New musical forms and experiments often challenge established structural forms in composition or reshape old ones into new uses, reordering fundamental musical building blocks to challenge the perception of music. The challenge of studying such diverse music in a principled and systematic manner within a context of neuroscience demands first to establish some fundamental and perhaps common mechanisms in music perception, a work that has been flourishing in the last decades. In this article, we focus on rhythm, one of these fundamental building blocks. Within new music, some artists work with rhythmic entropy while others try to dispel rhythm altogether. Insights into some basic mechanisms of the “listening apparatus” in our nervous system as they pertain to perception of musical rhythms might perhaps still be of use to both researchers and artists, either as a starting point for more advanced research or as a starting point of artistic defiance.

In short, auditory repetition suppression and omission activation, which this study address, are opposite neural phenomena and manifestations of principles of predictive processing. Repetition suppression is the reduction of neuronal activity during listening to a repeated sound that is present, omission activation is the increased neuronal activity that occurs when an expected sound is not present. Listening activates bilateral superior temporal cortices, independent of whether we are exposed to tones, words, animal, and instrumental sounds (Specht and Reul, 2003), while prolonged listening to an unchanging sequence of sounds quickly leads to a deactivation of neural activity, a phenomenon known as repetition suppression (Grill-Spector et al., 2006). Auditory repetition suppression is a robust, experience-dependent adjustment of neural functions (Grotheer and Kovacs, 2016), predictability (Costa-Faidella et al., 2011; Cacciaglia et al., 2019), and prior expectation (Summerfield et al., 2008; Todorovic et al., 2011). It is modulated by a range of factors, such as time scales of presentation rates, sequence position and stimuli similarities (Kovács and Schweinberger, 2016), and stimuli-specific characteristics (Linke et al., 2011), and is also task-dependent (Arnott et al., 2005). Repetition suppression also neatly demonstrates principles of predictive coding (Friston, 2005; Baldeweg, 2006; Auksztulewicz and Friston, 2016), where repetition suppression biases the activity in sensory cortices (Sreenivasan et al., 2014). For auditory stimuli, repetition suppression can be seen in the temporal cortices, including Herschel’s gyrus (HG), superior temporal gyrus (STG), and middle temporal gyrus (MTG) (Auksztulewicz and Friston, 2016; Cacciaglia et al., 2019).

Repetition suppression is related to working memory mechanisms, where attenuation of neural activity can be interpreted as a minimization of activity needed in a working memory maintenance stage (Kumar et al., 2016) through the suppression of irrelevant information (Linke et al., 2011; Ahveninen et al., 2017). It is, however, unclear whether working memory is dependent on persistent neural activity (for maintaining the information) (Huang et al., 2016) or on transient reorganization of synaptic weights (D’Esposito and Postle, 2015) in representational states (Myers et al., 2017; Sreenivasan and D’Esposito, 2019) or a combination of the two (Sreenivasan et al., 2014). A potential theoretical (or actual) difference between these two descriptions—working memory as either persistent or transient neural states [or what within a predictive coding framework can be called “representation units” (Clark, 2013)]—could lie in a distinction between auditory sensory memory (shorter low-level sensory cortical retention intervals) and higher-level working memory network organization (Nees, 2016). Different cortical activation for simple sequence processing and more complex, task-specific working-memory maintenance could point to such nuances (Brechmann et al., 2007). Neural correlates for auditory working memory have been shown in temporal cortices, including STG, HG, and planum temporale (PT) (Brechmann et al., 2007; Kumar et al., 2016). Furthermore, distinct neural differences between perceptual processing and active working memory tasks for melody and pitch (Zatorre et al., 1994), separate neural correlates for duration-based and beat-based auditory timing (Teki et al., 2011), and differentiations for melody and rhythm have been shown—with the right inferior frontal gyrus and insula particularly involved (Jerde et al., 2011).

Within a predictive coding framework, auditory prediction errors (Friston, 2005) must depend on working memory or sensory memory mechanisms, since they occur when an unchanging and predictable sequence of sounds suddenly changes, i.e., when predictions and expectations are breached, or when the incoming sensory signal does not match the “representation unit” (Clark, 2013); what we will henceforth call the representational maintenance. The concept of prediction errors [or “surprisal” (Clark, 2013)] draws on findings in EEG/ERP-studies on mismatch negativity (MMN) (Naatanen et al., 1978; Kompus et al., 2015). MMN potentials are measurable neural spikes triggered by deviant and rare stimuli in a chain of standard stimuli, where the difference between the deviant and the standard stimuli are proportional to the deviance [see (Näätänen, 1992) for a review]. This difference is the mismatch or prediction error (Friston, 2005). In fMRI, as in EEG/ERP studies, the size of activation reflects the magnitude of the MMN deviant (Mathiak et al., 2002; Liebenthal et al., 2003; Eichele et al., 2008). Omissions are a particularly interesting type of deviant stimuli, where omission activation (Raij et al., 1997; Wacongne et al., 2011) describes cortical activation as a result of missing stimuli in a predictable sequence of sounds, and can therefore be assumed to be generating internal responses based solely on expectancy or prediction, and not by a change in the deviant characteristics in the stimuli itself (Jongsma et al., 2005). As with repetition suppression, MMNs or prediction error magnitude is modulated by numerous factors (Näätänen, 1992), and as with repetition suppression and auditory working memory, specific parts of the temporal cortices have repeatedly been shown to be involved in the reporting of such prediction errors (Friston, 2005). Imaging and electrophysiological studies have consistently shown that a main source of MMN potentials is located in the intersection of STG, PT, and HG (Recasens et al., 2014), predominantly right lateralized (Tervaniemi et al., 2000; Opitz et al., 2002; Doeller et al., 2003; Rinne et al., 2005), also for omissions (Mustovic et al., 2003; Voisin et al., 2006; SanMiguel et al., 2013a, b), although one study has found omission activation predominantly on the left (Nazimek et al., 2013). Pertaining to our study, these cortical areas are sensitive to beat and pattern deviations, as shown in both EEG/ERP and fMRI studies (Tervaniemi et al., 2000; Opitz et al., 2002; Doeller et al., 2003; Mustovic et al., 2003; Rinne et al., 2005; Voisin et al., 2006; Nazimek et al., 2013; Recasens et al., 2014). Beat omission and positional saliency have been used to investigate rhythm and pattern-related phenomena with several imaging and neurophysiological techniques, indicating different levels of magnitude for salient and less salient metric beat positions, although findings are somewhat ambiguous (Winkler and Schroger, 1995; Jongsma et al., 2003, 2005, 2006; Ladinig et al., 2009; Wacongne et al., 2011; Salisbury, 2012; Bouwer et al., 2014; Damsma and van Rijn, 2016).

In the current study, we wanted to examine repetition suppression, representational maintenance, and omission activation during the perceptual processing of musical rhythms. In short, these three phenomena can be seen as manifestations of key principles in predictive processing frameworks, and musical rhythms are ideal stimuli to operationalize and demonstrate these principles because of their predictive, temporal nature (Vuust and Witek, 2014; Koelsch et al., 2019). Musical rhythms also facilitate operationalization of modulating factors such as contextual characteristics (simple or complex rhythms) and saliency (position of omission).

In addition, the neuronal mechanisms involved in the perception of musical rhythms are partly known, which makes it possible to compare our results with existing literature. Listening to rhythms activates cortical motor areas, such as premotor cortex and supplementary motor area (SMA), the basal ganglia, as well as large-scale networks across the brain (Grahn and Brett, 2007; Chen et al., 2008a, b; Bengtsson et al., 2009; Geiser et al., 2009, 2012; Grahn, 2009; Trost et al., 2014; Large et al., 2015).

We also wanted to examine the effect of pattern complexity on repetition suppression and omission activation and the effect of positional saliency on omission activation. To this end, two musical rhythms, one simple and one complex, were presented several times to the participants during scans. The first part of the stimuli presentation consisted of a short presentation (16 s) of continuous, unperturbed rhythmic repetition to examine repetition suppression, segueing into a longer continuous presentation of the same rhythm. In this second part, an overt target-detection task (with a quasi-randomly distributed deviant tone) was introduced to keep the participant attending to the rhythm, while quasi-randomly distributed beat omissions were used to covertly examine omission activation (see section “Materials and Methods” for more details on the stimuli and the paradigm).

Based on previous literature, we expected to see listening, repetition suppression, and omission activation in largely overlapping areas in the temporal cortices, with omission activation occurring in more posterior areas than suppression and maintenance. We hypothesized that representation maintenance would also occur in the temporal cortices, with additional activation of inferior frontal areas, insula, and premotor cortices. We hypothesized that complexity would differentially affect suppression, as the encoding stage presumably would be affected by a higher cognitive load for the complex rhythm. We also hypothesized that omission activation would be modulated by rhythmic context where the cognitive demand (i.e., higher neural activity in a representational maintenance) of the more complex rhythm would result in smaller omission sizes, and furthermore that positional saliency would modulate omission activation, where a more salient position (beat position number one) would show higher activation than the less salient position (beat position number two). Finally, we hypothesized that there would be an interaction between pattern complexity and beat position for omission activation.

All results are reported with a FWE-corrected threshold of p < 0.05 and at least 10 voxels per cluster. Figure 2 shows a detailed excerpt of the frontal right hemisphere of the findings listed below. Figure 3 shows a more detailed overview of the findings across the while brain.

Listening to musical rhythms predictably activated the bilateral STG as well as bilateral angular gyrus. Areas in the STG attenuated after the first 4 s of repetition of the rhythmic patterns overlapped exclusively with these areas, with the size of deactivation being larger in the left STG. In addition, activity in bilateral cerebellum, and the (predominantly right) basal ganglia, including caudate nucleus, putamen, and thalamus, decreased after the first 4 s. Since processing of music (Zatorre and Zarate, 2012) and particularly rhythm (Thaut et al., 2014) has been found to be right lateralized (Large et al., 2015), this larger deactivation in the left STG might reflect an asymmetric allocation of resources, where—after initial processing in sensory cortices—the processing of musical rhythms is predominantly done in the right STG. The rapid deactivation of the basal ganglia and cerebellar areas points to a role for these areas in initial beat detection (Peretz and Zatorre, 2005; Grahn, 2009).

During the prolonged listening to rhythms after the initial 16 s of encoding, larger activity was found in the SMA and the caudate nucleus, areas well known to be activated by rhythm (Grahn and Brett, 2007; Bengtsson et al., 2009). Areas related to attention, such as the anterior insula and frontal areas (the inferior frontal gyrus in particular, but also middle and superior frontal areas), were also activated. These areas have been directly implemented in rhythm perception (Chapin et al., 2010; Heard and Lee, 2020), in particular as crucial in working memory for rhythm (Jerde et al., 2011). Activity in the precuneus, the posterior part of the cingulate cortex, and middle prefrontal cortex has in addition been found to play a particular role in the maintenance of musical beats with high beat salience (Toiviainen et al., 2020), which the rhythms in the current study must be characterized as. The areas found to be more activated during maintenance (anterior insula, frontal areas, SMA, posterior cingulate, and precuneus), with more activity in the right hemisphere, were clearly distinct compared to the other conditions and could indicate that working memory mechanisms of representational maintenance are allocated in different areas of the brain, separate from primary cortices activated during initial sensory processing.

Omission activation was distinctly more posterior than listening, suppression, and maintenance, located predominantly in the MTG, and, as expected, omission activation was significantly bigger in the right MTG (Raij et al., 1997) (see Figure 3 and Table 4), with coordinates closely matching those found in previous fMRI (Mustovic et al., 2003; Voisin et al., 2006) and EEG (SanMiguel et al., 2013a, b) studies on omissions and silences in healthy controls and in our previous fMRI study on persons with Parkinson’s disease (Vikene et al., 2019).

In addition, at uncorrected levels, the more salient position of the omission showed a higher activation in the ACC, consistent with previous research implicating the ACC in saliency detection (Menon, 2011) (Supplementary Figure 1 and Supplementary Table 1).

On a theoretical level, our findings can be interpreted as manifestations of crucial principles in predictive processing frameworks, where repetition suppression (Friston, 2005; Baldeweg, 2006; Auksztulewicz and Friston, 2016) can be interpreted as model building; maintenance (Brechmann et al., 2007; Kumar et al., 2016) as a “representation unit” (Clark, 2013); and omission activation (Raij et al., 1997; Wacongne et al., 2011)—an internal response based solely on expectancy or prediction (Jongsma et al., 2005)—as a prediction error (Friston, 2005). Despite the limitations in this study, we will claim that it robustly shows results in line with previous findings on repetition suppression and omission activation—and the perception of musical rhythms in general—and that our paradigm, concretely or abstractly, can be used as a starting point for more refined studies of predictive processing mechanisms for the perception of musical rhythms.

Our study successfully replicated previous findings for repetition suppression and omission activation and shows that tailored musical stimuli can be used in an fMRI setting to robustly investigate such neural phenomena, even with a limited number of participants (n = 10). Importantly, our findings show a clear separation between repetition suppression and prediction error activation, and additionally indicate that representational maintenance activates areas different from those deactivated during repetition suppression. While listening and subsequent repetition suppression were located mainly in anterior parts of the superior temporal gyrus, prediction errors (omission activation) were clearly separated from these areas and located mainly in posterior parts of the MTG. Representational maintenance activated SMA, caudate nucleus, anterior insula, frontal areas, and middle and posterior cingulate cortex, potentially showing persistent representational maintenance activity in areas separate from initial listening (encoding), repetition suppression (attenuation), and prediction error (omission).

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Regional Committee for Medical and Health Research Ethics, REK-VEST (2014/1915). The patients/participants provided their written informed consent to participate in this study.

UF, KS, and KV contributed in the conception of the research project, did statistical analysis, and contributed to the final version of the manuscript. UF recruited the participants, organized the study, and wrote the first draft. KV designed the paradigm. All authors contributed to the article and approved the submitted version.

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.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.