Language experience shapes predictive coding of rhythmic sound sequences

According to predictive coding theories of perception, sensory processes and perceptual decisions are described as a process of inference, which is strongly shaped by prior knowledge and predictions (Clark, 2013; Friston, 2005; Rao and Ballard, 1999). Predictions can be derived from different sources of information, forming a hierarchical predictive system (de Lange et al., 2018). Each level of the predictive hierarchy houses an internal model encoding prior information about the structure of the external environment. When a prediction is violated, the prediction error is computed and used to adjust the corresponding prior and internal model. This results in a constantly evolving system that generates and refines predictions based on incoming sensory input and prior experience.

In the auditory domain, great progress in the understanding of the predictive capabilities of the auditory system has been made using the oddball design and its variations (see Heilbron and Chait, 2018, for a review). In these designs, participants are usually presented with sequences of tones encoding a certain rule that is then violated by a ‘deviant’ event. Such deviants elicit a sharp evoked response in the electroencephalography (EEG) signal which has been defined as ‘mismatch negativity’ (MMN). The MMN peaks at about 0.100–0.250 s from stimulus onset and exhibits enhanced intensity over secondary temporal, central, and frontal areas of topographic scalp maps (Sams et al., 1985; Garrido et al., 2009). Within the predictive coding framework, the MMN is putatively considered an index of cortical prediction error.

Functionally, the mechanism underlying the MMN operates over both conscious and preconscious memory representations. MMN responses to auditory violations are observable when the participant is not paying attention to the auditory task and have been reported even in states of sleep (Sallinen et al., 1994; Sculthorpe et al., 2009; Strauss et al., 2015) and coma (Fischer et al., 2000). Given its automatic nature, the anticipatory mechanism underlying the MMN has been suggested to reflect a form of ‘primitive intelligence’ in the auditory cortex (Näätänen et al., 2001). In the present study, we show that life-long experience with a spoken language can shape this automatic anticipatory mechanism.

Experimental studies using the oddball design and its derivations have been important to unveil the sensitivity of the auditory predictive system to local statistical regularities and transition probabilities (Heilbron and Chait, 2018). However, these studies have primarily examined so-called contextual (or short-term) predictive signals. These predictions are usually based on rules acquired in the context of an experimental task – that is, rules linked to short-term memory – and have a short-lived impact on sensory processing. Yet, one core assumption of current predictive coding models is that the brain also deploys predictions based on long-term memory representations (Seriès and Seitz, 2013; Yon et al., 2019; Teufel and Fletcher, 2020). Such long-term predictions may emerge via learning of regularities and co-occurring patterns that are relatively stable throughout the lifespan of an organism. Because arising over long timescales, these experiential priors become encoded into the tuning properties of sensory cortices, forming a computational constraint on bottom–up sensory processing (Teufel and Fletcher, 2020).

Long-term priors may have long-lasting effects on perception. One example from the visual domain is the systematic bias in humans toward the perception of cardinal orientations (Girshick et al., 2011). This bias has been linked to the presence of a long-term prior that mirrors the statistics of the visual environment, that is, the preponderance of cardinal orientations in visual input (Girshick et al., 2011). Monkey studies on visual processing have shown that the visual system employs long-term priors to generate long-term predictions of incoming data (Meyer and Olson, 2011). Yet, whether similar predictive coding schemes subserve cortical computation in the human auditory system remains unsettled.

Here, we take a cross-linguistic approach to test whether the auditory system generates long-term predictions based on life-long exposure to auditory regularities, using rules that extend beyond those acquired in the recent past. Currently, one critical behavioral example of the effect of long-term experience on auditory perception is the influence of language on rhythmic grouping (Iversen et al., 2008; Molnar et al., 2016): Sequences of two tones alternating in duration are usually perceived by speakers of functor-initial languages (e.g., Spanish, English) as repetition of short–long groups separated by a pause, while speakers of functor-final languages (e.g., Basque, Japanese) report a bias for the opposite long–short grouping pattern. This perceptual effect has been linked to the co-occurrence statistics underlying the word-order properties of these languages. Specifically, the effect has been proposed to depend on the quasi-periodic alternation of short and long auditory events in the speech signal – reported in previous acoustic analyses (Molnar et al., 2016) – which reflect the linearization of function words (e.g., articles, prepositions) and content words (e.g., nouns, adjectives, verbs). In functor-initial languages, like English or Spanish, short events (i.e., function words; e.g., un, a) normally combine with long ones (i.e., content words; e.g., ordenador, computer) to form ‘short–long’ auditory chunks (Figure 1A). In contrast, in functor-final languages like Japanese and Basque, short events (i.e., function words; e.g., bat, a) normally follow long ones (i.e., content words; e.g., ordenagailu, computer), resulting in ‘long–short’ phrasal units (Figure 1B). Regular exposure to such language-specific phrasal structures has been proposed to underlie the automatic grouping biases of non-linguistic sounds (Iversen et al., 2008; Molnar et al., 2016), suggesting the presence of an auditory ‘duration prior’ that mirrors the word-order and prosodic properties of a given language.

Figure 1

Download asset Open asset

We hypothesize that the auditory system uses the proposed ‘duration prior’ as an internal model to generate long-term predictions about incoming sound sequences. In predictive coding terms, our hypothesis posits that the human auditory system upweights neural activity toward the onset of certain high-level events, based on the statistics of a given language.

To test this hypothesis, two groups of Basque (n = 20) and Spanish (n = 20) dominant participants were presented with 30 s rhythmic sequences of two tones alternating in duration at fixed inter-stimulus intervals (Figure 1C), while magnetoencephalography (MEG) was monitoring their cortical activity. To measure prediction error, random omissions of long and short tones were introduced in each sequence. Omission responses allow to examine the presence of putative error signals decoupled from bottom–up sensory input, offering a critical test for predictive coding (Walsh et al., 2020; Heilbron and Chait, 2018).

If, in line with our hypothesis, the human auditory system uses long-term linguistic priors as an internal model to predict incoming sounds, the following predictions ensue. The omission of a long tone should represent the violation of two predictions in the Basque, but not in the Spanish group: a short-term prediction based on the statistics of the previous stimuli (i.e., a prediction about a new tone), and a long-term prediction based on the statistics of the Basque’s phrasal structure (i.e., a prediction about a new phrasal chunk). Consequently, such an omission response should lead to a larger prediction error in the Basque compared to the Spanish group (Figure 1E). An orthogonally opposite pattern is expected when the deviant event is reflected in the omission of a short tone (Figure 1E).

The expectation that stronger error responses would be elicited by the omission of the first element rather than the second element of a perceptual chunk (‘long’ for the Basque, ‘short’ for the Spanish group) is primarily based on previous work on rhythm and music perception (e.g., Ladinig et al., 2009; Bouwer et al., 2016; Brochard et al., 2003; Potter et al., 2009). These studies have shown that the amplitude of evoked responses is larger when deviants occur at the ‘start’ of a perceptual group and decline toward the end of the chunk, suggesting that the auditory system generates predictions about the onset of higher-level, internally formed auditory chunks.

We tested the predictions above against a control condition having the same alternation design as the experimental condition, but with the two tones alternating in pitch instead of duration (Figure 1D). Here, no difference between groups is expected, as both groups should rely on short-term, but not long-term priors (Figure 1F). Finally, we performed reconstruction of cortical sources to identify the regions supporting long-term auditory priors.

We first examined the responses evoked by the omissions of tones. MEG responses time-locked to the onset of the tones and omissions from the Basque and Spanish dominant groups were pulled together and compared via cluster-based permutation test analysis (Maris and Oostenveld, 2007). Cluster analysis was performed over several time points to identify spatiotemporal clusters of neighboring sensors where the two conditions differ (Methods). This analysis revealed an early effect of omission responses arising around 0.100 s from deviance onset (p < 0.0001), including several channels over the entire scalp (Figure 2A, B). The latency and topographical distribution of the effect resemble the one elicited by a classical mismatch response, with strong activations over left and right temporal regions (Figure 2A, B). A smaller cluster (p = 0.043) with lower amplitude was also detected in an earlier time window (~0.030–0.050 s post-stimulus/omission onset). This cluster primarily includes left centro-temporal channels. The directionality of the effect is the same as the later cluster, that is, larger responses to omissions compared to tones (Figure 2—figure supplement 1). This finding aligned with previous reports showing that the omission of an expected tone in a regular sequence of sounds generates larger event-related fields (ERFs) than an actual tone (e.g., Yabe et al., 1997; Raij et al., 1997).

Figure 2 with 2 supplements see all

Download asset Open asset

Our main question was on the presence of long-term predictions induced by the linguistic background of the participants during the processing of simple binary auditory sequences. To assess this, we tested for the presence of an interaction effect between the linguistic background of the participants (Basque, Spanish) and the type of tone omitted (long, short) in modulating the amplitude of the MMN. Omission MMN responses in this analysis were calculated by subtracting the ERF elicited by a given type of tone (i.e., long, short) from its corresponding omission (Garrido et al., 2009). A cluster-based permutation test was used to test the interaction effect between omission type (long vs short) and the linguistic background of the participants (Basque vs Spanish). As the cluster-based permutation test is designed to compare two conditions at a time, we tested for an interaction effect by subtracting the MMN elicited by the omission of a long tone from the MMN elicited by the omission of a short tone for each participant, and then compared the resulting differences between groups.

For this and the following contrasts, we selected a predefined time window of interest between 0.100 and 0.250 s, which covers the typical latency of the MMN (Näätänen et al., 2007; Garrido et al., 2009). Cluster analysis revealed a significant interaction (p = 0.03), which was particularly pronounced over left frontotemporal channels (see Figure 2C). To unpack the interaction, we averaged data samples for each participant and condition over the channels belonging to the significant cluster and time points of interest. This resulted in two ERFs for each participant, one for each type of omission MMN (long, short). We then compared ERFs for each omission type between the two groups using a one-sided independent sample t-test, testing the hypothesis that participants deploy long-term expectations about the onset of abstract language-like grouping units (Figure 1E). Specifically, we compared (1) the omission MMN responses generated by the omission of long tones in the Basque vs the Spanish group, and (2) the MMN responses generated by the omission of short tones in the Basque vs the Spanish group. Consistent with the hypothesis, we found that omissions of long tones generated a larger MMN response in the Basque compared Spanish group (t(38) = 2.22; p = 0.03 FDR-corrected; d = 0.70), while the omission of short tones generated a larger omission MMN in the Spanish compared to Basque group (t(38) = −2; p = 0.03 FDR-corrected; d = 0.63) (Figure 2D, E). Notice that, given the structure of our design, a between-group comparison (e.g., comparing the ERF between the Basque and Spanish groups) is more suited to test our hypothesis than a within-group comparison (e.g., comparing the ERF evoked by the omission of long vs short tones within the Basque group), as the pre-stimulus baseline activity is virtually identical across conditions only in the between-group contrast.

To further assess that the interaction was driven by the hypothesized long-term ‘duration prior’, the same analysis pipeline was applied to the data from the control condition. Here, no significant omission type × language background interaction was detected (no cluster with p < 0.05). To further check that no interaction was present in the control study, we averaged the data samples over the channels and time points in which we detected a significant interaction in the test condition and ran an independent sample t-test by comparing MMN responses elicited by the omission of high- and low-frequency tones in both groups (Methods). Even within this subset of channels, no between-group difference was detected between MMN responses evoked by omissions of high- (t(38) = −1.6; p = 0.12 FDR-corrected; d = −0.51), and low-frequency tones (t(38) = −1.1; p = 0.55 FDR-corrected; d = −0.04) (Figure 2F, G).

A linearly constrained minimum variance (LCMV) beamformer approach (Van Veen et al., 1997) was used to reconstruct the cortical sources of the MEG signal. We first focused on the source activity underlying the effect of omission (Figure 2A, B). Source activity was calculated for the epochs averaged in the 0.100–0.250 s interval for both tones and omissions. In order to isolate regions underlying the effect of omission, whole-brain maps for the ratio of the source activity associated with omission responses and tones were created (Figure 3A). A large network of regions showed a stronger response to omissions compared to auditory tones, including the bilateral inferior frontal gyri, premotor cortices, angular gyri, as well as the superior temporal gyri (Figure 3A).

Figure 3

Download asset Open asset

We then examined the cortical origin of the interaction effect that emerged at the sensor level (Figure 2C). In line with the MMN analysis at the sensor level, we focused on the source activity of the difference between omissions and tones for each omission type (short, long) and language group (Basque, Spanish). This analysis was aimed at identifying the cortical origin of the hypothesized long-term predictions. We hypothesized that such long-term priors might be linked to long-term experience with the rhythmic properties of the two languages. As such, they would be expected to arise around early auditory areas, such as the superior temporal gyrus (STG). An alternative hypothesis is that these priors are linked to the abstract syntactic structure of the two languages. Under this account, long-term predictions would be generated via long-range feedback from regions associated with syntactic processing, such as the left inferior frontal gyrus (IFG) (Ben Shachar et al., 2003). To disentangle these possibilities, we performed an analysis on three regions of interest (ROIs, based on the Brainnetome atlas, Fan et al., 2016) within the left STG, (Brodmann areas (BAs) 41/42 of the auditory cortex, rostral portion of BA 22, caudal portion of BA 22), and three ROIs within the left IFG (dorsal, ventral, and opercular portions of BA 44). We restricted our analysis to the left hemisphere only, which is where the significant interaction effect emerged at the sensor level (Figure 2C). We first performed a cluster-based permutation analysis in the whole left STG and the left IFG in the 0.100–0.250 s time interval, testing for the presence of an interaction between omission type and language group. This strategy is similar to the one performed at the sensor level, but is more time-sensitive as individual time points in the 0.100–0.250 s time interval are considered. No significant interaction was detected in the left IFG (no cluster with p < 0.05), while two temporally distinct effects emerged in the left STG: an early effect arising in the 0.110–0.145 s time interval (p = 0.01) and a later effect in the 0.200–0.220 s (p = 0.04) (Figure 3B–D). It is possible that these two clusters reflect different temporal responses of the two groups to long and short omissions. To better understand the nature of the interaction effect within the left STG, pairwise comparisons were performed on each ROI and cluster using one-sided independent sample t-tests, following the same contrasts that were performed at the sensor level. Despite the high redundancy of activity across neighboring brain regions due to MEG source-reconstruction limitations (Bourguignon et al., 2018), we were interested in verifying whether the more robust interaction effect, both in magnitude and reliability, was emerging in primary (BA 41/42) or associative (BAs 22) auditory regions. Given the limitations of the cluster-based statistical approach in the definition of the timing of significant effects (Sassenhagen and Draschkow, 2019), peak activity within each temporal cluster in the STG was selected and used for the pairwise comparisons on each ROI (early peak: 0.120 s; late peak: 0.210 s). Pairwise comparisons showed that the omission of long tones generated larger responses on the latency of the early peak (0.120 s) in the Basque compared to the Spanish dominant group over BA 41/42 and BA 22c BA 41/42 (t(38) = 1.88, p = 0.03, d = 0.59); BA 22c (t(38) = 2.04, p = 0.02, d = 0.64), while no difference emerged for later peak responses (0.210 s) (BA 41/42: t(38) = 0.59, p = 0.27, d = 0.32; BA 22c: t(38) = −1.14, p = 0.87, d = −0.36) (Figure 3C). On the contrary, the omission of short tones led to stronger responses over the later peak latency (0.210 s) in the Spanish compared to the Basque group in BA 41/42 (t(38) = −1.98, p = 0.02, d = −0.62), while no difference was observed on the latency of the earlier peak (BA 41/42: t(38) = 0.32, p = 0.62, d = 0.10) (Figure 3C). No significant effect emerged in BA22r. Overall, these findings suggest that segments of the left STG, in particular BA41/42, exhibit distinct sensitivity to omission responses in the two groups. The Basque group shows larger responses to the omission of long tones at an earlier time interval, whereas the Spanish group displays increased responses to the omission of short tones at a later time interval (Figure 3C).

Besides documenting an interaction, which was the analysis of interest of our study, we also searched for a main effect of omission type. A cluster-based permutation test was used to compare the MMN responses elicited by omissions of long tones and omissions of short tones averaged across the two groups. The results showed that long-tone omissions generated a larger omission MMN response than short-tone omissions (p = 0.03), with the cluster including several frontal channels (Figure 2—figure supplement 2). This effect was consistent in the Basque (p = 0.003), but not in the Spanish group (no clusters with p < 0.05). No main effect of language background was detected (no clusters with p < 0.05) (Figure 2—figure supplement 2). In the control condition, neither a main effect of omission type nor an effect of language background was detected (no clusters with p < 0.05).

By comparing MEG data from native speakers of functor-initial (i.e., Spanish) and functor-final languages (i.e., Basque) listening to simple binary sequences of tones with occasional violations, we show that experience with a given language can shape a very simple aspect of human perception, such as the neural processing of a binary rhythmic sound. This finding suggests that the human auditory system uses structural patterns of their native language to generate predictive models of non-linguistic sound sequences. This result highlights the presence of an active predictive system that relies on natural sound statistics learned over a lifetime to process incoming auditory input.

We first looked at the responses generated by sound omissions. In line with previous reports, we found that omissions of expected tones in an auditory sequence generate a sharp response in the ERF. Such omission responses have been suggested to reflect pure error signals decoupled from bottom–up sensory input (Hughes et al., 2001; Wacongne et al., 2011). On the other hand, other studies have proposed that omission responses could reflect pure predictions (Bendixen et al., 2009; SanMiguel et al., 2013). While the exact nature of such responses is currently debated and likely dependent on factors such as task and relevance, the latency and topography of the omission response in our data resemble those evoked by a classical mismatch response (Figure 2A, B). Analysis of cortical sources also supports this interpretation. Indeed, source activity associated with omissions leads to stronger responses compared to tones over a distributed network of regions, including the bilateral inferior frontal gyri, premotor areas, angular gyri, and right STG. Since sensory predictive signals primarily arise in the same regions as the actual input, the activation of a broader network of regions in omission responses compared to tones suggests that omission responses reflect, at least in part, prediction error signals.

Importantly, we showed that when an unexpected omission disrupts a binary sequence of sounds, the amplitude of the omission MMN varies orthogonally depending on the speaker’s linguistic background. Omissions of long auditory events generate a larger omission MMN in the Basque compared to the Spanish group, while omissions of short sounds lead to a larger omission MMN responses in the Spanish compared to the Basque group. We hypothesized that this effect is linked to a long-term ‘duration prior’ originating from the acoustic properties of the two languages, specifically from the alternation of short and long auditory events in their prosody. Importantly, no difference between groups was detected in a control task in which tones alternate in frequency instead of duration, suggesting that the reported effect was driven by the hypothesized long-term linguistic priors instead of uncontrolled group differences.

It is important to note that both Spanish and Basque speakers are part of the same cultural community in Northern Spain. These languages share almost the same phonology and orthography. However, Basque is a non-Indo-European language (an isolated language) with no typological relationship with Spanish. It is thus very unlikely that the current findings are driven by cultural factors that are not language-specific (e.g., exposure to different musical traditions or educational and writing systems).

How would such long-term priors arise? One possible interpretation is that long-term statistical learning of the duration prosodic pattern of native language shapes the tuning properties of early auditory regions, affecting predictive coding at early stages. Such language-driven tuning is arguably important for reducing the prediction error during the segmentation of speech material into phrasal units, as it allows the auditory system to generate a functional coding scheme, or auditory template, against which the incoming speech input can be parsed. Such an auditory template is likely to be recycled by the auditory system to build top–down predictive models of non-linguistic auditory sequences.

The idea that the auditory system implements long-term predictions based on the prosodic structure of the native language could explain the previously reported behavioral influence of language experience on rhythmic grouping (Iversen et al., 2008; Molnar et al., 2016): when listening to sequences of two tones alternating in duration, like those used in the present study, speakers of functor-initial languages report to perceive the rhythmic sequences as a repetition of ‘short–long’ units, while speakers of functor-final languages have the opposite ‘long–short’ grouping bias (Iversen et al., 2008; Molnar et al., 2016). Despite lacking a direct behavioral assessment (but see Molnar et al., 2016, for related behavioral evidence), our results indicate that this perceptual grouping effect can be explained within a predictive coding framework that incorporates long-term prior knowledge into perceptual decisions. Under such an account, the auditory system internalizes the statistics underlying the prosodic structure of language and uses this knowledge to make long-term predictions of incoming sound sequences. Such long-term predictions would bias auditory processing at early, rather than later decision-making stages, affecting how rhythmic sounds are experienced.

Our work capitalized on a specific aspect of natural sound acoustic – the duration pattern in Basque and Spanish prosody – as a testbed to assess the presence of long-term priors in the auditory system. Despite our work being restricted to this specific feature, it is likely that the auditory system forms several other types of long-term priors using the spectrotemporal features that dominate the auditory environment. Support for this claim comes from (1) studies showing that the human auditory system uses the statistics underlying the acoustic structure of speech and music to form perceptual grouping decisions (Młynarski and McDermott, 2019); and (2) behavioral experiments reporting off-line effects of language experience on auditory perception based on different acoustic features (Liu et al., 2023). For instance, native speakers of languages in which pitch carries phonemically meaningful information (i.e., tone languages, e.g., Mandarin Chinese) benefit from a behavioral advantage in non-linguistic pitch discrimination tasks as compared to speakers of non-tone languages like English (Bidelman et al., 2013). Similarly, speakers of languages that use duration to differentiate between phonemes (e.g., Finnish, Japanese) manifest an enhanced ability to discriminate the duration of non-linguistic sounds (Tervaniemi et al., 2006). Our results, in conjunction with these studies, suggest that the auditory system forms long-term priors and predictions over development, using the co-occurrences that dominate the natural stimulus statistics. Yet, our results leave open the question of whether these long-term priors can be updated during adulthood, following extensive exposure to new statistical dependencies. This can be tested by exposing adult speakers to natural sounds encoding rules that ‘violate’ the long-term prior (e.g., a language with opposite prosodic structure) and exploring the effects of such short-term exposure to behavioral and neural performance.

One potential alternative to the conjecture that the ‘duration prior’ is linked to the spectro-temporal features of a language is that the prior depends on abstract syntactic/word-order rules. This latter account would predict that violations of long-term predictions in our study would lead to larger error responses in regions sensitive to syntactic variables, such as the left IFG (Ben Shachar et al., 2003). Instead, the former account would predict that violations of long-term predictions elicit stronger responses in early left-lateralized auditory regions, which are putatively associated with early speech processing (Bhaya-Grossman and Chang, 2022). The reconstruction of cortical sources associated with the omission of short and long tones in the two groups showed that an interaction effect mirroring the one at the sensor level was present in the left STG, but not in the left IFG (Figure 3B–D). Pairwise comparisons within different ROIs of the left STG indicated that the interaction effect was stronger over primary (BA 41/42) rather than associative (BAs 22) portions of the auditory cortex. Overall, these results suggest that the ‘duration prior’ is linked to the acoustic properties of a given language rather than its syntactic configurations.

Our results are in line with predictive coding models stating that predictions are organized hierarchically. When two predictive signals, one short-term and one long-term are violated, the amplitude of the prediction error is larger compared to a scenario in which only one short-term prediction is violated. This result complements previous studies using the local–global design showing that the same deviancy presented in different contexts gives rise to different error signals, such as the MMN and the P3 (Bekinschtein et al., 2009; Wacongne et al., 2011). These studies provide empirical evidence that predictive coding of auditory sequences is organized at different functional levels, with early sensory regions using transition probabilities to generate expectations about the present, and frontal and associative regions inferring the global structure of an auditory event. Our results extend this work by providing direct evidence for the presence of a system in the auditory cortex that uses long-term natural sound statistics to generate long-term predictions. This interpretation is also supported by the reconstruction of cortical sources. Indeed, while the overall omission effect is larger in the right hemisphere (Figures 2A and 3A), the interaction effect arises in the left hemisphere (Figures 2C and 3B, C). This finding further suggests that distinct cortical systems, supporting different predictive models, underlie the generation of the omission MMN.

Our findings are also consistent with more recent predictive coding models incorporating the idea of ‘stubborn’ predictive signals – that is, predictions resilient to model updates. Unlike short-term expectations, long-term predictions are usually implemented as a computational constraint on input data, thus being largely unaffected by short-term experience (Teufel and Fletcher, 2020). In our study, the deployment of long-term predictions does not represent an effective coding strategy to perform the task. Yet, listeners still seem to assign different weights to incoming data, using a ‘default’ predictive coding scheme that resembles the segmentation strategy used to parse speech material. Why should a neural system rely on such stubborn priors even when irrelevant to solving a given perceptual task? One possibility is that implementing stable priors as a constraint on perception is computationally less expensive in terms of metabolic costs than recalibrating cortical internal models anew based on any type of novel experience. Another possibility is that relying on unchanging predictive schemes helps the system to form coherent models in front of environmental contingencies, thus reflecting an effective computational strategy for the reduction of the long-term prediction error. Defining how stubborn predictions emerge during learning and what their computational role is represents an important challenge to understanding the role of prior experience in perceptual inference.

We also reported a main effect of omission type, indicating that the MMN generated by the omission of a long tone was generally larger compared to that generated by the omission of a short one. Because such group effect was consistent only in the Basque group, it is possible that it merely reflects a larger sensitivity of the auditory system of this group to the omission of long events, in line with the interaction reported above. Alternatively, this effect could be driven by the fact that, during language processing, major predictive resources are invested in predicting the onset of long events, compared to short ones, as the formers usually refer to content words that is, semantically relevant events. Consequently, the auditory system may apply a similar predictive scheme also during the processing of non-linguistic sound sequences, independently of language background. Independently on the interpretation, the lack of a main effect of omission type in the control condition suggests that the long omission effect is driven by experience with the native language.

Our results also refine previous studies showing modulatory effects of (long-term) musical expertise on the MMN (e.g., Vuust et al., 2005; Vuust et al., 2009). These studies indicate that responses to violation during auditory rhythm perception are larger when the listener is an expert musician compared to a non-musician, pointing to the role of long-term auditory experience in shaping early predictive mechanisms. In our study, we manipulated long-term prediction orthogonally, with clear-cut predictions about the effect of language experience on early auditory predictive processing. Our results thus provide direct evidence for the presence of an active system in the auditory cortex that uses long-term priors to constrain information processing of incoming auditory stimuli.

Data and codes for reproducing the analysis and figures are publicly available via the Open Science Framework (OSF): https://osf.io/6jep8/.

1. 1. Morucci P
   2. Molinaro N

   (2023) Open Science Framework

   ID osf.io/6jep8. Language experience shapes predictive coding of rhythmic sound sequences.

   https://osf.io/6jep8/

1. 1. Allen M
   2. Poggiali D
   3. Whitaker K
   4. Marshall TR
   5. van Langen J
   6. Kievit RA

   (2019) Raincloud plots: a multi-platform tool for robust data visualization

   Wellcome Open Research 4:63.

   https://doi.org/10.12688/wellcomeopenres.15191.2

   • PubMed
   • Google Scholar
2. 1. Bekinschtein TA
   2. Dehaene S
   3. Rohaut B
   4. Tadel F
   5. Cohen L
   6. Naccache L

   (2009) Neural signature of the conscious processing of auditory regularities

   PNAS 106:1672–1677.

   https://doi.org/10.1073/pnas.0809667106

   • Google Scholar
3. 1. Bendixen A
   2. Schröger E
   3. Winkler I

   (2009) I heard that coming: event-related potential evidence for stimulus-driven prediction in the auditory system

   The Journal of Neuroscience 29:8447–8451.

   https://doi.org/10.1523/JNEUROSCI.1493-09.2009

   • PubMed
   • Google Scholar
4. 1. Ben Shachar M
   2. Hendler T
   3. Kahn I
   4. Ben Bashat D
   5. Grodzinsky Y

   (2003) The neural reality of syntactic transformations: evidence from functional magnetic resonance imaging

   Psychological Science 14:433–440.

   https://doi.org/10.1111/1467-9280.01459

   • PubMed
   • Google Scholar
5. 1. Bhaya-Grossman I
   2. Chang EF

   (2022) Speech computations of the human superior temporal gyrus

   Annual Review of Psychology 73:79–102.

   https://doi.org/10.1146/annurev-psych-022321-035256

   • PubMed
   • Google Scholar
6. 1. Bidelman GM
   2. Hutka S
   3. Moreno S

   (2013) Tone language speakers and musicians share enhanced perceptual and cognitive abilities for musical pitch: evidence for bidirectionality between the domains of language and music

   PLOS ONE 8:e60676.

   https://doi.org/10.1371/journal.pone.0060676

   • PubMed
   • Google Scholar
7. 1. Bourguignon M
   2. Molinaro N
   3. Wens V

   (2018) Contrasting functional imaging parametric maps: The mislocation problem and alternative solutions

   NeuroImage 169:200–211.

   https://doi.org/10.1016/j.neuroimage.2017.12.033

   • PubMed
   • Google Scholar
8. 1. Bouwer FL
   2. Werner CM
   3. Knetemann M
   4. Honing H

   (2016) Disentangling beat perception from sequential learning and examining the influence of attention and musical abilities on ERP responses to rhythm

   Neuropsychologia 85:80–90.

   https://doi.org/10.1016/j.neuropsychologia.2016.02.018

   • PubMed
   • Google Scholar
9. 1. Brochard R
   2. Abecasis D
   3. Potter D
   4. Ragot R
   5. Drake C

   (2003) The “ticktock” of our internal clock: direct brain evidence of subjective accents in isochronous sequences

   Psychological Science 14:362–366.

   https://doi.org/10.1111/1467-9280.24441

   • PubMed
   • Google Scholar
10. 1. Clark A

    (2013) Whatever next? Predictive brains, situated agents, and the future of cognitive science

    The Behavioral and Brain Sciences 36:181–204.

    https://doi.org/10.1017/S0140525X12000477

    • PubMed
    • Google Scholar
11. Book

    1. Cohen J

    (2013)

    Statistical Power Analysis for the Behavioral Sciences

    Routledge.

    • Google Scholar
12. 1. de Lange FP
    2. Heilbron M
    3. Kok P

    (2018) How do expectations shape perception?

    Trends in Cognitive Sciences 22:764–779.

    https://doi.org/10.1016/j.tics.2018.06.002

    • Google Scholar
13. 1. Fan L
    2. Li H
    3. Zhuo J
    4. Zhang Y
    5. Wang J
    6. Chen L
    7. Yang Z
    8. Chu C
    9. Xie S
    10. Laird AR
    11. Fox PT
    12. Eickhoff SB
    13. Yu C
    14. Jiang T

    (2016) The human brainnetome atlas: A new brain atlas based on connectional architecture

    Cerebral Cortex 26:3508–3526.

    https://doi.org/10.1093/cercor/bhw157

    • PubMed
    • Google Scholar
14. 1. Fischer C
    2. Morlet D
    3. Giard MH

    (2000) Mismatch negativity and N100 in comatose patients

    Audiology & Neuro-Otology 5:192–197.

    https://doi.org/10.1159/000013880

    • PubMed
    • Google Scholar
15. 1. Friston KJ
    2. Holmes AP
    3. Worsley KJ
    4. Poline JP
    5. Frith CD
    6. Frackowiak RSJ

    (1994) Statistical parametric maps in functional imaging: A general linear approach

    Human Brain Mapping 2:189–210.

    https://doi.org/10.1002/hbm.460020402

    • Google Scholar
16. 1. Friston K

    (2005) A theory of cortical responses

    Philosophical Transactions of the Royal Society B 360:815–836.

    https://doi.org/10.1098/rstb.2005.1622

    • Google Scholar
17. 1. Garrido MI
    2. Kilner JM
    3. Stephan KE
    4. Friston KJ

    (2009) The mismatch negativity: a review of underlying mechanisms

    Clinical Neurophysiology 120:453–463.

    https://doi.org/10.1016/j.clinph.2008.11.029

    • PubMed
    • Google Scholar
18. 1. Girshick AR
    2. Landy MS
    3. Simoncelli EP

    (2011) Cardinal rules: visual orientation perception reflects knowledge of environmental statistics

    Nature Neuroscience 14:926–932.

    https://doi.org/10.1038/nn.2831

    • PubMed
    • Google Scholar
19. 1. Heilbron M
    2. Chait M

    (2018) Great expectations: Is there evidence for predictive coding in auditory cortex?

    Neuroscience 389:54–73.

    https://doi.org/10.1016/j.neuroscience.2017.07.061

    • PubMed
    • Google Scholar
20. 1. Hughes HC
    2. Darcey TM
    3. Barkan HI
    4. Williamson PD
    5. Roberts DW
    6. Aslin CH

    (2001) Responses of human auditory association cortex to the omission of an expected acoustic event

    NeuroImage 13:1073–1089.

    https://doi.org/10.1006/nimg.2001.0766

    • PubMed
    • Google Scholar
21. 1. Iversen JR
    2. Patel AD
    3. Ohgushi K

    (2008) Perception of rhythmic grouping depends on auditory experience

    The Journal of the Acoustical Society of America 124:2263–2271.

    https://doi.org/10.1121/1.2973189

    • PubMed
    • Google Scholar
22. 1. Jung TP
    2. Makeig S
    3. Westerfield M
    4. Townsend J
    5. Courchesne E
    6. Sejnowski TJ

    (2000) Removal of eye activity artifacts from visual event-related potentials in normal and clinical subjects

    Clinical Neurophysiology 111:1745–1758.

    https://doi.org/10.1016/s1388-2457(00)00386-2

    • PubMed
    • Google Scholar
23. 1. Ladinig O
    2. Honing H
    3. Háden G
    4. Winkler I

    (2009) Probing attentive and preattentive emergent meter in adult listeners without extensive music training

    Music Perception 26:377–386.

    https://doi.org/10.1525/mp.2009.26.4.377

    • Google Scholar
24. 1. Liu J
    2. Hilton CB
    3. Bergelson E
    4. Mehr SA

    (2023) Language experience predicts music processing in a half-million speakers of fifty-four languages

    Current Biology 33:1916–1925.

    https://doi.org/10.1016/j.cub.2023.03.067

    • PubMed
    • Google Scholar
25. Conference

    1. Makeig S
    2. Bell A
    3. Jung TP
    4. Sejnowski TJ

    (1995)

    Independent component analysis of electroencephalographic data

    Advances in Neural Information Processing Systems.

    • Google Scholar
26. 1. Maris E
    2. Oostenveld R

    (2007) Nonparametric statistical testing of EEG- and MEG-data

    Journal of Neuroscience Methods 164:177–190.

    https://doi.org/10.1016/j.jneumeth.2007.03.024

    • PubMed
    • Google Scholar
27. 1. Meyer T
    2. Olson CR

    (2011) Statistical learning of visual transitions in monkey inferotemporal cortex

    PNAS 108:19401–19406.

    https://doi.org/10.1073/pnas.1112895108

    • Google Scholar
28. 1. Młynarski W
    2. McDermott JH

    (2019) Ecological origins of perceptual grouping principles in the auditory system

    PNAS 116:25355–25364.

    https://doi.org/10.1073/pnas.1903887116

    • Google Scholar
29. 1. Molnar M
    2. Carreiras M
    3. Gervain J

    (2016) Language dominance shapes non-linguistic rhythmic grouping in bilinguals

    Cognition 152:150–159.

    https://doi.org/10.1016/j.cognition.2016.03.023

    • PubMed
    • Google Scholar
30. 1. Näätänen R
    2. Tervaniemi M
    3. Sussman E
    4. Paavilainen P
    5. Winkler I

    (2001) “Primitive intelligence” in the auditory cortex

    Trends in Neurosciences 24:283–288.

    https://doi.org/10.1016/s0166-2236(00)01790-2

    • PubMed
    • Google Scholar
31. 1. Näätänen R
    2. Paavilainen P
    3. Rinne T
    4. Alho K

    (2007) The mismatch negativity (MMN) in basic research of central auditory processing: a review

    Clinical Neurophysiology 118:2544–2590.

    https://doi.org/10.1016/j.clinph.2007.04.026

    • PubMed
    • Google Scholar
32. 1. Oostenveld R
    2. Fries P
    3. Maris E
    4. Schoffelen JM

    (2011) FieldTrip: Open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data

    Computational Intelligence and Neuroscience 2011:156869.

    https://doi.org/10.1155/2011/156869

    • PubMed
    • Google Scholar
33. 1. Potter DD
    2. Fenwick M
    3. Abecasis D
    4. Brochard R

    (2009) Perceiving rhythm where none exists: event-related potential (ERP) correlates of subjective accenting

    Cortex; a Journal Devoted to the Study of the Nervous System and Behavior 45:103–109.

    https://doi.org/10.1016/j.cortex.2008.01.004

    • PubMed
    • Google Scholar
34. 1. Raij T
    2. McEvoy L
    3. Mäkelä JP
    4. Hari R

    (1997) Human auditory cortex is activated by omissions of auditory stimuli

    Brain Research 745:134–143.

    https://doi.org/10.1016/s0006-8993(96)01140-7

    • PubMed
    • Google Scholar
35. 1. Rao RP
    2. Ballard DH

    (1999) Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects

    Nature Neuroscience 2:79–87.

    https://doi.org/10.1038/4580

    • PubMed
    • Google Scholar
36. 1. Sallinen M
    2. Kaartinen J
    3. Lyytinen H

    (1994) Is the appearance of mismatch negativity during stage 2 sleep related to the elicitation of K-complex?

    Electroencephalography and Clinical Neurophysiology 91:140–148.

    https://doi.org/10.1016/0013-4694(94)90035-3

    • PubMed
    • Google Scholar
37. 1. Sams M
    2. Paavilainen P
    3. Alho K
    4. Näätänen R

    (1985) Auditory frequency discrimination and event-related potentials

    Electroencephalography and Clinical Neurophysiology/Evoked Potentials Section 62:437–448.

    https://doi.org/10.1016/0168-5597(85)90054-1

    • Google Scholar
38. 1. SanMiguel I
    2. Widmann A
    3. Bendixen A
    4. Trujillo-Barreto N
    5. Schröger E

    (2013) Hearing silences: human auditory processing relies on preactivation of sound-specific brain activity patterns

    The Journal of Neuroscience 33:8633–8639.

    https://doi.org/10.1523/JNEUROSCI.5821-12.2013

    • PubMed
    • Google Scholar
39. 1. Sassenhagen J
    2. Draschkow D

    (2019) Cluster-based permutation tests of MEG/EEG data do not establish significance of effect latency or location

    Psychophysiology 56:e13335.

    https://doi.org/10.1111/psyp.13335

    • PubMed
    • Google Scholar
40. 1. Sculthorpe LD
    2. Ouellet DR
    3. Campbell KB

    (2009) MMN elicitation during natural sleep to violations of an auditory pattern

    Brain Research 1290:52–62.

    https://doi.org/10.1016/j.brainres.2009.06.013

    • PubMed
    • Google Scholar
41. 1. Seriès P
    2. Seitz AR

    (2013) Learning what to expect (in visual perception)

    Frontiers in Human Neuroscience 7:668.

    https://doi.org/10.3389/fnhum.2013.00668

    • PubMed
    • Google Scholar
42. 1. Strauss M
    2. Sitt JD
    3. King JR
    4. Elbaz M
    5. Azizi L
    6. Buiatti M
    7. Naccache L
    8. van Wassenhove V
    9. Dehaene S

    (2015) Disruption of hierarchical predictive coding during sleep

    PNAS 112:E1353–E1362.

    https://doi.org/10.1073/pnas.1501026112

    • Google Scholar
43. 1. Tervaniemi M
    2. Jacobsen T
    3. Röttger S
    4. Kujala T
    5. Widmann A
    6. Vainio M
    7. Näätänen R
    8. Schröger E

    (2006) Selective tuning of cortical sound-feature processing by language experience

    The European Journal of Neuroscience 23:2538–2541.

    https://doi.org/10.1111/j.1460-9568.2006.04752.x

    • PubMed
    • Google Scholar
44. 1. Teufel C
    2. Fletcher PC

    (2020) Forms of prediction in the nervous system

    Nature Reviews. Neuroscience 21:231–242.

    https://doi.org/10.1038/s41583-020-0275-5

    • PubMed
    • Google Scholar
45. 1. Van Veen BD
    2. van Drongelen W
    3. Yuchtman M
    4. Suzuki A

    (1997) Localization of brain electrical activity via linearly constrained minimum variance spatial filtering

    IEEE Transactions on Bio-Medical Engineering 44:867–880.

    https://doi.org/10.1109/10.623056

    • PubMed
    • Google Scholar
46. 1. Vuust P
    2. Pallesen KJ
    3. Bailey C
    4. van Zuijen TL
    5. Gjedde A
    6. Roepstorff A
    7. Østergaard L

    (2005) To musicians, the message is in the meter pre-attentive neuronal responses to incongruent rhythm are left-lateralized in musicians

    NeuroImage 24:560–564.

    https://doi.org/10.1016/j.neuroimage.2004.08.039

    • PubMed
    • Google Scholar
47. 1. Vuust P
    2. Ostergaard L
    3. Pallesen KJ
    4. Bailey C
    5. Roepstorff A

    (2009) Predictive coding of music--brain responses to rhythmic incongruity

    Cortex; a Journal Devoted to the Study of the Nervous System and Behavior 45:80–92.

    https://doi.org/10.1016/j.cortex.2008.05.014

    • PubMed
    • Google Scholar
48. 1. Wacongne C
    2. Labyt E
    3. van Wassenhove V
    4. Bekinschtein T
    5. Naccache L
    6. Dehaene S

    (2011) Evidence for a hierarchy of predictions and prediction errors in human cortex

    PNAS 108:20754–20759.

    https://doi.org/10.1073/pnas.1117807108

    • PubMed
    • Google Scholar
49. 1. Walsh KS
    2. McGovern DP
    3. Clark A
    4. O’Connell RG

    (2020) Evaluating the neurophysiological evidence for predictive processing as a model of perception

    Annals of the New York Academy of Sciences 1464:242–268.

    https://doi.org/10.1111/nyas.14321

    • PubMed
    • Google Scholar
50. 1. Yabe H
    2. Tervaniemi M
    3. Reinikainen K
    4. Näätänen R

    (1997) Temporal window of integration revealed by MMN to sound omission

    Neuroreport 8:1971–1974.

    https://doi.org/10.1097/00001756-199705260-00035

    • PubMed
    • Google Scholar
51. 1. Yon D
    2. de Lange FP
    3. Press C

    (2019) The predictive brain as a stubborn scientist

    Trends in Cognitive Sciences 23:6–8.

    https://doi.org/10.1016/j.tics.2018.10.003

    • Google Scholar

Author details

1. Piermatteo Morucci

   1. Department of Fundamental Neurosciences, University of Geneva, Geneva, Switzerland
   2. Basque Center on Cognition, Brain and Language, Donostia-San Sebastian, Spain

   Contribution

   Conceptualization, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing - original draft, Project administration, Writing - review and editing

   For correspondence

   piermatteomorucci@gmail.com

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-4972-0864

2. Sanjeev Nara

   1. Basque Center on Cognition, Brain and Language, Donostia-San Sebastian, Spain
   2. Mathematical Institute, Department of Mathematics and Computer Science, Physics, Geography, Liebig-Universität Gießen, Gießen, Germany

   Contribution

   Formal analysis, Methodology, Writing - review and editing

   Competing interests

   No competing interests declared

3. Mikel Lizarazu

   Basque Center on Cognition, Brain and Language, Donostia-San Sebastian, Spain

   Contribution

   Supervision, Visualization, Methodology

   Competing interests

   No competing interests declared

4. Clara Martin

   1. Basque Center on Cognition, Brain and Language, Donostia-San Sebastian, Spain
   2. Ikerbasque, Basque Foundation for Science, Bilbao, Spain

   Contribution

   Conceptualization, Supervision, Writing - review and editing

   Competing interests

   No competing interests declared

5. Nicola Molinaro

   1. Basque Center on Cognition, Brain and Language, Donostia-San Sebastian, Spain
   2. Ikerbasque, Basque Foundation for Science, Bilbao, Spain

   Contribution

   Conceptualization, Resources, Formal analysis, Supervision, Funding acquisition, Visualization, Methodology, Project administration, Writing - review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-7549-6042