**Attention**: For participants who plan to train a spatial sAAD decoder: Once you have downloaded the training datasets (which will be available once Development Phase begins), please keep in mind to convert the attended azimuth to a 0-based binary label system, with `0` representing attended to left, and `1` to the right. The attended azimuth is negative if attended to left. The baseline example code is not a spatial decoder, and therefore no conversion function is provided. ## Causality requirements While there are **no restrictions on the training procedure**, models must operate with a **maximal of 1 second -lookahead** . This means that, for each provided label for a sample at time t, only samples up to t+1000ms, from that specific test trial, can be used. This restriction applies in seconds, regardless of the sampling rate you use. Please remember to take **all** lags in **all** steps in your proposed method into account. Participants freely choose whether they work with windows or predict the label on a per-sample basis, as long as in the end the labels per individual time sample at **128 Hz** are provided, regardless of the sampling rate used for inference. When predicting labels per window, the participants therefore need to convert these per-window labels to per-sample labels, while obeying the causality requirements. **Example for window-based decisions with windows > 1 second** For example, when using 10s-windows, the 1s-lookahead can still be guaranteed: 1) for non-overlapping windows: by assigning the window's attention label only to the last second in the window, and to the first 9 seconds of the next window. 2) by using overlapping windows (for example, with shifts of 1s), allowing to update the per-sample decision more frequently without violating the causality requirement. **Disclaimer**: the provided pre-processed data were generated using procedures that might have slightly violated the 1s-lookahead rule. We do not take this into account. However, if you decide to use your own pre-processing method, we require you to only use pre-processing methods that satisfy the 1s-lookahead rule. The pre-processing code should also be provided with your submission. Only the following methods are exempted from the 1s-lookahead rule: - the `filtfilt` function, with the same filter order used in the provided preprocessing scripts. - the `removeBadChannel` function, carried within each dataset's own folder. - the `repairBadSegment` function, carried within each dataset's own preprocessing scripts. - z-score normalization, carried out across a complete trial (as this is also used in the baseline) - the above functions are exempted from the causality restriction only once per preprocessing pipeline. ## Datasets ### Training Datasets All training datasets share the structure of **two competing talkers with known spatial directions and speech envelopes**. To keep cross-dataset evaluation tractable, all datasets use a 64-channel layout (except the KUL-Ultra-High-Density dataset), recorded with either BioSemi or NeuroScan. There is a slight mismatch between channels, which the participants need to take into account. Channel labels are provided. | Dataset | EEG System | Ch. | Eye-Gaze | Language | Details | |---|---|---|---|---|---| | DTU | BioSemi | 64 | Not Controlled | Danish | 18 subjects, ±60°, 60 trials/subject, ~50 min EEG/subject | | KUL | BioSemi | 64 | Not Controlled | Dutch | 16 subjects, ±90°, 20 trials/subject, ~72 min EEG/subject | | KUL-AV-GC | BioSemi | 64 | Controlled | Dutch | 13 subjects, eye-gaze controlled, 6–8 trials/subject, 10 min/trial | | KUL-Ultra-High-Density| BioSemi | 255 | Not Controlled | Dutch | 30 subjects, 4 trials/subject, ~ 26 min/subject | | USTC | NeuroScan | 64 | Not Controlled | Chinese | 18 subjects, earphones, ~40 min EEG/subject | Special note of the KUL-Ultra-High-Density Dataset: only 64 channels are used to train the decoder. However, 255 channels are provided. Channel selection is applied during training and test. All provide scalp EEG, speech envelopes, and binary left/right attention labels in a uniform format. Meanwhile, a unique dataset/subject/trial ID is assigned to help participants identify these datasets, while we do not guarantee any consistency between these subject or trial IDs and those presented in the original dataset. Disclaimer: The envelope labels were algorithmically validated as part of the dataset preparation process. The left/right attention labels were retained from the original published datasets and were not independently verified. While their correctness has therefore not been formally established here, no issues have been identified that would call their validity into question. **Attention**: For participants who plan to train a spatial sAAD decoder: Once you have downloaded the training datasets (which will be available once Development Phase begins), please keep in mind to convert the attended azimuth to a 0-based binary label system, with `0` representing attended to left, and `1` to the right. The attended azimuth is negative if attended to left. The baseline example code is not a spatial decoder, and therefore no conversion function is provided. ### Blind Test Sets To rigorously evaluate generalization, we provide **two held-out blind test sets** (KUL-DLAAD and NJU-SUSTech), one of which (NJU-SUSTech) is newly recorded. These held-out test sets will contain switches in attention. Summary: | Dataset | Ch. | Recording System | Eye-Gaze | Language | Details | |---|---|---|---|---| | KUL-DLAAD | 64 | BioSemi | Controlled | Dutch | 19 subjects, 8 trials/subject, ~50 min EEG/subject | | NJU-SUSTech-sAAD-GC | 64 | NeuroScan | Controlled | Chinese | 12 subjects, ~60 min EEG/subject | The NJU-SUSTech test set recording has been approved by the SUSTech Ethics Committee (Approval No. 20260006). All participants provided written informed consent. You (the participants) are not allowed to try to use the dataset ID, subject ID, or trial ID to improve your decoder during the test phase. See `Terms` for detailed description. ### Preprocessing We provide two versions of the EEG datasets: the raw data and a minimally preprocessed one. The raw data is derived from the original datasets and transformed into a unified format, the extended CND format. Downsampling to 128 Hz is applied during this conversion if the original sampling rate is not equal to 128 Hz. The raw data is provided to allow more flexible preprocessing, as long as you wish. The preprocessed data is derived from the raw data. A band-pass filter between 1–9 Hz is applied. For some datasets, bad channel removal or bad segment repair are applied. The data is further downsampled to 32 Hz. The linear baseline and its HMM-postprocessed version uses the 32-Hz preprocessed data. The data is distributed in the extended CND format. See [this link](https://cnsp-resources.readthedocs.io/en/latest/citation.html) and [this manual](https://box.nju.edu.cn/f/ad95c4fc8e0a46d89649/) for detailed information. A conversion script from the CND format to the Superhuge format is provided, the later is used for the DNN baselines. ### Submission **IMPORTANT**: You will be asked to submit your results in Codabench. Therefore, it is MANDATORY to complete the application process in Codabench to grant you the eligibility to submit a result. If you have not complete the process yet, you may go to `My Submission` page, and follow the instructions there. Please also fill in the ![registration form](https://table.nju.edu.cn/dtable/forms/da668b8e-1bcc-4802-ac59-152b091b6d22/). When you submit your prediction, you are asked to answer these questions: - Is your code public, or will you make it public after the challenge? (submissions with public code will be highlighted in the leaderboard) - How long is the lookahead buffer of your algorithm, including all post-stimulus lags in the decoder? (make sure it is below 1 second) - Does your decoder predict the attended-speaker label (based on provided envelopes) or the attended-direction label (left=0, right=1)? We allow both spatial AAD and/or stimulus reconstruction algorithms. Therefore, participants can choose whether they provide spatial labels (`0` for left and `1` for right) OR envelope labels (envelope 1 vs. 2 attended, see the code of the linear baseline). The official submission format will be communicated before the `Development Phase` submission begins. We will provide a submission script in due time.