Auditory EEG Challenge

Big challenge 04 · AI in Health

Two tracks: reconstruct speech from the brain, and match the brain to the speech.

Team registration openSubmissions open

EEG

Category: Artificial Intelligence in Healthcare
Format: Online, 9 November – 13 December 2026 · finale on 17 December at NIT and online
Tracks: 2 ranked · up to 2 per team
Teams: Solo or teams of up to 4
Submission deadline: 10 December 2026

Overview

Train on public EEG recorded while people listened to speech, then be tested on private, real recordings from people your model has never seen. The winning team is the one whose model still works on new brains.

The brain tracks the speech we hear, and EEG picks that up. Relating the two can help diagnose hearing problems in people who cannot be tested with ordinary methods, and it underpins smart hearing aids that follow which voice the wearer attends to.

The signal is very weak and differs a lot from person to person. In the 2023 ICASSP Auditory EEG Decoding Challenge, 14 of 85 subjects were held out. For envelope reconstruction, the top teams clearly beat the baseline on unseen stories but did not generalise to unseen subjects. That gap is the problem this challenge puts at the centre: your model is scored on recordings from people it has never seen, in a test set you cannot tune against.

Two practical pitfalls make results look better than they are, and the dataset authors warn about both: splitting segments of the same recording between training and testing, and relying on cues that do not depend on the speech. The rules and scoring are built to remove them.

The data

Accou et al., SparrKULee: A Speech-Evoked Auditory Response Repository from KU Leuven, Containing the EEG of 85 Participants — Data 9(8):94, 2024. The main training dataset.

Dataset source

Participants: 85
EEG channels: 64
Listening per participant: 90 to 150 minutes
Listeners: Young, normal-hearing
Stimuli: Natural Dutch speech
Licence: CC BY-NC 4.0 (non-commercial)

Also in the challenge

Second dataset
KUL
16 normal-hearing subjects, 64-channel EEG, four Dutch short stories narrated by different male speakers; 5.2 GB. Use version 2.1, the current release, dated 17 July 2026.
zenodo.org/records/21415059 Licence CC BY-NC-SA 4.0
Optional
DTU
18 subjects, 64-channel EEG at 512 Hz, Danish speech, trials of about 50 seconds with two competing talkers.
zenodo.org/records/1199011 Licence CC BY-NC 4.0

Test data · never released

The private real-data test

Private, real EEG recordings from subjects who appear in none of the public datasets. The test is split by subject into a validation set, used for feedback in the online phase, and a final test set, used only for the final ranking. Neither set is ever published, and you will not be given the recordings.

What you are told: The input format of the tracks and the pseudonymous subject_id of each window.
What you are not told: How many subjects there are, the recording equipment, the language or speakers, or how windows are grouped.
Privacy: All data is pseudonymised. Because you submit a model and not a data request, the recordings stay on the organisers' machines.

Download the datasets yourself from the official sources: the organisers do not redistribute them. Read each dataset's terms before use.
Getting SparrKULee: follow the README of its code repository. Clone it, install the requirements, then mount the dataset with rclone using the instructions linked there; the older direct-download route is marked deprecated. Copy the preprocessed data to local disk before training rather than training from the mount.
A few SparrKULee files from 11 participants (107 recordings) are restricted for privacy. You do not need them, but you can request them from sparrkulee@kuleuven.be, stating your use.
Trial fingerprints, a warning from the dataset authors: models can recognise which recording a segment came from. If segments of one recording appear in both training and validation, scores are inflated and can collapse under a proper split. Split by whole recording, story or, best, subject.
Gaze bias, a warning from the dataset authors: classifiers that predict attended direction can exploit eye movements rather than the brain's response to speech. Tie your decoder to the speech signal itself, as both tracks here do.
A good habit: train on SparrKULee and test on KUL before trusting any in-dataset score.

Tracks

A team may enter up to 2 of these. Each ranked track has its own leaderboard.

01Track 1

Leaderboard · per-subject correlation

Reconstruction

Predict the speech envelope the person was hearing, from their EEG.

Input: An EEG window of 64 channels × 640 samples (10 s at 64 Hz), float32, and metadata.csv with a pseudonymous subject_id.
Output: 640 values per window at 64 Hz: the predicted envelope, as defined by the reference code.
Chance level: Correlation near 0.
The question: How much of the heard speech can be recovered from a new person's EEG?

Evaluation

Metric	Purpose
Track scorePrimary	0.6 × average term + 0.4 × robustness term: the score that ranks the track.
Per window	Pearson correlation between predicted and true envelope; a constant prediction scores 0.
Subject score	Mean correlation over that subject's windows.
Average term	Mean of subject scores over all test subjects.
Robustness term	Mean of the subject scores of the lowest 25% of subjects.

Every score is first computed per subject, so a few easy subjects cannot carry a model. Two teams are tied when the 95% interval of the difference between their scores, from a bootstrap that resamples subjects 1,000 times, includes zero. The tie-break is the higher median subject score. Reported, not ranked: the distribution of your subject scores.

Scoring

Average term — mean subject score over all test subjects60%
Robustness term — mean subject score of the weakest 25% of subjects40%

Expected output

/output/envelopes/<window_id>.npy for every window: 640 values, the predicted envelope at 64 Hz.

You submit

Frozen model package
A Docker image (preferred) or a zip with an environment file, weights and a run.sh. It is run as ./run.sh /input /output and writes /output/envelopes/<window_id>.npy. Upload a zip of up to 10 MB; share the image or a larger zip as a link.
ZIP · up to 10 MB, or a link

02Track 2

Leaderboard · per-subject accuracy

Match-mismatch

Given the EEG and 5 candidate speech segments, pick the one that was actually heard.

Input: The same EEG window, plus 5 candidate speech envelopes, 5 × 640. One candidate is time-aligned with the EEG; the other 4 come from other parts of the speech.
Output: 5 probabilities per window that sum to 1.
Chance level: 20% accuracy.
The question: Can the model tell whose brain activity belongs to which speech, on a new person?

Evaluation

Metric	Purpose
Track scorePrimary	0.6 × average term + 0.4 × robustness term: the score that ranks the track.
Per window	Correct or not.
Subject score	Accuracy over that subject's windows.
Average term	Mean of subject scores over all test subjects.
Robustness term	Mean of the subject scores of the lowest 25% of subjects.

Every score is first computed per subject, so a few easy subjects cannot carry a model. Two teams are tied when the 95% interval of the difference between their scores, from a bootstrap that resamples subjects 1,000 times, includes zero. The tie-break is the lower log-loss on the 5 probabilities. Reported, not ranked: the distribution of your subject scores and your accuracy per subject.

Scoring

Average term — mean subject score over all test subjects60%
Robustness term — mean subject score of the weakest 25% of subjects40%

Expected output

/output/match.csv with the columns window_id,p1,p2,p3,p4,p5: five probabilities per window that sum to 1.

You submit

Frozen model package
A Docker image (preferred) or a zip with an environment file, weights and a run.sh. It is run as ./run.sh /input /output and writes /output/match.csv. Upload a zip of up to 10 MB; share the image or a larger zip as a link.
ZIP · up to 10 MB, or a link

Grand Prize

How the tracks add up.

In each track you enter, your track score gives a rank r among N ranked teams, and a rank score of 100 × (1 − (r − 1) / N). The automatic score A is the mean of your rank scores, so a team that enters one track competes on equal terms.

The jury gives the top 6 teams by A a jury score J from 0 to 100.

Final = 0.6 · A + 0.4 · J

Jury score J · 100 points

Method — soundness and justification of choices30pts
Generalisation — how you avoided leakage, your train-on-SparrKULee, test-on-KUL result, your subject-score distribution, honest limits30pts
Reproducibility — clean package, clear method note, declared data and models20pts
Presentation and answers to questions20pts

The top 6 teams are told privately on 16 December at 18:00 and present to the jury at the finale on 17 December.

How to submit

Register your team, then upload each deliverable from the participant portal. Your latest upload before the deadline is the one evaluated.

Track 1 · Reconstruction

Frozen model package
A Docker image (preferred) or a zip with an environment file, weights and a run.sh. It is run as ./run.sh /input /output and writes /output/envelopes/<window_id>.npy. Upload a zip of up to 10 MB; share the image or a larger zip as a link.
ZIP · up to 10 MB, or a link

Track 2 · Match-mismatch

Frozen model package
A Docker image (preferred) or a zip with an environment file, weights and a run.sh. It is run as ./run.sh /input /output and writes /output/match.csv. Upload a zip of up to 10 MB; share the image or a larger zip as a link.
ZIP · up to 10 MB, or a link

Every team

Method note (at most 2 pages)
One note for the team, covering every track you enter: data used, split strategy, results of training on SparrKULee and testing on KUL, and known limits. Describe any per-subject adaptation, list every dataset and pretrained model, disclose any use of generative AI and cite the datasets.
PDF · up to 10 MB
Training codeoptional
Repository link. Optional at the freeze: the top 3 teams per track share their training code within 3 days of a request.
Link

Technical requirements

Both tracks use the same input: a 10-second window from a 64-channel cap, band-limited and downsampled to 64 Hz by the reference preprocessing published on 9 November. The test is delivered already in this form, so every team sees identical input.
The exact envelope definition and the preprocessing code are published on 9 November with the baseline models and the scoring code, so that predictions and targets are computed the same way for everyone.
Submit a frozen package: a Docker image (preferred), or a zip with an environment file, weights and a run.sh. It is run as ./run.sh /input /output.
/input/eeg/<window_id>.npy holds a float32 array, 64 × 640. For Track 2, /input/candidates/<window_id>.npy holds the 5 candidate envelopes, 5 × 640. /input/metadata.csv has window_id,subject_id.
Track 1 writes /output/envelopes/<window_id>.npy with 640 values. Track 2 writes /output/match.csv with columns window_id,p1,p2,p3,p4,p5.
The package runs with no network and one GPU (specification announced on 9 November), with a total time limit for the whole set. A package that crashes, times out or writes malformed files scores zero for that track.
Per-subject adaptation is allowed if it uses only the unlabelled EEG in /input and is described in your method note.
On 9 November the organisers publish format_check/: ten windows built from public KUL data in the exact layout above, so you can test your pipeline end to end.
Real-data validation opens on 16 November: submit a package up to once every 3 days and receive the overall validation score within 24 hours. The final test is run once, on the package you froze on 13 December.

Competition rules

Who can enter

Students and researchers from Algerian universities and research centres, in teams of 2 to 4. Each person joins one team. If you have no team, tell us and we will match you.

Registration

Entry is free. At least one member must be registered for ARC'26 under the fees in the call for abstracts. Register by Friday 27 November: team name, members, affiliations, one contact email, tracks entered and the signed honour code.

Online teams welcome

Your package is run for you on the organisers' machine, and the finale on 17 December takes place at NIT, Rahmania, Algiers, and online.

Non-commercial use only

The public datasets are under non-commercial terms, so everything built from them, including trained models, may be used only for non-commercial purposes. Company teams are not eligible.

Not eligible for prizes

Organisers, jury members and their direct supervisees or lab collaborators. They may take part unranked.

Honour code

Signed at registration. Sharing models between teams or breaking the rules leads to disqualification.

1Allowed data: public auditory EEG datasets and public pretrained models only, so everyone has the same access. Download them yourself from the official sources; do not share copies. List everything you used in your method note.
2Splits: never place windows from the same recording, story or subject on both sides of a train and validation split. Your method note must state how you split.
3Forbidden: trying to obtain, recover or infer the private test recordings; code that detects the test set, hard-codes outputs or opens network connections.
4Frozen package: one package per track, uploaded by 13 December, 23:59, with a method note of at most 2 pages: data used, split strategy, results of training on SparrKULee and testing on KUL, known limits.
5Generative AI: disclose any use in code, writing or data work. The team stays responsible for all content.
6Reproducibility: the top 3 teams per track share training code on request within 3 days. Prizes stay provisional until that check passes.

Cite Accou et al. (SparrKULee, 2024), Biesmans et al. (the KUL data, 2016) and Monesi et al. (Auditory EEG Decoding Challenge for ICASSP 2023, 2024) in your method note and in any publication. The full references are at the bottom of this page.

You may publish your method after the results are announced, citing the datasets and the challenge. The private test data cannot be published.

Timeline

Key dates.

Call opens: registration, scoring code, preprocessing code, format_check/, baseline notebooks
Mon 9 Nov
Real-data validation opens
Mon 16 Nov
Registration closes (same day as ARC'26 registration)
Fri 27 Nov
Final team list published
Mon 30 Nov
Model freeze: packages and method notes uploaded; validation closes
Sun 13 Dec, 23:59
Organisers run every frozen package on the final test set; results stay sealed
Mon 14 – Tue 15 Dec
Day 1: challenge desk open for check-in. At 18:00 the top 6 teams are told privately
Wed 16 Dec
Finale at NIT, Rahmania, Algiers, and online
Thu 17 Dec
Reproducibility check done; prizes confirmed
Mon 21 Dec

Recognition

What’s at stake.

Grand Prize
Track 1 podium
Track 2 podium
Certificates for all participants

Prizes stay provisional until the reproducibility check passes; they are confirmed on 21 December.

FAQ

Something else? Ask the organisers — answers of general interest are posted here.

Important

Research use only: nothing built in this challenge is a medical device or may inform clinical decisions.

Glossary

Speech envelope: The slow rise and fall of the loudness of speech over time. Track 1 reconstructs it; its exact definition is in the reference code.
Match-mismatch: Deciding which of several speech segments was heard during an EEG window. Here, one of 5 candidates is time-aligned with the EEG.
Unseen subject: A person whose EEG was never used for training. The test set contains only such people.
Pearson correlation: How closely two signals rise and fall together, from −1 to 1; 0 means no linear relation.
Trial fingerprint: Recording-specific cues that let a model recognise where a segment came from, which inflate scores when one recording is split across training and validation.
Per-subject adaptation: Adjusting the model to each test subject using only that subject's unlabelled EEG: allowed here, if described in the method note.
Bootstrap: Resampling the test subjects many times (here 1,000) to tell whether the difference between two teams is real.
Log-loss: A penalty on predicted probabilities in which confident wrong answers cost the most; Track 2's tie-break.

References

[1]Accou et al., SparrKULee: A Speech-Evoked Auditory Response Repository from KU Leuven, Containing the EEG of 85 Participants. Data 9(8):94, 2024.doi.org/10.3390/data9080094
[2]SparrKULee, KU Leuven Research Data Repository (RDR).doi.org/10.48804/K3VSND
[3]SparrKULee code and download tools.github.com/exporl/auditory-eeg-dataset
[4]Biesmans et al., Auditory-inspired speech envelope extraction methods for improved EEG-based auditory attention detection in a cocktail party scenario. IEEE Transactions on Neural Systems and Rehabilitation Engineering 25(5):402–412, 2016 (the KUL data).doi.org/10.1109/TNSRE.2016.2571900
[5]KUL auditory attention dataset (Auditory Attention Detection Dataset KULeuven), version 2.1, Zenodo.zenodo.org/records/21415059
[6]DTU auditory attention dataset (EEG and audio dataset for auditory attention decoding), Zenodo.zenodo.org/records/1199011
[7]Monesi et al., Auditory EEG Decoding Challenge for ICASSP 2023. IEEE Open Journal of Signal Processing, 2024.doi.org/10.1109/OJSP.2024.3376296
[8]Auditory EEG Decoding Challenge, ICASSP 2023.exporl.github.io/auditory-eeg-challenge-2023/

Auditory EEG Challenge · ARC'26

Bring your best model.

Metric

Purpose

Track scorePrimary

0.6 × average term + 0.4 × robustness term: the score that ranks the track.

Per window

Pearson correlation between predicted and true envelope; a constant prediction scores 0.

Subject score

Mean correlation over that subject's windows.

Average term

Mean of subject scores over all test subjects.

Robustness term

Mean of the subject scores of the lowest 25% of subjects.

Metric

Purpose

Track scorePrimary

0.6 × average term + 0.4 × robustness term: the score that ranks the track.

Per window

Correct or not.

Subject score

Accuracy over that subject's windows.

Average term

Mean of subject scores over all test subjects.

Robustness term

Mean of the subject scores of the lowest 25% of subjects.

Who can enter

Students and researchers from Algerian universities and research centres, in teams of 2 to 4. Each person joins one team. If you have no team, tell us and we will match you.

Registration

Online teams welcome

Your package is run for you on the organisers' machine, and the finale on 17 December takes place at NIT, Rahmania, Algiers, and online.

Non-commercial use only

The public datasets are under non-commercial terms, so everything built from them, including trained models, may be used only for non-commercial purposes. Company teams are not eligible.

Not eligible for prizes

Organisers, jury members and their direct supervisees or lab collaborators. They may take part unranked.

Honour code

Signed at registration. Sharing models between teams or breaking the rules leads to disqualification.

1Allowed data: public auditory EEG datasets and public pretrained models only, so everyone has the same access. Download them yourself from the official sources; do not share copies. List everything you used in your method note.
2Splits: never place windows from the same recording, story or subject on both sides of a train and validation split. Your method note must state how you split.
3Forbidden: trying to obtain, recover or infer the private test recordings; code that detects the test set, hard-codes outputs or opens network connections.
4Frozen package: one package per track, uploaded by 13 December, 23:59, with a method note of at most 2 pages: data used, split strategy, results of training on SparrKULee and testing on KUL, known limits.
5Generative AI: disclose any use in code, writing or data work. The team stays responsible for all content.
6Reproducibility: the top 3 teams per track share training code on request within 3 days. Prizes stay provisional until that check passes.

You may publish your method after the results are announced, citing the datasets and the challenge. The private test data cannot be published.

Key dates.

Call opens: registration, scoring code, preprocessing code, format_check/, baseline notebooks
Mon 9 Nov
Real-data validation opens
Mon 16 Nov
Registration closes (same day as ARC'26 registration)
Fri 27 Nov
Final team list published
Mon 30 Nov
Model freeze: packages and method notes uploaded; validation closes
Sun 13 Dec, 23:59
Organisers run every frozen package on the final test set; results stay sealed
Mon 14 – Tue 15 Dec
Day 1: challenge desk open for check-in. At 18:00 the top 6 teams are told privately
Wed 16 Dec
Finale at NIT, Rahmania, Algiers, and online
Thu 17 Dec
Reproducibility check done; prizes confirmed
Mon 21 Dec

64 channels. 10 seconds. Unseen brains.

SparrKULee

Choose your track.

How the tracks add up.

Ship it from your team workspace.

Play fair, show your work.

Who can enter

Registration

Online teams welcome

Non-commercial use only

Not eligible for prizes

Honour code

Key dates.

What’s at stake.

Questions, answered.

Do I need a GPU?

Can I use KUL for validation?

Can I use a pretrained model?

Will I see the test recordings?

Can I enter only one track?

Can I publish my method?

I have no team. Can I still take part?

Where do I ask a question?

Bring your best model.

Auditory EEG Challenge

64 channels. 10 seconds. Unseen brains.

SparrKULee

Choose your track.

How the tracks add up.

Ship it from your team workspace.

Play fair, show your work.

Who can enter

Registration

Online teams welcome

Non-commercial use only

Not eligible for prizes

Honour code

Key dates.

What’s at stake.

Questions, answered.

Do I need a GPU?

Can I use KUL for validation?

Can I use a pretrained model?

Will I see the test recordings?

Can I enter only one track?

Can I publish my method?

I have no team. Can I still take part?

Where do I ask a question?

Bring your best model.