MAESTRO is a multimodal auditory-attention-decoding (AAD) dataset collected at The Ohio State University (OSU). In each trial a listener hears several talkers speaking at once from fixed locations in a room and attends to one of them; the goal is to decode which talker they attended from brain and behavioural signals. Every trial is recorded simultaneously across five modalities — scalp EEG, eye-gaze, head IMU, the spatialized multi-speaker audio stimulus, and an egocentric scene video — together with a post-trial comprehension label.

16 subjects, each: 5 familiarization + 100 evaluation trials (~35 s each)
~1,680 trials total · ~16 h of synchronized multimodal recordings · ~42 GB
Released CC-BY-NC-SA-4.0, with a companion loader (pip install maestro-loader) that returns perfectly cross-modal-aligned segments and trial-disjoint LOSO / intra-subject splits.

Why this dataset

Most AAD datasets are EEG-only. MAESTRO pairs EEG with every other signal that co-varies with auditory attention — where the eyes point, how the head moves, what the scene looks like, and the exact per-speaker audio — so that the "EEG ↔ each modality" relationship can be studied directly, and so overt behaviours (gaze, head orienting) can be honestly controlled for rather than silently inflating decoding scores. It targets two uses: (i) a benchmark for multimodal AAD, and (ii) supporting MAESTRO-Net, a stimulus-aware match–mismatch decoder.

Experiment paradigm

Each subject completes 5 training (familiarization) trials followed by 100 main (evaluation) trials. In every trial, ~35 s of simultaneous speech is played from fixed spatial loudspeakers; the subject attends one talker and answers a comprehension question afterward.

Custom python UI for recording the data

3 stereo audio devices → 6 fixed loudspeakers. Each device plays one stereo file whose two channels are two distinct talkers (L and R). Demuxed to a fixed 6-speaker numbering:

Speaker	Device	Channel	Azimuth	Role
1	1	L	−22.5°	attendable (Left, outer-near)
2	1	R	+22.5°	attendable (Left, inner)
3	2	L	−67.5°	attendable (Right, inner)
4	2	R	+67.5°	attendable (Right, outer)
5	3	L	−135°	distractor only
6	3	R	+135°	distractor only

Only speakers 1–4 are ever attended (attended_speaker ∈ {1,2,3,4}); speakers 5–6 are always distractors → free hard negatives for match–mismatch.

The attended speaker is balanced 25/25/25/25 across the 100 main trials. Per-speaker power and trial SNR are in metadata/trials.csv.

Target SNR distribution across trials of the attended loudspeaker with respect to the other loudspeakers

Modalities

Modality	Source	Rate	Shipped as
EEG	ANT Neuro 32-ch	500 Hz	parquet, raw ADC units (clipped ±0.0839) + 32 `ch_*` columns
Eye-gaze (per-eye) + pupil	Tobii Glasses 3	~50 Hz	parquet: `gaze2d/3d`, per-eye origin/direction, pupil diameter
Head IMU	Tobii Glasses 3	~100 Hz	parquet: accelerometer (`ax,ay,az`) + gyroscope (`gx,gy,gz`)
Scene video	Tobii Glasses 3	~25 fps	egocentric `.mp4` (room with loudspeakers; no faces)
Audio stimulus	3 stereo FLAC → 6 speakers	16 kHz	per-speaker mono `.flac` (no mixing)
Behaviour	post-trial question	per trial	comprehension correctness

EEG montage (channel order in metadata/eeg_channels.json): Fp1, Fpz, Fp2, F7, F3, Fz, F4, F8, FC5, FC1, FC2, FC6, M1, T7, C3, Cz, C4, T8, M2, CP5, CP1, CP2, CP6, P7, P3, Pz, P4, P8, POz, O1, Oz, O2. IMU and video share the gaze (Tobii) recording clock.

Sample EEG recording from Subject S08, Trial eval_017. (a) A 5-second excerpt of bandpass-filtered (1–45 Hz) and common-average-referenced EEG traces across all 30 scalp channels (mastoid electrodes M1 and M2 excluded), with a 50 μV spacing between channels for clarity. (b) Power spectral density per channel (gray) and their mean (blue), computed across the full trial. The characteristic 1/f spectral slope and alpha band prominence (8–13 Hz, shaded) confirm physiologically plausible signal quality.

Benchmark Task in the Paper

Target	Classes	Chance
Attended speaker (4-class)	1 / 2 / 3 / 4	0.25

Labels (attended_speaker, hemisphere, inout, 4-class) are attached to every segment by the loader.

Layout

data/eeg/subject=S01/trial=eval_001.parquet     # 500 Hz; t_sec(internal)+sample_idx+32 ch_*
data/gaze/subject=S01/trial=eval_001.parquet    # Tobii per-eye gaze, recording-relative t
data/imu/subject=S01/trial=eval_001.parquet     # Tobii accel+gyro
media/audio/eval_001/speaker{1..6}_dev{d}_{L|R}_spkid{ID}.flac   # per-trial, shared across subjects
media/video/subject=S01/eval_001.mp4            # egocentric scene video
media/timing/subject=S01/trial=eval_001.json    # unix anchors for cross-modal alignment
metadata/{trials,subjects,bad_channels,trials_per_subject}.csv
metadata/{eeg_channels,audio_layout,audio_manifest}.json
splits/loso/fold_00..15.json   splits/within/fold_0..4.json

Subjects are S01..S16; trials are training_001..005 and eval_001..100. The audio stimulus is identical across subjects, so it is stored once per trial.

How the modalities are aligned

Each stream runs on its own clock (EEG internal clock, audio playback clock, Tobii recording clock). Every trial ships media/timing/*.json with the wall-clock (unix) anchors — EEG first-sample time, per-device audio playback_start, and the Tobii recording start. From these the loader clips all streams to the single window where EEG, audio playback and the Tobii recording are simultaneously live, and (optionally) resamples them onto one shared time grid so EEG / gaze / IMU / audio come out the same length and sample-aligned. IMU and video share the gaze clock. You never have to do this by hand — but the anchors are public if you want to.

Quick start

pip install maestro-loader            # core
pip install maestro-loader[all]       # + torch, mne (EEG preproc), av (video)

from maestro_loader import load_aad, get_dataloaders

# Stream straight from the Hub (lazy download + cache); or pass local_path=...
ds = load_aad(subjects=[1, 2, 3], trials="main", modalities="all",
              segment_length=5.0, overlap=0.5, preprocess=False)
seg = ds[0]
seg["eeg"]    # (32, 2500) @ 500 Hz      seg["gaze"]  # (19, ~250) @ ~50 Hz
seg["imu"]    # (6, ~500)  @ ~100 Hz     seg["audio"] # (6, 80000) @ 16 kHz (6 speakers)
seg["attended_speaker"], seg["hemisphere"], seg["inout"]

More examples

# 1) Model-ready: decoder EEG pipeline + everything on one 64 Hz grid, 28-band mel audio
ds = load_aad(subjects="all", modalities=["eeg", "audio"],
              segment_length=5.0, preprocess=True)          # eeg (32,320), audio (6,28,320)

# 2) Aligned RAW signals at native rates (bring your own preprocessing)
ds = load_aad(subjects=[1], modalities=["eeg", "gaze", "imu"], preprocess=False)

# 3) Custom EEG filtering / reference (override the defaults)
ds = load_aad(subjects=[1], modalities=["eeg"],
              preprocess={"l_freq": 0.5, "h_freq": 30, "notch": 60, "reference": "average"})

# 4) Pick a common rate yourself, and choose the audio representation
ds = load_aad(subjects=[1], modalities=["eeg", "audio"], target_sfreq=128,
              audio_feature="waveform")                     # or "mel" / "envelope"

# 5) Window options: 1 s windows, 75% overlap; or one segment per whole trial
ds = load_aad(subjects=[1], modalities=["eeg"], segment_length=1.0, overlap=0.75)
ds = load_aad(subjects=[1], modalities=["eeg"], segment_length=None)   # whole trial

# 6) Per-segment z-scoring and torch tensors
ds = load_aad(subjects=[1], modalities=["eeg"], normalize="zscore", return_format="torch")

# 7) Selecting subjects / trials / training block
load_aad(subjects=[1, 5, 9], trials=[1, 2, 3])          # specific main trials
load_aad(subjects="all", trials="training")             # the 5 familiarization trials
load_aad(subjects=[2], trials="all")                    # training + main

# 8) Video frames decoded and aligned to the segment (needs maestro-loader[video])
ds = load_aad(subjects=[1], modalities=["video"], segment_length=2.0,
              video_frames=True, video_max_frames=8)
ds[0]["video"]            # (N, H, W, 3) uint8, frames within the window
# (default: items carry video_path + video_frame_range only — decode yourself)

Train / test splits — trial-disjoint, LOSO and intra-subject

The atomic unit of every split is a (subject, trial) — a trial is never divided across train and test, so windows cannot leak between them.

from maestro_loader import load_aad, get_dataloaders, assert_trial_disjoint

# Leave-one-subject-out (fold = held-out subject, 0..15)
train_dl, test_dl = get_dataloaders(setting="loso", fold=3,
                                    modalities=["eeg", "audio"],
                                    segment_length=5.0, preprocess=True, batch_size=64)

# Intra-subject k-fold (trial-disjoint within each subject)
train_dl, test_dl = get_dataloaders(setting="intra", fold=0, n_folds=5,
                                    scheme="chrono",       # or "random"
                                    subjects=[1], modalities=["eeg"], segment_length=5.0)

# Or get the datasets (numpy) instead of loaders, and verify the guarantee:
train, test = load_aad(subjects="all", modalities=["eeg"],
                       split={"setting": "loso", "fold": 0})
assert_trial_disjoint(train.units, test.units)

Ready-made split definitions are also shipped as JSON under splits/loso/ and splits/within/. Note: the 105 stimuli are shared across subjects by design, so LOSO holds out the subject (all of their recordings) — no (subject, trial) is ever in both splits; a stimulus-disjoint LOSO is not possible. For a held-out stimulus set, restrict trials= to disjoint id ranges.

Notes for users

Coverage. EEG and audio are complete for all 16 subjects on every trial. The Tobii streams (gaze, IMU, video) are present for 1,678 of the 1,680 trials; the two exceptions are S02/eval_077 and S02/eval_078, for which the glasses recording was not captured — the loader simply omits those streams for those two trials.
EEG units. EEG ships as raw ADC units (clipped ±0.0839); apply preprocessing (the loader's preprocess=True, or your own pipeline) to obtain microvolt-scale, referenced signals.
Referencing. preprocess=True applies a robust reference — linked mastoids when available, common-average otherwise — and interpolates any channels flagged in metadata/bad_channels.csv, all automatically.
Gaze. The eye-tracker is uncalibrated, so gaze is subject-relative.

Citation & license

Collected at The Ohio State University. Released under CC-BY-NC-SA-4.0 — please cite the accompanying paper when using the dataset -

@misc{hassan2026maestromultimodalauditoryattentionegocentric,
title={MAESTRO: a Multimodal Auditory-attention Egocentric Speech-TRacking Open corpus},
author={K M Naimul Hassan and Ali Alavi and Donald S. Williamson},
year={2026},
eprint={2609.31898},
archivePrefix={arXiv},
primaryClass={eess.AS},
url={https://arxiv.org/abs/2609.31898},
}

Downloads last month: 1,754

Total file size:

41.9 GB

Paper for aspire-osu/maestro-eeg-dataset

MAESTRO: a Multimodal Auditory-attention Egocentric Speech-TRacking Open corpus

Paper • 2609.31898 • Published 15 days ago