BCIwiki (bciwiki.com) — A single EEG network called NEUROTOKEN identified which speaker a listener was attending to 9% to 16% more accurately than the strongest non-generative baseline across 3 public datasets at 5-second windows, Ali Alavi and Donald S. Williamson reported in a preprint posted to arXiv on September 30, 2026. The work has not been peer reviewed.
Auditory attention decoding (AAD) targets the cocktail-party problem: working out from EEG whose voice a listener is following, so a hearing aid or brain-computer interface (BCI) knows what to amplify. The authors argue the field is fragmented. Directional AAD classifies which side attention points to but does not map that to a stream; regression-based source AAD ranks candidates by a single Pearson correlation that is noisy at the 1- to 5-second windows real devices need; and envelope reconstruction has no native decision rule.
In NEUROTOKEN, 3 heads share one EEG front-end. A conditional flow-matching head, ATTUNEFLOW, scores candidate streams by an integrated velocity-residual likelihood ratio. Two inference-time ensembles, QUADTRACK, which combines 4 complementary statistics, and ENV-FLOW, which z-normalizes and merges QUADTRACK with ATTUNEFLOW, are designed to absorb the failure modes of individual statistics.
On the KU Leuven, DTU and NJU datasets, ATTUNEFLOW also cut across-subject variance by about 3x, and trial-level fusion exceeded 93% on 2 of the 3 datasets. Reproducing widely cited direction-AAD results, the authors found that accuracies of 95% to 97% fell by 17% to 45% under a strict trial-disjoint protocol, which they say shows why a likelihood-based formulation is needed.