How do stimulus priority and EEG evidence drive auditory attention?

A dataset built for instruction-free competition

Most open EEG benchmarks for auditory attention ask participants to attend to a designated stream, so they measure top-down selection. This paper targets the opposite situation: two sounds arrive at the same time, one in each ear, and nobody says which to follow. The authors call this stimulus-driven auditory attention and release the SAAD-EEG dataset to study it.

Twenty-three adults (14 male, 9 female; mean age 27.0 ± 2.7 years) listened to 480 dichotic pairs across three sessions of 160 trials. Each pair lasted 2 seconds and was drawn from four main sound categories — high-ecological environmental sounds, low-ecological environmental sounds, speech, and music — each split into two subcategories. To keep participants from deliberately selecting a stream, they watched a silent feature film and were told not to monitor the sounds. After a 1-second delay they reported which side was more attention-capturing.

That label definition matters for reading every result below. The recorded outcome is a self-report of orienting, not an externally verified attended ear, and the paper explicitly frames it as a behavioral measure of stimulus-driven orienting rather than an objective measure of attentional state.

Sound priority generalizes to unseen pairings

The behavioral core of the paper is a Bradley–Terry model, a pairwise-comparison method that assigns each sound a latent strength from a series of head-to-head contests. The question was whether a priority estimated from some pairings would carry over to pairings the model had never seen.

It did. Held-out prediction improved from an AUC of 0.577 to 0.718 when the estimated sound priority was used. In other words, the relative standing a sound earned in earlier competitions contained information about how it would fare against a new opponent, which is what makes a single priority value useful rather than a per-pair lookup table.

Category-level selection rates show the same ordering in a simpler form. Against an unbiased expectation of 0.5, speech was selected at 0.61 and high-ecological sounds at 0.58, while music sat at 0.48 and low-ecological sounds at 0.33; all four deviated significantly after Bonferroni correction, and all pairwise comparisons between categories were significant. The authors read this as consistent with the greater semantic salience and ecological relevance of speech and high-ecological sounds, and with music and low-ecological sounds being perceived as more background-like. That reading is the authors' interpretation, not a separately tested mechanism.

Acoustic differences explain only part of the choice

A natural alternative explanation is that participants simply chose the louder or more textured sound. Pair-level binomial GLMs tested this by relating left–right differences in nine perceptual features to the reported choice, with each feature z-scored across pairs.

Several features did predict choices. Loudness, spectral modulation, spectral centroid, and temporal modulation were positively associated with the selected side (all p<0.001), while spectral flatness and f0 were negatively associated (p<0.001 and p<0.01 respectively), and harmonicity showed a weak positive association (p<0.05). The remaining features were not significant.

So acoustics are real contributors, but the paper's framing is that they are not the whole account. The EEG result below is presented as information that remains predictive beyond acoustic asymmetry, which is why the proposed model keeps stimulus evidence and neural evidence as separate inputs rather than collapsing them.

Centro-temporal lateralization tracks the reported side

The EEG analysis found mid-to-late centro-temporal lateralization associated with the side participants reported as selected. Lateralization here means a difference in response between electrodes over the left and right hemispheres, and its timing in the mid-to-late portion of the response places it after early sensory encoding.

Two supporting checks make the dataset's neural content more credible. Comparing the 0–2 second stimulation window with a −3 to 0 second baseline, spectral power increased across 0–40 Hz, with theta (4–8 Hz) and alpha (8–13 Hz) showing the strongest increases and beta (13–30 Hz) a smaller but consistent one. Alpha and beta increases were most prominent over parieto-occipital regions, theta over anterior frontal regions.

Decoding gives a separate, trial-level measure. Using EEGNet under within-subject five-fold cross-validation, balanced accuracy averaged 68.38% (SD = 9.5% across participants), with per-participant means ranging from roughly 57% to 88%. Balanced accuracy is used because the self-reported labels can be imbalanced across individuals, so it weights classes equally rather than rewarding the majority class.

Decoding survives stricter train-test splits

A random trial split can leak information: the same stimulus pair may appear in training with one ear assignment and in testing with the other, letting a decoder exploit ear-of-presentation or stimulus-specific cues. The authors therefore added two grouping schemes.

In the mirror-constrained split, mirrored trials from the same pair were kept in the same fold, so the model never saw the opposite ear assignment of a pair during training. In the pairing-type held-out split, six-fold cross-validation trained on 20 of the 24 subcategory-level pairing types and tested on the 4 held out.

The three schemes produced comparable results: random split 68.4%, mirror-constrained 68.2%, pairing-type held-out 68.3%. The small gaps are the paper's evidence that decoding was not primarily driven by ear cues or memorized pairings. Note what the held-out scheme does and does not test: it measures generalization to unseen pairings, not to entirely novel sound categories.

Gating lets neural evidence adjust stimulus priority

The proposed framework follows the two findings. It first estimates a latent priority for each competing sound and forms relative stimulus evidence from the difference between the two priorities. A multi-scale EEG pathway then extracts trial-specific neural evidence through two complementary readouts: a signed readout, which can carry direction, and a power-based readout, which captures magnitude.

These streams are combined at the decision level through gated integration, meaning the neural evidence is allowed to modulate the decision rather than being averaged in with a fixed weight. The framework was evaluated under the mirror-constrained and pairing-held-out protocols alongside acoustic, EEG, and multimodal baselines, plus systematic ablations.

The resulting account is that spontaneous auditory selection reflects an interaction between generalizable stimulus priority and trial-specific neural variability. That is a computational account supported by the reported comparisons, not a claim that priority causes the neural lateralization or vice versa.

What the 68% figure does and does not mean

The headline decoding number is moderate, and the paper says so: reliable but moderate decodability, with substantial inter-individual spread from about 57% to 88% mean balanced accuracy. Within-participant variability across folds was relatively small for most individuals, so the spread is mainly between people rather than between test splits.

Several boundaries travel with these results. The sample is 23 university students who were native Chinese speakers fluent in English, with normal hearing and no reported neurological disorders, so the priority ordering and decoding levels are established for that group. The labels are self-reports collected after a 1-second delay, which the design used to keep motor responses out of stimulus-related EEG. And the pairing held-out test covers unseen combinations of the same eight subcategories, not new sound types.

Behavioral quality itself was high: only 15 trials were excluded under the 5,000 ms response criterion, 0.14% of all trials, and median reaction time fell from 554 ms in Session 1 to 469 ms in Session 2 and 440 ms in Session 3, which the authors attribute to growing task familiarity with minimal fatigue.

Priority is the reusable signal, EEG is the trial-level correction

The most decision-relevant split in this paper is between what transfers and what must be measured fresh. A sound's priority, estimated from prior pairwise competitions, transferred to unseen pairings and lifted held-out AUC from 0.577 to 0.718 — that is the reusable part, and it is what a system could carry into new pairings without new neural data. The EEG contribution is the opposite kind of quantity: trial-specific, varying with the participant and the moment, and worth about 68% balanced accuracy on its own under strict splits.

For anyone building on this dataset, the practical consequence is to treat the two evidence types as answering different questions. Stimulus priority tells you which sound is generally favored in a matchup; the lateralization and gated neural readout tell you what happened on this trial for this listener. The paper's contribution is showing that both are needed, and that the neural part still carries information after acoustic asymmetry and pairing memorization are controlled.

Sources used in this answer