Participants were twenty native speakers of German, recruited from the participant database of the Max Planck Institute (12 females, 8 males, average age 25 ± 5 years). The sample size was based on previous studies with similar designs (Colin et al., 2009). All participants were right-handed, with scores >90% on the Edinburgh Handedness Inventory (Oldfield, 1971). None of the participants reported a history of hearing or neurological problems and participated for monetary compensation (€ 10 per hour). The study was approved by the local Ethics Committee and in accordance with the declarations of Helsinki. Prior to the experiment, participants provided written informed consent and were informed about legal aspects of the study as well as data handling policies in written and spoken form.

Trochaic and iambic stimuli were disyllabic pseudowords, starting with the voiced velar stop [g] and followed by the round, back high vowel [u], i.e., “gugu.” First, complete pseudowords were recorded in the carrier-sentence “Er soll nun gugu sagen (he shall say gugu now),” with “gugu” either pronounced with a strong initial syllable (N = 10) or a strong final syllable (N = 10). Carrier-sentences and pseudowords were spoken by a phonetically trained female speaker and recorded with 44.1 kHz temporal and 16 bit amplitude resolution in a silent recording chamber of the Max-Planck-Institute for Empirical Aesthetics in Frankfurt (Germany). From the entire set of 20 recordings, we selected those gu-syllables that had the most comparable pitch changes between strong and weak versions and showed the least difference in intensity. We decided to use stimuli that approximate typical trochaic and iambic disyllabic words without differing too much in acoustic terms, for any change of acoustic properties would modulate the omission MMN. We arrived at four syllables from one trochaic and one iambic pseudoword, of which the weak syllables had a very comparable pitch contour, differing from the strong counterparts by about 35 Hz in average pitch height. Final full-word stimuli were cross-spliced in that the original strong syllable from the selected trochaic pseudoword was combined with the weak syllable from the selected iambic pseudoword, resulting in a trochaic cross-spliced test stimulus. Vice versa, the weak syllable from the trochaic pseudoword was combined with the strong syllable of the iambic pseudoword, resulting in an iambic cross-spliced test stimulus. All syllables were trimmed to 250 ms with the phonetic software PRAAT, using the overlap-add algorithm. This was done in order to avoid MMN asymmetries that arise solely by differences in stimulus or stimulus part durations. Longer stimuli in deviant compared to standard position elicit a smaller MMN than vice versa, i.e., shorter stimuli in deviant compared to standard position (Takegata et al., 2008; Colin et al., 2009). Furthermore, all syllables were set to an internal intensity of 70 dB, corresponding to a comfortable listening level at ∼70 dB SPL when played during the experiment. Wave forms and pitch tracks of the experimental full-word stimuli are displayed in Figure 1A. Due to identical syllable durations, each disyllabic word had a duration of 500 ms.

We also analyzed the phonetic timing in the trochaic and iambic words. Due to the cross-splicing, this timing was identical across conditions. First, closure durations as measure from technical stimulus beginning until onset of the consonantal burst were 50 ms. Second, the time from the onset of the consonantal burst until the beginning of the vowels was 40 ms. In acoustic terms, this means that syllables were separated by 50 ms-pauses (corresponding to the consonantal closure durations).

Omissions were realized as syllable omissions. All omissions were created by truncating the cross-spliced pseudowords at their respective mid-points. For instance, an omission in first-syllable position of a trochee resulted in a weak syllable that originally stemmed from an iamb. In total, due to two foot types and two positions, four truncated pseudowords realized the four types of omissions.

The stimulus material was arranged in a typical oddball paradigm, where stimuli could occur in standard or deviant position, distributed over several blocks. Blocks as displayed in Figure 1B. were further split in half in order to guarantee manageable experiment times. Thus, in 2 × 4 blocks, standards consisted of disyllabic (full) pseudowords (Figure 1B) and deviants of truncated pseudowords. Truncations resulted in either first-syllable omissions or second-syllable omissions. If syllables were omitted in first position, the deviant effectively started with silence. In each block, there were 350 standards and 50 deviants (translating into 87.5% standards and 12.5% deviants). The stimulus material was pseudo-randomized, with different randomization for each participant. Constraints on randomizations were as follows: (1) minimally four consecutive standards; (2) maximally 10 consecutive standards; (3) no immediate repetition of identical standard numbers, e.g., five standards and then five standards again. The first three standards per block and standards immediately following a deviant were discarded from further analyses. The stimulus material, arranged in 2 × 4 blocks, therefore constituted a 2 × 2 design, with two levels of foot type (trochee, iamb) and two levels of omission position (first syllable, second syllable). In order to match the number of occurring strong and weak syllables, we additionally included four blocks where standards were single syllables (strong and weak syllables from trochees and iambs) and deviants were full (i.e., disyllabic) pseudowords. These four blocks were not analyzed further.

All stimuli were presented with a constant inter-stimulus interval of 300 ms. This means that inter-stimulus differences were 300 ms (measured from the end of the second syllable of one disyllabic word to the beginning of the first syllable of the next disyllabic word). This translates into a Stimulus Onset Asynchrony (SOA) of 800 ms. Note that deviants with word-initial omissions effectively resulted in an SOA of 1,050 ms measured from the beginning of the standard immediately before the deviant and the beginning of the truncated deviant syllable.

Stimuli were presented over open-field loudspeakers placed symmetrically 1 m in front of the participants. Participants were seated in electrically and acoustically shielded EEG-cabins. Next to the loudspeakers, a flat-screen was placed 1.2 m in front of the participants. This screen was used to display a silent movie during the passive oddball paradigm.

After EEG-setup, participants passively listened to the 12 blocks of standard-deviant trains. There was no task except the request to ignore the sounds as best as possible, while watching a silent movie (without subtitles). After each block, the experimenter allowed for a short break. In the middle of the experiment, the break was longer and the air in the EEG cabin was refreshed. Each block lasted for about 5 1/2 min; the entire experiment in the cabin about 65 min.

Continuous EEG was recorded from 64 Ag/AgCl electrodes, arranged on a nylon cap following the extended 10–20 system (Oostenveld et al., 2011). EEG signals were amplified with a BIOSEMI ActiveTwo amplifier. Two electrodes placed left and right posterior to Cz were used as online-reference and as ground during the recording. EEG signals were recorded with a sampling rate of 500 Hz and filtered between DC and 250 Hz within the ActiveView BIOSEMI software.

Electroencephalogram raw data were analyzed within fieldtrip (Oostenveld et al., 2011), running on Matlab (Mathworks, 2016). Electrophysiological responses were analyzed in time windows from 200 pre-stimulus onset to 800 ms post-stimulus onset. These epochs were defined on the basis of the full disyllabic words and underwent automatic artifact detection implemented within fieldtrip (Oostenveld et al., 2011). This involved detecting muscle and eye-movement (electro-oculogram) artifacts as well as epochs with amplitudes exceeding 150 μV (peak-to-peak). Automatic artifact detection led to the exclusion of individual epochs, but in no participant or condition did the exclusion rate exceed 25% of the total number of epochs (mean exclusion rate: 9.27%). Subsequently, epochs were band-pass filtered between 0.3 and 30 Hz (Hamming-window digital Butterworth filter) and re-referenced to electrodes in close proximity to the mastoids (TP9, TP10) in order to approximate a linked-mastoid reference, as is common for MMN studies (Näätänen and Alho, 1997; Schröger, 2005; Winkler, 2007). For baseline correction, the mean amplitude of the pre-stimulus window (−200 to 0 ms) was subtracted from the epoch. Responses to standards and deviants in the first-syllable and second-syllable omission conditions were then averaged separately.

The mismatch negativity is defined as the difference between deviant and standard responses. In order to establish electrodes and time-points at which differences between standard and deviant responses are indeed significant, we used a multi-level, non-parametric cluster statistics approach (Henry and Obleser, 2012; Strauß et al., 2014), implemented in fieldtrip (Oostenveld et al., 2011). At the first level, we calculated independent-samples t-tests between single-trial amplitude values for standards and single-trial amplitude values for deviants, separately for the first-syllable and the second-syllable omission conditions. We thereby obtained uncorrected by-participant t-values for all time points and all electrodes. These t-values were subsequently tested against zero using dependent-sample t-tests at the second, i.e., group level, of our cluster-analysis. We estimated type I-error controlled cluster significance probabilities (at p < 0.05) by a Monte-Carlo non-parametric permutation method with 1,000 randomizations. The resulting matrix of t-values (electrodes × time points) was then analyzed between 100 and 200 ms post word onset for the first-syllable omission condition, and between 350 and 450 ms post word onset for the second-syllable omission condition. These time windows represent the expected temporal location of the omission MMN, measured from stimulus onset (Bendixen et al., 2014; Scharinger et al., 2017). Within these time windows, electrodes-time point clusters were determined by neighboring electrodes and neighboring time points for which t-values were above the significance threshold (p < 0.05). In the first-syllable omission condition, this led to a cluster of 20 electrodes, showing significant standard-deviant differences between 130 and 180 ms post-stimulus onset (Figure 1). In the second-syllable omission condition, we obtained a cluster of 28 electrodes, yielding significant standard-deviant differences between 400 and 450 ms post-stimulus onset (Figure 1). Note that the latter time window corresponds to time points between 150 and 200 ms post-deviance onset. The final electrode selection for further analyses was then based on the intersection of the two electrode clusters, yielding 18 fronto-central electrodes (AF3, AF4, AF7, F1, F2, F3, F4, F5, F6, F7, FC1, FC2, FC3, FC4, FC5, FC6, FCz, and Fz).

Next, we calculated the omission MMN as difference between deviant and standard responses for the aforementioned electrodes, and in the two temporal regions as determined from the cluster statistics, separately for each participant and meter type (trochee, iamb). This resulted in mean MMN values for each participant, electrode, omission position and meter. Additionally, within the two time windows of the omission MMN, we determined the peak amplitude and the time point (latency) of this peak amplitude. Peak amplitudes were selected automatically by determining the minimum value of the deviant-standard difference in the respective time windows, and by manually inspecting the plausibility of the peaks. The automatic approach performed well, and only in three cases manual adjustment was necessary.

Cortical sources of the omission MMN were estimated using Variable Resolution Electromagnetic Tomography (VARETA; Bosch-Bayard et al., 2001; Scharinger et al., 2017). The VARETA algorithm attempts a reconstruction of cortical sources by looking for a discrete spline-interpolated solution to the EEG inverse problem. This is achieved by obtaining estimates of the spatially smoothest intracranial primary current density (PCD) distribution that is compatible with the observed scalp voltage distribution. Possible solutions are restricted to gray matter on the basis of the probabilistic brain tissue maps available from the Montreal Neurological Institute (MNI, Evans et al., 1993). First, possible sources are modeled as a pre-defined grid of voxels with 7 mm spacing. The 64 electrodes were co-registered with the average probabilistic brain atlas developed at the MNI, assuming a head radius of 85 mm. The difference ERPs of standards and deviants in the MMN time window as established by the cluster statistics were transformed into source space. Statistical parametric maps (SPMs) of the PCD estimates were then constructed based on a voxel-by-voxel Hotelling T² test against zero (with df = 19).

Omission MMN mean amplitudes, peak amplitudes and latencies were then submitted to linear-effects mixed models (LMMs), calculated with the statistical software R (R Development Core Team, Vienna, Version 3.2.2). Results are reported as mixed-effects analysis of variance (ANOVAs) with F-values that were estimated by the lmerTest package (Kuznetsova et al., 2014), using the Satterthwaite’s method. These models used the fixed effects POSITION (omission of first syllable, omission of second syllable), FOOT TYPE (Trochee, Iamb), ELECTRODE (AF3, AF4, AF7, F1, F2, F3, F4, F5, F6, F7, FC1, FC2, FC3, FC4, FC5, FC6, FCz, and Fz) and the random effect SUBJECT in a full-factorial design (i.e., including all possible interactions).

Omission MMNs were reliably elicited in the typical time windows between 100 and 200 ms after deviance onset (between 100–200 ms and 350–450 ms post-stimulus onset, Figure 2). When looking at each expression of the factors POSITION (first vs. second syllable) and FOOT TYPE (trochee, iamb), topographies of omission MMNs showed typical fronto-central distributions, with sources in left and right temporal areas, including primary and secondary auditory cortex, planum temporale and parts of superior and middle temporal gyrus (Figure 3).

Statistical analyses on amplitudes are summarized in Table 1.

For both mean and peak amplitudes, omission MMNs were larger for omissions in the second syllable (mean amplitude: −2.28 μV, peak amplitude: −3.06 μV) than in the first syllable (mean amplitude: −1.89 μV, peak amplitude: −2.84 μV). The interaction of the effects POSITION and FOOT TYPE also showed similar patterns for mean and peak amplitudes (Figure 4). Notably, in first position, trochees elicited larger MMN responses (mean amplitude: −2.05 μV, peak amplitude: −2.95 μV) than iambs (mean amplitude: −1.73 μV, peak amplitude: −2.73 μV), while in second position, iambs elicited larger MMN responses (mean amplitude: −2.50 μV, peak amplitude: −3.29 μV) than trochees (mean amplitude: −1.93 μV, peak amplitude: −2.82 μV). When the interaction was decomposed according to FOOT TYPE, Iambs [mean amplitude: F_(1,665) = 38.97, p < 0.001; peak amplitude: F_(1,665) = 20.57, p < 0.001], but not trochees (all Fs < 1, n.s.), showed higher amplitudes for omissions in second-syllable, compared to omissions in first-syllable position. All effects and interactions were independent of the electrodes.

Naturally, MMN latencies differed between omissions of the first and omissions of the second syllable. Overall, iambs showed longer latencies than trochees; this, however, depended on the effect of POSITION, as seen from the decomposition of the interaction of the effects of POSITION and FOOT TYPE (Table 2).

The main effect of FOOT TYPE reflected on average an eight millisecond earlier omission MMN for trochees than for iambs. This effect was driven by foot type difference in the first-syllable omission condition, with significantly earlier latencies for trochees than for iambs. Here, omission MMNs occurred at 146 ms for trochees and at 161 ms for iambs. Latencies in the second-syllable omission condition did not differ between trochees and iambs (Figure 5).

The omission MMN has been identified as prediction error response to rare omissions in tone sequences whose inter-onset intervals would not exceed a specific temporal window of integration of about 125–150 ms (Yabe et al., 1997, 1998). Subsequent work has shown that the omission response can also be elicited by speech material (Bendixen et al., 2014) and in cases where the temporal window of integration is in fact exceeded (Scharinger et al., 2017). Here, we provide evidence that omissions of syllables whose duration by far exceeded the 125–150 ms integration window can also elicit a robust omission MMN. This is the basis for our following interpretations, since we take the omission response to reflect a violation of a syllable-based prediction.

The difference between omissions in first- and second-syllable position in our experiment may—at first sight—be based on differences in temporal predictions. Second-syllable omissions are characterized by a violation of the word-internal timing. In all disyllabic pseudowords, the onset-to-onset interval of the two syllables is 250 ms. In addition to the prediction that the syllable in second position is a repeated version of the syllable in first position, there is also a strong temporal prediction that the onset of the second syllable is 250 ms after the onset of the first syllable. Temporal predictions in audition are particularly fostered by regular acoustic contexts, such as provided by oddball paradigms (Tavano et al., 2014; Auksztulewicz et al., 2018; Lumaca et al., 2019; Pinto et al., 2019). Tavano et al. (2014) and Auksztulewicz et al. (2018) explicitly refer to the need of temporal regularity for higher-order predictions, possibly supported by the brain’s dynamic sensitivity to different processing frequencies (Arnal et al., 2014), related to motor-areas (Auksztulewicz et al., 2018) or subcortical, thalamo-cerebellar circuits (Schwartze et al., 2012). In our study, omissions in first and second position may differ on the basis of temporal predictability. While second-syllable omissions may rather be sensitive to word-internal temporal regularity, first-syllable omissions should be sensitive to between-word temporal regularity. However, since word-internal as well as word-external timing is constant throughout the experiment, the interpretation of the stronger effects in second position would be that a violation of within-word temporal regularity causes a stronger prediction error than a violation of between-word temporal regularity.

A more plausible interpretation of the differences between first- and second position is based on the interaction of the effects of position and foot type. This interaction indicates that the position effect crucially depends on foot type: Only for iambs, the omission of the second syllable resulted in a larger omission MMN. That is, there is not a position effect per se, but rather a strong prediction of when strong syllables occur in either trochaic or iambic words. In iambic words, the strong syllable appears in second position, thus, the omission of the second syllable should result in a stronger prediction error, if the omission MMN is sensitive to prosodic properties such as syllable weight. This is supported by the results of our experiment, where indeed second-syllable omissions in iambs resulted in stronger MMNs than first-syllable position omissions. The same, complementary pattern, held for trochees: Here, omissions in first-position resulted in stronger omission MMNs than omissions in second-position. Put differently, the omission MMN is not only sensitive to syllable omissions and their temporal position, but also to the prosodic properties of these syllables, following the different foot types. This partially replicates the findings of Brochard et al. (2003) who demonstrated that the MMN depends on the subjective accenting of sound sequencing, with stronger MMNs in strong positions compared to weak positions. In our study, strong and weak positions are encoded in the acoustics of the experimental material. To this end, trochees consisted of initial syllables with a higher pitch than their final syllables, while iambs consisted of final syllables with a higher pitch than their initial syllables. In both cases, the syllables with higher pitch are likely to be interpreted as strong syllables, and the respective omission of the strong syllables resulted in a larger MMN than the omission of the corresponding weak syllables. Future studies may take this as a starting point when examining to what extent these prediction violations co-vary with higher-order, aesthetic processing. Existing studies strongly suggest an interactive effect of prosodic expectations and aesthetic appreciation (Obermeier et al., 2013, 2015; Menninghaus et al., 2015). The omission paradigm can offer a new way to quantify this correlation.

Finally, when looking at the MMN latencies, our patterns of results suggest that trochees take a specific role in that the omission of their (strong) first syllable results in an earlier MMN than the omission of the (weak) first syllable of iambs. Hence, the omission of a strong syllable in first position results in a particularly salient prediction violation. Of course, it is impossible to base this effect on foot type because foot type and the position of strong syllable are confounded. To disentangle these effects, future work is necessary. However, in combination with the amplitude data, the conclusion seems warranted that trochees have a specific influence on the omission response in that this response is not only elicited at earlier latencies but also with a stronger amplitude when the strong syllable is omitted. Note that the omission of the strong syllable in iambs led to an even stronger MMN, indicating that at least the amplitude pattern does not depend on whether the syllable occurred word-initially or word-finally. Therefore, we conclude that the particular pattern elicited by trochees reflects their preferred status in German prosody (Wiese, 1996; Wagner, 2012; Wiese and Speyer, 2015). Furthermore, the latency effect in first-syllable position may also be driven by the Iambic-Trochaic Law (ITL) according to which foot beginnings are marked by higher pitch and/or higher syllable intensities, while foot endings are marked by longer syllable durations (Hay and Diehl, 2007; de la Mora et al., 2013; Crowhurst and Olivares, 2014; Crowhurst, 2020). Since we only modified pitch in our experiment, we cannot fully explore the ITL here, but suggest that earlier sensitivity to the omission of the higher-pitched syllable in trochees compared to the lower-pitched syllable in iambs is in accordance with this law. A likely articulatory explanation of this effect is that due to the respiratory cycle, word- and phrase initial syllables can be produced with higher intensities and higher pitch just because more air and more pressure is available after inhalation (see Tierney et al., 2011 for a similar explanation for song patterns in humans and non-humans).