Experimental stimuli

The present study employed naturalistic auditory tone sequences with pitch fluctuations exhibiting statistical regularities similar to those prevalent in natural stimuli^26,27. Specifically, each sequence consisted of 34 concatenated pure tones presented without temporal overlap or gap. Within the same sequence, tone pitches were temporally dependent upon each other (i.e., autocorrelated over time), allowing for the prediction of future tone pitches as a function of previously presented tone pitches. The degree of autocorrelation within each sequence was determined by β, which defines the relationship between the frequency of pitch fluctuations over time and the power of fluctuations at the respective frequency, such that P ≈1 / f ^β (i.e., the temporal power spectrum of pitch fluctuation). Consequently, a β of 0 means that pitch values between any two time points are uncorrelated, while a high β implies temporally adjacent tone pitches are positively dependent on one another. Further details regarding the tone sequence creation are described in detail in ref. ²⁵. The present auditory tone sequences were constructed with three levels of autocorrelation strength β: 0.5, 0.99, and 1.5.

In accordance with ref. ²⁵, each tone series was scaled such that its pitches ranged from log(220) to log(880). Tone series were discretized so that each tone was assigned to one of 25 values evenly spaced on the log scale with semitone distance. A circulant embedding algorithm⁵¹ was used to create nine unique 33-element long series, three for each β level:

where each element xj of xβ,i is taken to represent the pitch of the j^th tone in the sequence. Importantly, the choice of autocorrelation strength β lies within the range of natural acoustic signals, for which β commonly ranges between 0 and 2²⁶. The full set of tone sequences can be downloaded at:

https://med.nyu.edu/helab/sites/default/files/helab/Baumgarten_etal_stim_wav_files_and_figs.zip

All tone sequences converged on an identical penultimate (33rd) tone pitch (440 Hz), p33. This allowed us to disentangle sensory processing of p33 from predictive processing relying on p1−32. Specifically, since p33 was held constant across trials, it can be excluded from a regression which seeks to explain differences in neural activity during the presentation of the 33rd tone as a function of the previous tone sequence and the predicted upcoming tone pitch based on it.

For each tone sequence, a specific theoretically predicted final (34th) tone pitch (p34*; see refs. ^10,25 for further details) was computed, representing the optimally fitting final tone pitch given the pitch information provided by the first 33 tones. Nine unique sequences (Fig. 1b) were selected to represent all combinations of temporal autocorrelation level β (0.5, 0.99, 1.5) and three bins of theoretically predicted final tone pitch (p34*: low [370 Hz, 392 Hz], medium [440 Hz], high [494 Hz, 523 Hz]).

To probe subjects’ ability to predict the final tone pitch, the actually presented 34th tone of each sequence (p34) was independently drawn from one of six possible pitches located four [349 Hz/554 Hz], eight [277 Hz/699 Hz], or twelve [220 Hz/880 Hz] semitone steps below/above the mean pitch value of 440 Hz. Consequently, for a listener who can optimally extract the sequence information provided by the temporal autocorrelation within a given tone sequence, the tone pitch distance between p34 (i.e., the presented final tone) and p34* (i.e., the theoretically predicted final tone) should determine if p34 is considered likely or unlikely given the information provided by p1−33.

Identical tone sequences were presented in different tone duration conditions, comprising short (150 ms per tone/5.1 s per sequence), medium (300 ms/10.2 s), or long (600 ms/20.4 s) tone duration. The medium condition was used as the representative condition to determine sensor clusters of interest in later analyses.

In total, nine unique sequences (3 β levels × 3 p34*) × 3 tone durations constituted 27 distinct auditory sequences. Each distinct sequence was presented once within each of 12 blocks in random order, resulting in a total of 324 trials per subject.