EEG-sAAD Data Story
ICASSP 2027 EEG-sAAD Grand Challenge · EDA and method roadmap

Integrate the whisper, starve the shortcuts

A signal-processing read of the challenge training data: what the raw trials reveal, why a 1-second decision sits near 55%, which shortcuts to starve, and which published methods are most likely to beat the linear-decoder + causal-HMM baseline.

Bottom line. The ICASSP 2027 EEG-sAAD training data carries a weak attention signal and several strong signals that have nothing to do with attention. The attention signal is small but consistent. In a within-subject test on one KUL-AV-GC trial, 1-s windows gave an evidence signal-to-noise ratio of SNR₁ ≈ 0.16, and accuracy grew with window length close to the Φ(SNR₁·√w) law. The official cross-dataset baseline (53.94% at 1 s) fits the same law with a lower SNR₁ of about 0.10. The other signals are large and specific to each lab:

Three changes are most likely to beat the linear-decoder-plus-causal-HMM baseline without breaking the ≤1 s lookahead rule, and each has published evidence behind it:

  1. Remove the lab "dialects" causally: align channels by name, then apply running normalisation and causal filtering.
  2. Raise SNR₁ with pooled-covariance or CCA decoders and within-trial, causal adaptation.
  3. Retune the temporal integrator. On synthetic switches scored with the official-style SDT rule, the best H came with SDTs of about 4–7 s (smoothing memory of roughly 5–10 s). The official baseline's SDT is 18.5 s.

Spatial (direction-only) decoders and DNNs that have not been checked for causality are the riskiest routes on gaze-controlled blind data.

Two surprises should change how teams validate. First, in the KUL-AV-GC example trial, the tracked talker does not change at the 5-min spatial switch. The switch is a change of direction, not of speaker. Second, the latency score L has a high chance floor: a decoder fed sign-randomised (pure-chance) evidence still scored L ≈ 0.49–0.57 in our re-implementation of the scorer. Every example-trial result below rests on one trial per dataset (n = 1 subject), and every switch-simulation result uses synthetic switches. Treat both as hypotheses to test at scale, not as estimates.

The story in five acts

The exploratory analysis ran at two levels. The corpus level used summary statistics for all 1,979 raw trials. The illustration level used one complete 128 Hz example trial per dataset (subject 1, trial 1). The five findings build on each other, and each one points to a design decision.

Act What the data shows Design consequence
1. Five labs, five dialects of "raw" 16× scale spread, DC offsets, different pre-filtering, only 57 shared channel names Causal, per-trial normalisation and name-based channel alignment come before any modelling
2. The signal is a whisper Channel-level tracking only 1–2× above a null; SNR₁ ≈ 0.16; accuracy follows √w ~55% at 1 s is a statistical limit; gains must come from higher SNR₁ or time integration
3. The 5-minute switch is not what it seems In the example KUL-AV-GC trial the attended talker is constant across the direction switch Speaker-label and direction-label targets need different training labels; genuine speaker switches may not exist in training
4. Shortcuts waiting to be learned Loud attended column, parity labels, covariance fingerprint, gaze/EOG Randomise stream order, z-score envelopes, avoid spatial-only decoders, validate leave-one-dataset-out
5. Time is the free lunch, up to a point Synthetic switches: best H at moderate smoothing; too sticky loses both terms Tune the integrator for ~5–10 s memory against the official scorer

Five labs speak five dialects of "raw"

The organisers describe the "raw" release as data derived from the original recordings and downsampled to 128 Hz. A second, "preprocessed" release adds a 1–9 Hz band-pass, 32 Hz resampling and bad-channel handling (Codabench). The corpus statistics show that "raw" is not one processing stage. KUL-AV-GC and KUL2016 trials are already zero-mean (RMS/SD = 1.00). DTU, KUL-UHD and USTC still carry large electrode DC offsets, with median RMS/SD of 210, 58 and 94. That pattern fits unreferenced, DC-coupled amplifiers. The median channel SD ranges from 8.3 (KUL-AV-GC) to 132.1 (KUL-UHD), a gap of about 16×. The DTU example trial also drifts by a median of ≈592 units/min, which is why 98% of its median-channel power sits below 4 Hz.

Dataset Median channel SD Median channel RMS RMS/SD Median max|x| Envelope inter-stream r
KUL-AV-GC 8.3 8.3 1.00 242 0.002
DTU 39.2 9,220 210 24,829 −0.008
KUL2016 9.6 9.6 1.00 282 0.010
KUL-UHD (253 ch stored) 132.1 8,330 58 44,929 0.032
USTC 61.5 5,315 94 23,096 0.012

Corpus-level, all 1,979 raw trials (this EDA).

Amplitude scale and DC offset by dataset
Figure 1. Left: the median channel SD differs about 16× across datasets. Right: DTU, KUL-UHD and USTC carry DC offsets far larger than their signal SD, while the two KUL-AV-GC/KUL2016 releases are already zero-mean. Any model that sees un-normalised data can identify the dataset from amplitude alone.

The DTU trials also vary a lot internally. Channel SD spans 17–808 across trials. Within the DTU example trial it runs from 34 (CP1) to 870 (FC5), so the "raw" data still contains bad trials and bad channels. The recording protocols explain part of this spread. KUL2016 is distributed high-passed at 0.5 Hz and MWF-cleaned (Zenodo 4004271). USTC is average-referenced, filtered at 0.1–45 Hz and ICA-cleaned (Zenodo 18366436). KUL-AV-GC is stored unreferenced (Zenodo 11058711).

The example-trial spectra (n = 1 per dataset) point the same way. KUL-AV-GC has a steep low-pass near 40 Hz, where its PSD falls about 10 orders of magnitude by 45 Hz, plus a sub-1 Hz high-pass. The 50 Hz mains peak-to-neighbour ratio is 10.8 (KUL-UHD), 9.3 (USTC), 1.6 (DTU), 1.0 (KUL2016) and 0.6 (KUL-AV-GC). The alpha peak above the 1/f fit is strong in KUL2016 (11.25 Hz, +0.77 log₁₀) and USTC (10 Hz, +0.75) but nearly absent in KUL-AV-GC (+0.07).

The edges of trials are hazardous. The USTC and KUL-UHD example trials jump by 9,249 and 13,011 units per sample at t = 0, and again at the end of the trial. Median channel kurtosis is 268 in USTC, against 3.4 in KUL-UHD and −0.2 in DTU. KUL-UHD also has an event at about 170 s and a bad P7 channel (SD 2,381 vs 71 at Cz). None of this matters for 1–9 Hz envelope tracking. All of it matters for a filter's warm-up, for robust scaling, and for any model that could learn "which lab is this?".

Example-trial spectra
Figure 2. Each example trial (n = 1 per dataset) shows a different pre-filtering history: a steep ~40 Hz low-pass in KUL-AV-GC, strong 50 Hz mains in KUL-UHD and USTC, and different alpha prominence. The shaded 1–9 Hz band is where the baseline works.

Montages add a third dialect. The three BioSemi-64 datasets (DTU, KUL2016, KUL-AV-GC) share channel names and order, although AFz is spelled "Afz". USTC (NeuroScan) uses upper-case names. It lacks AF7, AF8, AFZ, IZ, P9 and P10, and adds M1, M2, CB1, CB2, PO5 and PO6. KUL-UHD stores 253 channels, not the advertised 255. Its names follow the 5% system (POOz, FTT8h), and it lacks AFZ and FCZ. A case-insensitive intersection across all five datasets leaves 57 channels, with 58 for BioSemi∩USTC and 62 for BioSemi∩UHD. Pooling channels by index therefore scrambles montages without warning. Channels have to be matched by upper-cased name, with interpolation or masking where one is missing.

The envelopes differ too. DTU, KUL-UHD and USTC envelopes are standardised per trial; in DTU, both streams have SD 0.2368. KUL2016 and KUL-AV-GC envelopes are not standardised (SDs of 0.13–0.14 and 0.018). All envelopes are band-passed and zero-mean. The KUL-AV-GC envelope is high-passed below 1 Hz, which fits the 1–40 Hz band-pass described for its release (Zenodo 11058711). In the single example trial, the Mandarin USTC envelope has relatively more energy at 3–10 Hz than the Danish and Dutch sets. The challenge releases envelopes only, not audio, and the fetched rules do not say how the envelopes were computed (Codabench). This constrains the feature roadmap below, because richer speech features cannot be computed from test data that contains no audio.

The attention signal is a whisper (SNR₁ ≈ 0.16, the √w law)

Neural tracking is barely visible at the level of single channels. The example trials were preprocessed with the 57 common channels, a 1-s edge trim, median removal, a common-average reference, a 1–9 Hz Butterworth filter (zero-phase for this figure only), resampling to 32 Hz, and a robust z-score clipped at ±5. On that data, the RMS-over-channels cross-correlation between EEG and envelope peaks at only 0.02–0.09 at lags from −250 to +750 ms. That is only 1–2× above a circular-shift null of about 0.014–0.045. The attended stream beats the ignored stream at the peak in DTU (0.092 vs 0.060), KUL-UHD (0.026 vs 0.020) and the first half of KUL-AV-GC (0.032 vs 0.025). It does not in KUL2016 or USTC. A single channel is a whisper; a multichannel spatial filter is required.

Channel-level tracking
Figure 3. RMS-over-channels envelope–EEG correlation per example trial (n = 1 each). The grey band is the 95% circular-shift null. The attended–ignored gap is a few hundredths at best.

A regularised backward decoder concentrates this signal but does not escape the statistics. The decoder used lags of 0–500 ms, ridge regression and 32 Hz input, and was scored with within-trial blocked 5-fold cross-validation. Accuracy at 1 s was 50–58% across datasets and regularisation settings. For example, KUL2016 with causal filtering and α = 0.1 reached 58.2% (1 s), 64.0% (5 s) and 68.6% (10 s). Causal (sosfilt) and zero-phase filtering made no systematic difference at this sample size, which is good news for the lookahead budget. Each trial offered only about 2–6 minutes of training data.

Pooling across labs is harder. A leave-one-dataset-out test trained on the four other example trials (about 15 min from 4 people) and tested on the fifth. It was at chance: 44–54% at 1 s, with no consistent gain from longer windows. For comparison, the official baseline trained on about 60 h from four datasets reaches 53.94% at 1 s and 73.20% at 50 s on KUL-AV-GC (Codabench).

The cleanest way to see why accuracy at 1 s sits near 55% is to treat the decoder output as evidence. The cross-half KUL-AV-GC decoder was trained on one half of the trial and tested on the other, with column 0 as the attended stream. Its evidence is d = r(att) − r(ign), computed in 1-s windows with a 0.25-s hop. That evidence has mean 0.065 and SD 0.410, so the per-window SNR₁ is 0.159, and 1-s accuracy is 54.8%. If windows are independent, averaging w of them scales the SNR by √w, and accuracy should follow Φ(SNR₁·√w). The measured values track that prediction closely over nearly two decades of window length. The 30-s point, with only 19 windows, is the noisy exception.

Window Predicted Φ(0.159·√w) Measured (non-overlapping) n windows
1 s 56.3% 54.5% 598
2 s 58.9% 55.2% 299
5 s 63.9% 61.3% 119
10 s 69.2% 66.1% 59
20 s 76.1% 72.4% 29
30 s 80.8% 63.2% 19
60 s 89.1% 88.9% 9

KUL-AV-GC example trial, within-subject cross-half decoder (this EDA; n = 1 trial).

Accuracy versus decision window
Figure 4. Window accuracy follows Φ(SNR₁·√w) from a 1-s SNR of 0.16. The stars mark the official cross-dataset baseline at 1 s and 50 s.

The law can be inverted as a back-of-envelope check, with a caveat: it assumes independent windows and one extrapolation from a single trial. The official baseline's 53.94% at 1 s implies SNR₁ ≈ 0.10. That value predicts about 76% at 50 s, against 73.20% measured. So the cross-dataset baseline behaves like a weaker copy of the same evidence stream, not a broken one. The gap between 0.16 and 0.10 is worth about 2.6× in integration time. At SNR₁ ≈ 0.10, reaching 70% by integration alone takes roughly 28 s. At 0.16 it takes roughly 11 s. Reaching 60% within a single 1-s window would need SNR₁ ≈ 0.25.

With 32 samples per window and correlation gaps of a few hundredths, the sampling noise of Pearson's r swamps the effect. About 55% at 1 s is a statistical limit of the evidence, not an implementation bug. Every improvement has to either raise SNR₁ (a better decoder, better features, more or better-aligned training data) or integrate evidence over time (the HMM's job). The literature agrees on the scale. Figure-read backward-decoder accuracies on DTU are about 57–65% at 1 s in within-subject settings (Wong et al. 2018). Subject-independent linear decoders are much slower to switch than subject-specific ones: median MESD is 57.9 s vs 17.5 s on KUL (Geirnaert et al., JBHI 2021).

One leave-one-dataset-out result needs an audit. When USTC was held out, the reconstruction correlated more with column 1 than with column 0 (r = 0.083 vs 0.026 at α = 1; 0.060 vs 0.032 at α = 0.1). The evidence is weak: one trial, and the SE of r is about 0.03. Still, a stream-order check across all USTC trials is cheap and should be done before trusting any USTC-trained model.

The 5-minute switch changes direction, not talker

KUL-AV-GC is the only training set with gaze control and a within-trial switch: spatial attention switches at 5 min in each 10-min trial (Zenodo 11058711). That makes it the natural proxy for the blind sets, which contain "attention switches at random times within trials" and eye-gaze control (Codabench). The EDA tested what actually switches. It trained on one half of the example trial, tested on the other, and scored the predictions under two labelling hypotheses.

If the attended stream moves from column 0 to column 1 at 5 min, accuracy falls far below chance as the window grows: 41.6% (1 s), 30.3% (5 s), 25.4% (10 s), 5.3% (30 s), 0.0% (60 s). If column 0 is attended for the whole trial, accuracy rises normally: 54.5%, 61.3%, 66.1%, 63.2% and 88.9% (the 60-s point has n = 9). In this trial, the talker being tracked stays the same across the switch. The switch is in direction, which fits the two speakers swapping sides.

KUL-AV-GC switch test
Figure 5. Cross-half decoding on the KUL-AV-GC example trial. Labels that assume a speaker switch at 5 min fall toward 0% as windows lengthen. Labels that assume no speaker switch rise toward 89%. One trial; confirm against the CND attendedStream and azimuth vectors.

This is one trial, so it is a hypothesis. The consequences are large enough to check it at once. For the direction-label target (left = 0, right = 1), KUL-AV-GC's trial-level label, which gives the onset side, is wrong for the second half of every trial where the hypothesis holds. For the speaker-label target, which the rules also allow (Codabench), there is no switch in this trial at all. If the hypothesis generalises, the training corpus contains no genuine speaker switches. The speaker-switch side of the latency metric could then be validated only on synthetic switches.

The spatial side of the story is also flat. Parieto-occipital alpha lateralisation barely moves across the switch (mean index 0.040 → 0.045, Mann–Whitney p = 0.77), and neither does a horizontal-EOG proxy (AF7/F7/FT7 − AF8/F8/FT8 below 1 Hz: 0.00 → −0.00, p = 0.98). There is one caveat: the data are already high-passed, so any DC EOG shift would have been removed anyway. In this gaze-controlled trial, no spatial signature of the switch is visible. That fits published evidence that spatial AAD methods "do not achieve significant above-chance performances" on this dataset, while linear stimulus reconstruction works in every condition (Rotaru et al. 2024, arXiv 2412.01401).

Alpha lateralisation and EOG proxy across the switch
Figure 6. Alpha lateralisation (left) and the horizontal-EOG proxy (right) in 10-s bins across the 5-min spatial switch in the KUL-AV-GC example trial. Neither shifts detectably.

Four shortcuts are waiting to be learned

The organisers designed the blind sets to prevent "overfitting shortcuts". The rules forbid using subject or trial IDs, and they require a single forward pass over the test data (Codabench). The training data still offers four shortcuts that a flexible model will find on its own.

The loud column. In the training files, column 0 holds the attended envelope. In KUL2016, column 0 has the larger SD in 65% of 320 trials (Wilcoxon p = 2.3 × 10⁻⁹). In KUL-AV-GC the figure is 59% of 99 (p = 0.0095). DTU (50%), KUL-UHD (50%) and USTC (49%) are balanced (p > 0.5). The KUL2016 and KUL-UHD ratios fall on a few discrete levels. That means they encode which story pair was played, not attention. Pearson correlation is scale-invariant, so the linear baseline is immune. A DNN that sees raw envelopes can learn "larger-variance stream = attended", or learn story identity. The fixes are cheap. Z-score each stream per window, and swap the two stream columns at random during training (flipping the target to match), so that column position carries no information.

Envelope spectra and attended-column loudness
Figure 7. Left: envelope modulation spectra differ by dataset (n = 1 each). Right: across all trials, the attended column is slightly "louder" in KUL2016 and KUL-AV-GC, and its ratio falls on discrete, story-pair-specific levels.

Parity labels. In KUL-UHD and USTC, the trial-level label is exactly determined by trial-ID parity: odd trials are left, even trials are right (60/60 and 180/180). DTU, KUL2016 and KUL-AV-GC are balanced against parity. In an alternating-block design, anything that varies slowly over the session (drift, electrode impedance, fatigue) varies with the label. Published work shows DNNs learn exactly this temporal autocorrelation from block designs. Their accuracy drops significantly on non-block data, while linear stimulus reconstruction stays robust (Yan et al., Interspeech 2025).

The covariance fingerprint. On 5-s epochs, 57-channel 1–9 Hz covariances with log-Euclidean tangent features identify the dataset with 99.7% accuracy (chance 39%). In this analysis, dataset and subject are the same thing. After per-recording Euclidean alignment (R^−1/2 C R^−1/2), identification falls to 76.1%. Second-order spatial statistics are therefore a near-perfect domain fingerprint, and re-centering removes much of it but not all. The same effect appears in the literature. CSP-LDA trained to predict trial identity on KUL-AV-GC scores 99.6–100%. Under leave-one-trial-out it falls below chance (41–45%) for attended direction (Rotaru et al. 2024, J. Neural Eng.).

Covariance fingerprint before and after Euclidean alignment
Figure 8. Tangent-space covariance features form five tight dataset clusters before alignment (left). Per-recording Euclidean alignment merges them substantially but not fully (right). Example trials only, so dataset = subject here.

Gaze. KUL2016's own documentation warns of a "relatively strong eye gaze bias". It advises against evaluating direct-classification algorithms on that dataset, and against splitting a trial between training and testing (Zenodo 4004271). On KUL-AV-GC, subject-independent CSP reaches 69.5% at 5 s only in the static-video condition, where gaze and attention are congruent. It gets 50.0–56.9% in the incongruent conditions (Rotaru et al. 2024).

The blind sets are gaze-controlled. A decoder that learned gaze, trial order or lab identity will lose that signal at test time, or even have it reversed. The general pattern is well documented. The Vandecappelle CNN falls from 81.47% with a within-trial split to 58.60% on held-out data at 1 s (Puffay et al. 2023). Recent spatial DNNs report 94–97% at 1 s on KUL under within-trial splits but 61–78% under leave-one-subject-out. On DTU they are near chance (55.5–56.8%) (ListenNet, IJCAI 2025).

Time is the free lunch, but only up to about ten seconds

The metric is H = 2·Acc·L/(Acc + L), where L averages exp(−SDT/60) over switches (Codabench). L is lenient. An SDT of 5 s gives ℓ ≈ 0.92, 10 s gives ≈ 0.85, 18.5 s gives ≈ 0.73, and 60 s gives ≈ 0.37. Temporal smoothing therefore pays: the official HMM lifts 1-s accuracy from 53.94% to 68.4% at SDT 18.53 s (H = 0.7002).

The question is where the optimum sits. The EDA built synthetic switches from real evidence. It took the KUL-AV-GC evidence stream d(t) (the second half plus a time-reversed copy) and flipped its sign at random switch times drawn every U(30, 90) s, making 200 synthetic trials. It added +0.5 s of decision latency and fitted the HMM emission Gaussians on the first half only. Scoring re-implemented the official description (Codabench): within [t₀, t₀ + 100 s] (capped here at the next synthetic switch), τ is the split point that maximises agreement with an old→new step; the detection is valid only if the 0.5 s after τ all carry the new label, otherwise the switch scores 0; SDT = τ* − t₀; L is the median of exp(−SDT/60); H as above. This is a re-implementation, not the official scorer.

Rule (SYNTHETIC switches) Acc L ≈SDT H
plain 1-s argmax 59.7% 0.895 6.7 s 0.716
causal MA 1 s 61.1% 0.913 5.5 s 0.732
causal MA 5 s 70.8% 0.929 4.4 s 0.804
causal MA 10 s 73.5% 0.888 7.1 s 0.804
causal MA 20 s 75.6% 0.820 11.9 s 0.787
causal MA 40 s 68.6% 0.719 19.8 s 0.702
HMM p_stay = 0.9 (per 0.25 s) 64.6% 0.910 5.7 s 0.755
HMM p_stay = 0.98 73.7% 0.896 6.6 s 0.809
HMM p_stay = 0.995 75.2% 0.811 12.5 s 0.780
HMM p_stay = 0.999 68.0% 0.716 20.0 s 0.698
HMM p_stay = 0.9995 66.0% 0.689 22.3 s 0.674
Null: argmax on sign-randomised evidence 50.0% 0.486 — —
Null: HMM p_stay = 0.98 on sign-randomised evidence 49.2% 0.574 — —

≈SDT is back-computed from L as −60·ln L, so it summarises the median per-trial score, including missed detections (scored 0).

Within-subject decoder, one trial, synthetic switches, hop and lag approximations (this EDA). Not comparable with the official cross-dataset H.

Accuracy–latency trade-off on synthetic switches
Figure 9. SYNTHETIC switches on real KUL-AV-GC evidence, official-style SDT rule. Grey curves are iso-H contours. Moderate smoothing (effective memory ~5–10 s) wins. Very sticky integrators (≈SDT 20–22 s, the range of the official baseline's 18.5 s) lose both accuracy and latency. The grey cross is a chance-level decoder, which still earns L ≈ 0.49.

Two lessons follow. First, both the HMM and the moving average peak at moderate smoothing. Very sticky settings lose accuracy as well as latency, because every missed or delayed switch spends many seconds on the wrong label. The official baseline's SDT of 18.5 s sits in the regime that lost on synthetic data. That comparison has a limit. The baseline runs on weaker cross-dataset evidence (SNR₁ ≈ 0.10, against 0.16 here), and weaker evidence moves the optimum toward stickier settings. The right conclusion is "tune it on cross-dataset evidence", not "use p_stay = 0.98".

Second, the latency term alone does not prove switch tracking. Because τ* is chosen as the best-fitting split point, even a decoder with no information finds one somewhere in the 100-s window: on sign-randomised evidence, L was 0.49 (argmax) and 0.57 (HMM). An L of 0.5 therefore corresponds to chance, not to half-decent tracking, and the useful dynamic range of L is roughly 0.5–1. This is a validation pitfall for teams and organisers alike. Report L next to its null, and compare H using the official scorer on genuine switches.

Published real-time systems fall in the same moderate-memory range. A DTU/COCOHA closed-loop CCA system reported mean switch latencies of 4.7 s and 12.6 s in two experiments (bioRxiv 2024.03.01.582668). Geirnaert's MESD analysis finds the fastest stable gain control at short windows with modest accuracy (τ = 2.54 s, p = 0.62, N = 7 states). That beats long, accurate windows, and a decoder trained on averaged correlation matrices switches about 3× faster than one built by averaging per-trial decoders (Geirnaert et al., TNSRE 2020).

A prioritised roadmap from findings to fixes

The 1-second lookahead budget comes first

The rule covers "all lags in all steps" (Codabench), so lookahead should be budgeted like latency in a real-time system.

A simple unit test enforces the budget: perturb all inputs after t + 1 s and assert that every label up to t is unchanged.

Roadmap table

Priority EDA finding addressed Fix Key evidence Lookahead / rule compliance Expected benefit Main risk
P0 16× scale, DC offsets, drift, edge jumps, 57-name montage intersection Upper-cased name alignment (57 common, or interpolate/mask to a larger set); causal high-pass with warm-up; 1-s edge trim; running robust z-score and clipping; CAR This EDA; channel-adaptation benchmark (unverified, arXiv 2604.23091) Causal; within-trial statistics only Removes the amplitude fingerprint; prerequisite for everything else Warm-up transients at trial start; interpolation can hurt flexible models
P0 Loud column, envelope normalisation differences Per-window z-scored envelopes; random stream-column swap during training This EDA No test-time cost Removes a DNN shortcut; neutral for PCC None material
P0 Integrator memory; SDT 18.5 s baseline; √w law Retune HMM p_stay, window and emission temperature on leave-one-dataset-out folds against the official scorer; calibrate Δr emissions with running per-trial variance This EDA (synthetic); MESD; Miran 2018; Heintz HMM, arXiv 2506.24024 Causal forward pass; ≤ remaining budget for fixed-lag Largest cheap gain in H if the baseline is over-sticky Over-tuning to synthetic switches; weaker cross-dataset evidence shifts the optimum
P1 SNR₁ ≈ 0.10 cross-dataset Subject-independent backward decoder on pooled covariance ("avgcorr"); shrinkage/ridge (choice matters little) Geirnaert SPM 2021: avgcorr MESD 19.8 s vs avgdec 32.2 s (KUL); Wong 2018 0–250 ms EEG lags recommended Higher SNR₁ at zero test-time cost Dataset imbalance (DTU 1,080 short trials) dominates pooled covariance
P1 Same CCA with lagged/filter-bank transforms of EEG and stimulus Geirnaert SPM 2021: CCA MESD 15.1 vs 19.8 s (KUL), 16.3 vs 23.2 s (DTU); de Cheveigné 2018; code Stimulus-side lags use past context; keep EEG-side lags ≤ budget ~24–30% lower MESD within-subject No published leave-one-subject-out or cross-dataset CCA-vs-ridge test
P1 Covariance fingerprint 99.7% → 76.1% after EA Causal running Euclidean alignment (whitening by running mean covariance), applied identically in training and test He & Wu EA; Revisiting EA (MI/ERP gains; no AAD evidence) Allowed only within-trial and causally; per-subject pooling across test trials is not Reduces lab/subject shift in the spatial filter's input Unproven for stimulus reconstruction; warm-up period; negative transfer possible
P2 SI decoders far worse than subject-specific Within-trial time-adaptive unsupervised decoder (SI init, forgetting factor, soft labels from HMM posterior) Unsupervised AAD, JBHI 2021: MESD 19.6 s vs 57.9 s SI (KUL); time-adaptive; init-bias fixes, arXiv 2509.14764 Explicitly permitted "within-trial" if causal Potentially large for long trials Needs minutes of data (≈20 min to converge from random init); error reinforcement
P2 Switch timing vs stickiness Adaptive inertia: two-rate HMM / N-state MESD ladder / BOCPD run-length modulating p_stay MESD; BOCPD; SGAD (unverified, arXiv 2608.01618) Causal Better switch timing without losing stable-period accuracy No AAD-specific BOCPD evidence; adds tuning knobs
P3 Whisper-level SNR; unseen subjects Causal DNN envelope decoder (VLAAI-style) with future receptive field ≤ budget, contrastive loss, ensembling; output feeds the same PCC → HMM chain VLAAI: SI r 0.19 vs 0.125 linear; contrastive transformer Must be retrained with truncated lookahead Possibly the largest SNR gain Official VLAAI baseline at 30 s scored only 0.5136; shortcut learning; unseen-subject gains often vanish
Avoid / auxiliary only Gaze control, parity labels, covariance fingerprint, flat alpha/EOG at switch Spatial-only (CSP/RGC/DNN locus) decoders; EEG foundation models Rotaru 2024; arXiv 2412.01401; ICLR 2026 FM study (unverified, OpenReview) Compliant but fragile Low on gaze-controlled blind data Trial/gaze fingerprints collapse or invert at test time

P0: clean the dialects before touching the model

Preprocessing is the cheapest fix and it removes a real problem. Five datasets that differ 16× in scale, carry DC offsets up to 210× the signal SD, and share only 57 channel names will teach any learned model to identify the lab.

The pipeline is straightforward. Match channels by upper-cased name. Interpolate or mask missing channels rather than dropping temporal sites. Apply a causal high-pass, and discard or warm up its first second, because the USTC and KUL-UHD edge jumps would otherwise ring through an IIR filter. Re-reference to the common average. Scale with a running robust estimator and clip at a few SDs. Every statistic must be causal and within-trial. The rules forbid test-time updates "except within-trial and if the causality restrictions are satisfied", and they disqualify the use of subject or trial IDs (Codabench). Per-subject normalisation computed across a subject's test trials is therefore off the table.

A learned spatial filter can use a larger interpolated montage. Interpolation is not free, though. A 2026 benchmark (unverified) found that external channel adaptation degraded 24.8% of fine-tuning runs for flexible models (arXiv 2604.23091). For a linear decoder, the 57-channel intersection is the safe default.

On the stimulus side, only envelopes are released. Features that published match–mismatch work finds most useful, such as wav2vec 2.0 and mel spectrograms (Wang et al., arXiv 2401.04964), cannot be computed from the test data. The options left are transforms of the given envelope: per-window z-scoring, a half-wave-rectified derivative (an onset proxy; onsets are studied in J. Neural Eng. 2024), and a small causal filter bank on the envelope itself. None of these has a quantified 1-s AAD gain in the retrieved literature, so treat them as cheap experiments.

P0: retune the integrator against the real metric

The √w law says the HMM is doing the only thing that works at fixed SNR: integrating evidence. The synthetic simulation says the baseline probably integrates for too long. The recipe is concrete. Feed the HMM soft evidence rather than hard window decisions: a per-hop log-likelihood ratio from Δr, with class-conditional Gaussians fitted on training subjects, as Miran et al. did with log-normal markers (Miran 2018). Normalise Δr with a running per-trial variance, so that emissions stay calibrated when a new amplifier or subject changes the correlation scale. Add a single temperature parameter to absorb the dependence between overlapping windows. Then grid-search p_stay, window length and fixed-lag depth on leave-one-dataset-out folds, scoring with the official scorer.

Choose the operating point on the worst or median held-out dataset, not the mean. The MESD ladder, an N-state Markov chain over gain levels, is a principled alternative to the two-state HMM. It is designed precisely for short windows with modest accuracy (Geirnaert TNSRE 2020).

The organisers' own HMM line is the natural reference. It is a causal HMM over speakers with low switch probability and linear-time cost (Heintz et al., arXiv 2506.24024). A related Markov-switching-model preprint reports supervised subject-independent switch times of 33.7 s (MSM) vs 51.5 s (HMM) on segmented KUL2016. It uses a different switch-time definition from the challenge (unverified, arXiv 2602.13447). It shows that post-processing design alone can move latency by tens of seconds. It should not be read as a challenge-metric estimate.

P1: raise SNR₁ with better linear decoders and alignment

Among linear methods, CCA has the strongest published evidence. On KUL and DTU it significantly outperformed every backward stimulus-reconstruction variant (p < 0.001), with median MESD of 15.1 s vs 19.8 s (KUL) and 16.3 s vs 23.2 s (DTU) against the best backward decoder. The choice of regulariser (ridge vs lasso) had no significant effect (Geirnaert et al. SPM 2021). A stepwise match–mismatch study showed that each refinement mattered, from plain cross-correlation to full CCA with lags and filter banks on both sides. The 5-s error rate fell from about 30% to about 3% on single-talker data (de Cheveigné et al.).

The most transferable lesson from the same benchmark concerns how multi-subject training data is combined. Pooling correlation matrices ("avgcorr") beat averaging per-trial decoders significantly on both datasets. The MESD paper reports that this choice made switching about 3× faster (Geirnaert SPM 2021; MESD).

In multi-dataset training, pooling interacts with imbalance. DTU contributes 1,080 short trials against 99 for KUL-AV-GC, so the pooled covariances need per-dataset weighting. A 2026 preprint on unbalanced datasets is relevant but was not read (unverified, arXiv 2605.25605).

These results are within-subject. No retrieved study compares CCA and ridge under leave-one-subject-out or cross-dataset conditions, so the size of the gain in this challenge is an open empirical question.

Euclidean alignment targets the covariance fingerprint in Figure 8. It is label-free and cheap, and in motor-imagery and ERP BCIs it has produced gains of 1–26 and 36–40 percentage points (Revisiting EA; He & Wu). For a stimulus-reconstruction decoder there is no published AAD evidence, and the EDA shows alignment removes only part of the fingerprint (99.7% → 76.1%). Two design rules follow. The test-time version must be a causal running covariance within each trial. Training must use the same causal estimator, so that the decoder never sees a whitening quality it will not get at test time.

For direction decoding specifically, Riemannian parallel transport raised poor subjects' locus-of-attention accuracy from 45.84% to 67.92% (Wilroth et al. 2023). That supports alignment as a concept, but the evidence comes from a different task.

P2: adapt within the trial, carefully

The gap between subject-independent and subject-specific linear decoders is the largest in the literature. Median MESD on KUL is 57.9 s vs 17.5 s. Unsupervised self-adaptation initialised from the subject-independent decoder closes most of it (19.6 s, not significantly different from supervised). It works by iterating: predict, re-label, retrain (Geirnaert et al., JBHI 2021; code). A time-adaptive version updates its correlation matrices recursively with forgetting factors. It matches a fixed supervised decoder at 30-s windows (80.2–83.0% vs 80.2%) and tolerates the loss of 2–32 electrodes. It takes about 20–22 minutes to converge from random initialisation (time-adaptive AAD).

The rules explicitly allow within-trial, causal adaptation. The catch is data volume. Blind-set trial lengths are unknown, and KUL-AV-GC trials are 10 minutes, so adaptation starting from the SI decoder will help late in a trial more than early. Initialisation bias is a documented risk: iterations reinforce their own errors. Sum-of-speakers initialisation and soft labels mitigate it at 1.0–1.5× cost (arXiv 2509.14764). The HMM posterior is a natural source of soft labels.

Adaptive inertia is the other P2 lever: letting the integrator become less sticky when a switch looks likely. Bayesian online change-point detection keeps a run-length posterior that could modulate p_stay (Adams & MacKay 2007). No AAD application was found. One 2026 preprint reports that a learned switch-probability gate raised switch-F1 from 53.7% (EMA) to 67.3% at similar accuracy (unverified, arXiv 2608.01618). Treat it as a lead, not a basis for a decision.

P3: DNNs only with causal receptive fields and shortcut defences

DNNs can raise SNR₁ on large single-speaker corpora. VLAAI lifted subject-independent envelope correlation from 0.125 to 0.19 and held up on held-out subjects (0.18) and on DTU (0.17) (Accou et al. 2023). Contrastive transformer ensembles reached 86.40% match–mismatch accuracy on held-out subjects at 3 s (Sci. Rep. 2025).

Three facts temper this. First, VLAAI's published receptive field extends about 2.19 s into the future, so it must be retrained with a truncated one. Second, the official VLAAI baseline scored only 0.5136 at 30 s (Codabench). Third, unseen-subject gains often vanish. The 2023 regression winner scored 0.0976 against the linear baseline's 0.0960 on held-out subjects (HappyQuokka). In a two-speaker benchmark, "none of the tested nonlinear methods" was competitive on both KUL and DTU (Geirnaert SPM 2021).

A DNN worth submitting reconstructs the envelope (keeping the PCC → HMM chain, which is scale-invariant). It is trained with stream-swap augmentation on z-scored inputs, and it is validated leave-one-dataset-out. EEG foundation models are a low priority. A secondary summary of an ICLR 2026 study reports fine-tuned models reaching Pearson R near 0 on DTU envelope regression (unverified, OpenReview). The case against them rests mainly on the lookahead constraint and the general DNN evidence above.

Why spatial-only decoders are the riskiest bet

The rules let teams submit direction labels, and direction decoders post the most impressive published numbers. Every thread of this analysis argues against relying on them.

The canonical direction signal is parieto-occipital alpha lateralisation, with an individual peak of 9.32–11.19 Hz (Deng et al. 2020). It lies outside the 1–9 Hz tracking band. In the gaze-controlled example trial it showed no detectable shift across the switch. Combined with neural tracking it helped only at 25–50 s windows (JSLHR 2021).

Spatial decoders also exploit exactly the confounds found in the EDA. CSP identifies trials at ~100%, falls to 41–45% under leave-one-trial-out, and transfers across subjects only when gaze matches attention (Rotaru 2024). A predecessor challenge on cross-subject locus decoding reports winning scores of 56.50% and 56.99% (unverified, EEG-AAD 2026).

The KUL-AV-GC trial-level labels add another trap. If the direction-switch finding generalises, the labels are wrong for the second half of each trial, so a direction decoder trained on them learns noise. A spatial stream is defensible only as a small auxiliary evidence term fused into the HMM. It should be restricted to alpha-band parieto-occipital features, exclude frontal and EOG-proxy channels, and be kept only if it improves leave-one-dataset-out H on KUL-AV-GC.

Validate against the test conditions, not the training conditions

The validation protocol should mimic the blind sets' three shifts at once: new lab, new subjects, and within-trial switches under gaze control.

The outer loop is leave-one-dataset-out over the five training sets, with KUL-AV-GC as the primary proxy fold. It is the only gaze-controlled set with a within-trial switch, and it is the organisers' own hold-out, so its 53.94% (1 s) and H = 0.7002 are directly comparable reference points (Codabench). Hyperparameters, including HMM settings, go in an inner leave-one-subject-out loop. Nested schemes avoid the optimistic bias of a single loop (arXiv 2510.19174). Never score windows drawn from a trial that also contributed training data. Within-trial splits inflated a published CNN from 58.60% to 81.47% (Puffay et al. 2023). A causal time-series split is the realistic alternative (Ivucic et al., EMBC 2024).

Score with the official scorer from the organisers' Docker image, at 128 Hz, with subject-averaged accuracy and median-per-subject L. Report Acc, L, the full SDT distribution and the miss rate separately, not just H.

Keep genuine and synthetic switches apart. KUL-AV-GC offers genuine direction switches, but only once its labels are corrected for the second half, if the direction hypothesis holds. It may contain no genuine speaker switches. Speaker-switch latency therefore has to be estimated with synthetic switches. Build them at the evidence level, as in this EDA (sign-flipping d(t)), or by label-consistent stream swapping of the envelope columns. Avoid splicing raw EEG from different trials: splice discontinuities are detectable artefacts and would be exactly the kind of shortcut the blind sets remove.

Two synthetic-switch cautions apply. Evidence from a within-subject decoder is stronger than cross-dataset evidence, so synthetic results overstate performance and understate the optimal stickiness. Always pair a reported L with a null L, the same pipeline run on sign-randomised evidence, to expose latency credit that a chance-level decoder would also earn.

Finally, run the lookahead unit test (perturb inputs after t + 1 s and assert that labels up to t are unchanged) on the full submission container. Confirm that no step reads trial IDs, subject IDs or cross-trial statistics.

Items to audit before trusting any number

Several data and metric questions are unresolved. Each one could quietly bias training or validation.

Item Evidence so far Why it matters Cheap check
USTC stream order Held-out USTC reconstruction tracks column 1 (r 0.083 vs 0.026); n = 1, SE ≈ 0.03 A swapped attended column would poison 360 training trials Subject-independent decoder correlation with each column across all USTC trials
KUL-AV-GC direction vs speaker labels Example trial: talker constant across 5-min direction switch Trial-level label wrong for second half (direction target); no genuine speaker switches in training Compare with CND attendedStream/azimuth vectors; repeat cross-half test on all 99 trials
KUL-UHD amplifier and channel count Codabench lists BioSemi/255; Zenodo lists Compumedics SynAmps RT (Zenodo 4518754); 253 channels stored Affects reference, DC behaviour and montage mapping Inspect CND metadata; confirm which two channels are missing
Chance floor of L and search-window edge cases Official window is [t₀, t₀ + 100 s] with a best-split τ*; our re-implementation gives L ≈ 0.49–0.57 for a chance decoder; behaviour when two switches fall within 100 s is not documented L near 0.5 is chance, not partial success; closely spaced switches may be scored against the wrong label Run the official scorer on null (sign-randomised) evidence; ask organisers how overlapping windows are handled
Causality of the "preprocessed" 1–9 Hz release Filter type and phase not documented A zero-phase filter would silently exceed the lookahead budget Read the CND manual; re-filter 128 Hz raw data causally
Envelope computation and scaling Per-trial standardised in DTU/UHD/USTC, not in KUL sets; KUL-AV-GC high-passed Scale-sensitive models learn dataset or story identity Per-window z-score; check test-envelope statistics when released
Shared stimuli across KUL2016 and KUL-UHD Both Dutch stories, male talkers, ER3A at 60 dB (unverified overlap) Story-level leakage across folds Cross-correlate envelopes across datasets
Parity-locked labels KUL-UHD 60/60, USTC 180/180 Slow drifts co-vary with label; folds by trial order degenerate Fold by subject; test a drift-only classifier as a sanity check
Bad channels and segments in raw data DTU trial SD up to 808; KUL-UHD P7 SD 2,381 Outliers dominate pooled covariances Robust channel-quality screen before pooling
Blind-set protocol Public: KUL-DLAAD 19 subjects, 8 trials, ~50 min EEG each (BioSemi); NJU-SUSTech-sAAD-GC 12 subjects, ~60 min each (NeuroScan); switch rate and cueing unknown Switch rate, cueing and azimuth set the HMM prior Re-tune only on released test-set documentation, never on test labels

Conclusion

Seen as a signal-processing problem, this challenge is an evidence-accumulation task at very low SNR, run under a hard latency budget and heavy domain shift. The EDA suggests that almost everything distinctive about each dataset (scale, offsets, montage, envelope loudness, trial order, covariance) is a liability, and that the attention evidence itself is weak but consistent. That points to a less glamorous strategy than the recent direction-decoding literature: strip away the dialects causally, spend engineering effort on raising SNR₁ within the linear/CCA family, and treat the integrator's memory as the main hyperparameter of the metric.

The most consequential open question is not a modelling one. If the KUL-AV-GC switch is purely spatial, as the one example trial suggests, the training data holds no genuine speaker switch. Every team's speaker-switch latency then rests on synthetic validation, the latency term has a chance floor near 0.5, and the blind sets will be the first real test of switch tracking. Verifying the direction labels, the USTC stream order and the scorer's behaviour on null evidence should come before any model search.