comparisons between the Match/Mismatch conditions on the one hand, and the Passive condition on the other, should be treated with a degree of caution. This is because data for these conditions came from different trial blocks, and (most importantly) differed in terms of the task that participants were required to perform (i.e., produce a covert/overt phoneme at the precise sound-time, versus passively listen to the audible phoneme). Notwithstanding the fact that the N1-suppression effect has previously been found to be robust to variations in attention (Timm et al., 2013; Saupe et al., 2013) and trial structure (Baess et al., 2011), this nevertheless raises the possibility that differences in participants’ attention or task-preparation may have contributed to the observed differences between the ‘active’ conditions and the passive condition. We note that this limitation is not restricted to the current study – it potentially applies to any procedure that attempts to measure sensory suppression by comparing active and passive conditions, which constitutes the vast majority of studies that have examined sensory suppression to overt speech. Critically, however, this limitation does not apply to the key contrast between the Match and Mismatch conditions. This is because data for these conditions came from the same trial blocks, in which participants were required to perform exactly the same task on each trial. For example, in blocks in which participants were required to produce the inner phoneme /ba/, they experienced both Match trials (in which the audible phoneme was /BA/) and Mismatch trials (in which the audible phoneme was /BI/), and there was no way for participants to predict whether the current trial would be a Match or Mismatch trial prior to the onset of the audible phoneme. Hence attention, task-preparation, etc., during the pre-stimulus period could not differ systematically between Match and Mismatch trials. Our factorial design also ensured that the differences between the Match and Mismatch conditions were not driven by differences in the inner phoneme (i.e., /ba/ vs /bi/) or differences in the audible phoneme (i.e., /BA/ vs /BI/). Consequently, we can conclude with confidence that the observed difference in N1 amplitude between Match and Mismatch trials reflects the impact of participants’ inner speech on their sensory response to the audible phoneme. It is this contrast that demonstrates that inner speech, like overt speech, is associated with a precise, content-specific efference copy.

The results of the inner speech experiment mirror those of previous studies which have found sensory attenuation to overt speech to be reduced or eliminated if auditory feedback deviates from what would normally be expected; e.g., by pitch-shifting the feedback or providing foreign-language feedback (Heinks-Maldonado et al., 2006; Behroozmand and Larson, 2011; Behroozmand et al., 2011; Behroozmand et al., 2016; Larson and Robin, 2016). To confirm this pattern using the current procedure, we performed a supplementary experiment in which participants were required to overtly vocalize the phoneme at the sound-time. The key result from the overt speech experiment was that N1-amplitude elicited by an audible phoneme was significantly smaller when participants simultaneously produced the same phoneme in their overt speech (Match condition), compared to when they produced a different phoneme in their overt speech (Mismatch condition). This result, which is consistent with numerous previous studies in the sensory attenuation literature (Heinks-Maldonado et al., 2006; Behroozmand and Larson, 2011; Behroozmand et al., 2011; Behroozmand et al., 2016; Larson and Robin, 2016), was identical to the corresponding contrast in the inner speech experiment, in which participants had to produce phonemes in inner, as opposed to overt, speech.

A second notable finding of the overt speech experiment was that N1-amplitude was significantly smaller in the Passive condition compared to both active conditions (i.e., Match and Mismatch). A potential ‘low-level’ explanation for this result lies in the fact that participants were required to make an overt motor action in the overt speech experiment but not the inner speech experiment. While the marked difference in pre-stimulus activity between the active and passive conditions (Figure 5a) is consistent with this explanation, the fact that between-condition differences remained when correcting for the motor-associated activity (Figure 5b) stands against this possibility (though see Horváth, 2015; Sams et al., 2005 for a discussion of the challenges associated with motor correction when comparing active and passive conditions in studies of sensory attenuation). An alternative explanation for why N1 was smallest in the Passive condition is based on the idea that the auditory N1-component is (in addition to sound intensity, discussed previously) also sensitive to stimulus predictability, with predictable sounds evoking a smaller N1 than unpredictable sounds (Behroozmand and Larson, 2011; Bäss et al., 2008; Lange, 2011). Our task differed from other willed vocalization tasks in the literature in that the audible phoneme delivered to the headphones was: (a) of a different person’s voice, and (b) much louder than the actual vocalization, as participants were instructed to vocalize quietly in order to minimize bone conduction. In other words, a substantial discrepancy between the predicted and actual sound existed, even in the Match condition. This sensory discrepancy was even larger in the Mismatch condition, as the content of the sound was also different. Consistent with the idea that N1-amplitude is sensitive to stimulus predictability, it is possible that the larger N1-amplitude in the active compared to the passive conditions was due to prediction-errors as to the expected quality of the audible phoneme. It is further possible that such detailed predictions as to phoneme quality do not occur in the context of inner speech (a suggestion for which there is some empirical support – Oppenheim and Dell, 2008), which may account for why N1-amplitude was not reduced in the passive condition in the inner speech experiment.

In summary, both the inner speech and overt speech experiments showed the same basic pattern of results with respect to the key contrast: N1-magnitude was smaller if the phoneme generated by the participant (either covertly or overtly) matched the audible phoneme than if it mismatched. These findings suggest that inner speech – like overt speech – is associated with a precise, content-specific efference copy, as opposed to a generic and non-specific prediction. Taken together, our results provide support for the contention that inner speech is a special case of overt speech, which does not have an associated motor act. The notion that inner speech generates an IFM in the absence of an overt motor act has been hypothesized previously across several different literatures (Jones and Fernyhough, 2007; Feinberg, 1978; Scott, 2013; Guenther and Hickok, 2015; Ford et al., 2001b; Sams et al., 2005; Kauramäki et al., 2010). However, this hypothesis has been notoriously difficult to test empirically, due to the covert nature of inner speech. Ford et al., (Ford and Mathalon, 2004) played participants repeated sentences over 30 s and asked them to reproduce the same sentences in inner speech. Ford et al., found that the sentences elicited a smaller N1-component when participants engaged in inner speech compared to when they did not, consistent with the results of the present study. More recently, Tian and Poeppel (Tian and Poeppel, 2010) used MEG to show that the auditory cortex was activated immediately following production of an inner phoneme in the absence of auditory feedback, which they took as evidence of an inner-speech-initiated efference copy. In a subsequent study, which was a strong influence on our own, Tian and Poeppel (Tian and Poeppel, 2013) asked participants to produce an inner phoneme within a 2.4 s window. This window was followed by an audible phoneme that could either match or mismatch the content of the inner phoneme. The authors found no difference in the amplitude of the M100 between the match and mismatch conditions, inconsistent with the results of the present study. However, given the width of the temporal window in which participants were asked to produce their inner phoneme (2.4 s), the efference copy and auditory feedback would not necessarily be expected to coincide under these conditions, in which case M100-suppression would not be expected to occur. Tian and Poeppel (Tian and Poeppel, 2015) asked participants to signal their production of an inner phoneme via a button-press, and measured the amplitude of the M100 component evoked by a pre-recorded audible phoneme of their own voice which matched the content of the inner phoneme, but which could be pitch-shifted or delayed. They found evidence of M100 suppression to unshifted, undelayed audible phonemes relative to a passive baseline condition, consistent with the results of the present study. Pitch-shifting or delaying the auditory phonemes was found to increase M100 amplitude above baseline levels. While this study’s design enabled the timing of the inner phoneme to be precisely specified, the fact that it was specified by means of an overt motor action (i.e., a button-press), which is known to be associated with N1-suppression per se (Hughes et al., 2013), raises the possibility of the motor action and inner speech being confounded. Finally, Ylinen et al., (Ylinen et al., 2015) asked participants to mentally rehearse tri-syllabic pseudowords in inner speech. After several mental repetitions, an audible pseudoword was played which had either matching or mismatching beginnings and endings to the rehearsed pseudoword. The results revealed that audible syllables that were concordant with participants’ inner speech elicited less MEG activity than discordant syllables, a result which is broadly consistent with the results of the present study.

The current experiment holds some methodological advantages over previous designs. Firstly, the experimental stimuli (animation, audible phonemes, rating-scale, etc.) were physically identical across all conditions, as was the nature of participants’ task (i.e., to fixate on the screen and produce an inner phoneme at a designated time). The only thing that differed between the different trial-types was the inner phoneme that participants were asked to produce. This meant that the observed differences in sensory attenuation could not have been due to any physical differences between the conditions (Luck, 2005). Secondly, the fact that it was impossible to predict which of the two audible phonemes would be presented on any given trial meant that it was impossible to distinguish Match from Mismatch trials a priori. This meant that the observed results could not have been due to between-condition differences in, for example, demand characteristics. Thirdly, the ‘ticker tape’ feature of the current protocol enabled participants to very accurately time-lock the onset of their inner phoneme to match the onset of the external sound. In the current protocol, the position of the trigger line refreshed every 8.3 ms, which presumably enabled participants to time the onset of their inner phoneme far more accurately than would be possible with a countdown sequence, mental rehearsal or open temporal window, such as have been used in previous studies (Tian and Poeppel, 2013; Ford et al., 2001b; Ylinen et al., 2015). Finally, the current protocol did not require participants to make an active movement to signal the onset of their inner phoneme, such as by pressing a button. This is a significant advantage over previous studies which have employed a motor condition to signal the onset of covert actions, as it avoids the potential confound associated with having temporally-overlapping auditory-evoked and motor-evoked potentials – see Horvath (Horváth, 2015) and Neszmélyi and Horvath (Neszmélyi and Horváth, 2017) for a discussion of the challenges associated with ‘correcting’ for motor activity in studies of sensory attenuation. In light of its methodological features, the present study provides arguably the strongest evidence to date that inner speech results in sensory attenuation of the N1-component of the auditory-evoked potential, even in the absence of an overt motor response. Perhaps the most important strength of this paradigm is that all the above issues were controlled within a single task, thereby removing any reliance on cross-experimental inferences.

This study’s focus on the N1 component is consistent with the majority of the existing literature on electrophysiological sensory attenuation. The rationale for focusing on N1 lies in the fact that the amplitude of this component is volume dependent; that is, other things being equal, loud sounds evoke N1-components of larger amplitude than do soft sounds (Näätänen and Picton, 1987; Hegerl and Juckel, 1993). In prior studies of N1-suppression, participants have typically generated sounds through overt actions such as overt speech, button-presses etc. The observation of N1-suppression in such studies thus implies that the brain processes self-generated sounds as though they were physically softer than identical external sounds. The N1-suppression demonstrated in the present study extends this idea by suggesting that the brain also processes imagined sounds as though they were physically softer than identical, unimagined sounds. In addition to providing evidence that inner speech is associated with an IFM of similar nature to overt speech, this finding provides evidence that mental state influences perception at a fundamental level (Gregory, 1997).

With regards to the question of what mechanism could underlie sensory attenuation to inner speech: a recent study by Niziolek, Nararajan and Houde (Niziolek et al., 2013) on sensory attenuation in the context of overt speech production found that the degree of sensory attenuation was stronger when participants produced vowel sounds that were closer (in terms of their acoustic properties) to their median production of these sounds, compared to when they produced vowel sounds that were, for them, less typical. These results suggests that the efference copy associated with overt speech production represents a sensory goal (i.e., ‘a prototypical production at the center of a vowel’s formant distribution’, p. 16115), and that the distance (in formant space) of any given utterance from this ‘sensory prototype’ determines its degree of sensory attenuation. If inner speech is, in fact, a special case of overt speech (as we have suggested above), then this raises the question of the nature of the sensory goal in the context of inner speech production. One possibility, based on the results of Niziolek et al., is that in the present study (in ‘imagine /ba/’ trials, for example) the sensory goal was of a prototypical /ba/, which was presumably covertly ‘spoken’ in the participant’s own voice (though see below for a discussion of the validity of this assumption). In this case, the participant’s prototypical /ba/ would never match perfectly with the audible phoneme, as the audible phoneme would never be the participant’s own voice.

The fact that the present study observed N1-suppression in the Match condition but not the Mismatch condition is nevertheless consistent with the Niziolek et al., 2013 account, in that the distance, in formant space, between an inner /ba/ and an audible /BA/ would be smaller than the distance between an inner /bi/ and an audible /BA/, even though the inner phoneme did not match the audible phoneme perfectly in either case (as the audible phoneme was always produced by the same unknown speaker). The fact that while Niziolek et al., 2013 observed maximal levels of sensory attenuation to prototypical vowel sounds, they still observed significant (i.e., non-zero) levels of sensory attenuation to atypical vowel sounds is also consistent with this idea. A prediction of this account is that participants should show even greater levels of sensory attenuation in the Match condition if the audible phoneme is presented in their own voice rather than the voice of an unknown stranger; testing this prediction may be a worthwhile endeavor in future studies.

In regards to the assumption that the sensory goals of inner speech are the same as overt speech, and that a person’s inner voice is the same as their actual voice, there is some evidence in support of this conjecture: Filik and Barber (2011) provided evidence that people produce inner speech in the same regional accent as their overt speech. However, other studies have reported evidence suggesting that inner speech has impoverished acoustic properties relative to overt speech (Oppenheim and Dell, 2008). It is also possible that inner speech can consist of several distinct ‘voices’, with each having specific auditory properties; the fact that people with auditory-verbal hallucinations often report hearing multiple voices with different auditory properties (McCarthy-Jones et al., 2014) is consistent with this idea, if – as discussed further below – auditory-verbal hallucinations ultimately reflect inner speech being misperceived as overt speech. Finally, it is also possible that the acoustic properties of inner speech are not fixed. Specifically, in the context of the present study, it is possible that the acoustic properties of the audible phonemes began to influence the inner phonemes, such that after numerous repetitions, participants began to imagine themselves producing an inner phoneme with the acoustic properties of the audible phoneme. Testing these possibilities may also be worthwhile in future studies.

While the primary focus of the paper was on the N1-component of the auditory-evoked potential, between-condition differences were also observed in the amplitude of the P2 and P3 components (see Figures 3 and 4). A likely explanation for the observed results in these later components involves another ERP component, the N2, whose spatial and temporal distribution typically overlaps with that of the P2 (Griffiths et al., 2016). The N2 and P3 components are among the most heavily investigated components in the ERP literature (Näätänen and Picton, 1986; Polich, 2007), and are typically elicited by tasks – such as the auditory oddball and Go/NoGo tasks – in which the participant is asked to identify (by means of a button-press, for example) ‘target’ stimuli which are nested among ‘non-target’ stimuli (Smith et al., 2010; Spencer et al., 1999). Critically, the N2 and P3 can also be elicited by tasks in which a mental response is required, such as when target stimuli have to be mentally counted as opposed to signaled with a button-press (Mertens and Polich, 1997). We suggest that, in the two inner speech conditions of our study, participants made a mental response – possibly a ‘template-matching response’ along the lines of whether the audible phoneme matched their inner phoneme (Griffiths et al., 2016) – which they did not make in the Passive condition. In this case, the audible phoneme in the inner speech conditions might be expected to elicit an N2 and a P3 component, which would not be present in the Passive condition. The occurrence of a (negative-going) N2 in the inner speech condition would then interact and compete with the expression of a (positive-going) P2 component elicited by the audible phoneme. The result would be the absence of a distinct P2, but presence of a P3, in the inner speech conditions but not the Passive condition – as observed empirically. It is also worth noting that Tian and Poeppel (Tian and Poeppel, 2013) observed a larger M200 component in their match vs. mismatch comparison, consistent with the enhanced P2 observed in the Match vs. Mismatch comparison in the present study. Taken together, these results suggest that the M200/P2 component may index something other than a sensory prediction, possibly involving a cognitive ‘template matching’ process.

The implied existence of an efference copy to inner speech holds important implications for how to best understand some of the psychotic symptoms associated with schizophrenia. Some of the most characteristic of these symptoms seem to reflect the patient misattributing, to external agents, self-generated motor actions (e.g., delusions of control) and self-generated thoughts (e.g., delusions of thought insertion, auditory-verbal hallucinations – Feinberg, 1978; Frith, 1987). An influential account of these experiences argues that they arise because of an abnormality in the IFM associated with both physical and mental actions (Feinberg, 1978; Frith, 1992). This IFM abnormality leads to an inability to predict and suppress the consequences of self-generated actions, which leads to confusion as to their origins. This hypothesis has a strong theoretical foundation: for example, the distinctive symptom of thought echo, in which the patient hears their own thoughts spoken out loud by an external voice, can be well explained as the patient’s own inner speech being misattributed and misperceived as an external voice (Frith, 1992). However, while numerous studies have provided empirical evidence showing that schizophrenia patients exhibit subnormal levels of sensory attenuation to their own physical actions (Whitford et al., 2011; Blakemore et al., 2000b; Shergill et al., 2005), including subnormal levels of N1-suppression to overt speech (Ford et al., 2001a; Ford et al., 2007b; Ford et al., 2001c), there is little empirical evidence that schizophrenia patients show sensory attenuation deficits to self-generated mental actions such as inner speech. Furthermore, the few studies that did report sensory attenuation deficits to inner speech in patients with schizophrenia did not include a ‘mismatch’ condition, raising the possibility that these sensory attenuation deficits ultimately reflect attentional deficits in the patient group (Ford and Mathalon, 2004; Ford et al., 2001d). We suggest that the failure to identify electrophysiological sensory attenuation deficits to inner speech in schizophrenia patients is not because the deficits do not exist, but rather because previous experimental protocols have been insufficiently sensitive to detect them. In order to maximize the chances of detecting sensory attenuation deficits to inner speech in schizophrenia patients, we suggest that future experiments should: (a) ensure that the onset of inner speech is precisely time-locked to the audible sound, without reverting to using a willed action (e.g., a button-press) to signal inner speech-onset, as this could potentially lead to the auditory and motor responses being confounded; (b) limit the content of inner speech to phonemes rather than entire sentences as this enables the onset and content of the inner speech to be more tightly controlled; (c) investigate patients exhibiting those symptoms that seem to most clearly reflect misperceived inner speech (e.g., thought echo, other auditory-verbal hallucinations), rather than grouping patients with different symptom profiles into a clinically-heterogeneous ‘schizophrenia’ group. By providing an optimized protocol for quantifying N1-suppression to inner speech, our hope is that the present study can provide a methodological framework for identifying and assessing sensory attenuation deficits in inner speech in patients with schizophrenia.

In conclusion, the present study demonstrated that engaging in inner speech resulted in sensory attenuation (specifically, N1-suppression) of the electroencephalographic activity evoked by an audible phoneme, but only if the content of inner speech matched the content of the audible phoneme. These results suggest that inner speech evokes an efference-copy-mediated IFM, which is both content-specific and time-locked to the onset of inner speech, which is consistent with the existing literature on sensory attenuation to overt speech. Cumulatively, this implies that inner speech may ultimately be ‘a kind of action’, and a special case of overt speech, as long suggested by prominent models of language. Accordingly, these findings not only provide insight into the nature of inner speech, but also provide an experimental framework for investigating sensory attenuation deficits in inner speech, such as have been proposed to underlie some psychotic symptoms in patients with schizophrenia.