A dataset built for moving conversations
Most auditory attention decoding (AAD) research asks a listener's brain response to track one of several competing voices. The decoding target is usually the speech envelope, the slow amplitude contour of a voice, and the standard test is whether the reconstructed envelope correlates more strongly with the attended stream than with the unattended one. The paper argues that this test has usually been run in static, simplified scenes that poorly match everyday listening, where talkers move and conversations turn-take.
MOV-AAD addresses that gap by recording 64-channel EEG together with synchronized eye tracking, respiration airflow and effort, galvanic skin response, heart rate, peripheral oxygen saturation, skin temperature, body motion via accelerometry, and photoplethysmography. Fifty participants with normal hearing (age 24±4.5 years) completed the protocol at a 1200 Hz sampling rate. Speech sources moved across the frontal horizontal plane from -90° to +90°, rendered through head-related transfer functions, so the spatial position of a talker changed over the course of a trial rather than staying fixed.
Tasks that keep listeners engaged
The dataset separates several listening situations rather than pooling them. A repeated sentence task (120 sentences) served as an initial EEG reliability check. A localization task asked participants to report where a source was, using nine discrete azimuth locations spanning -90° to +90°. The conversation tasks comprised 40 single-conversation trials and 56 multi-conversation trials, each about 40±2.0 s long, with diotic background noise (pedestrian noise or speech babble) set 9 or 12 dB below each conversation stream and streams matched in root-mean-square power.
Attention was measured behaviorally through repeated-word detection: a space-bar press within 2 s of a repeated word in the attended stream counted as a hit, and a press with no attended repeated word in the preceding 2 s counted as a false alarm. This matters for interpreting the decoding results, because it shows whether participants were actually following the instructed stream rather than merely hearing it.
Localization holds at the group level
Localization accuracy was defined as the proportion of correctly identified source locations, with chance at 1/9, or about 11.1%. Mean absolute error compared reported and true azimuth, with random guessing expected to produce roughly 66.7° of error. Most participants stayed above chance accuracy and well below chance error, and perceived angles remained largely aligned with the correct hemifield even in the lower-accuracy third of participants.
The paper reports reduced discriminability at extreme lateral positions, which it attributes to cue saturation of interaural time and level differences and to the use of non-individualized head-related transfer functions. The practical reading is that spatial cues were preserved well enough to support moving-speaker tracking at the group level, while individual spatial precision varied substantially.
Attended streams are behaviorally separated
Repeated-word detection confirmed that the attention manipulation worked. F1 scores were significantly higher for the attended stream in the multi-conversation condition than for the unattended stream (Wilcoxon signed-rank test, p<10⁻⁴, r=0.87). F1 was also significantly higher in the single-conversation condition than in the attended multi-conversation stream (p<10⁻⁴, r=0.52), a smaller but reliable drop that the authors read as slightly reduced detection when interfering conversations are present.
This comparison is the behavioral anchor for everything that follows: the attended and unattended streams differ in what listeners did, not only in how the stimuli were labeled.
Envelope decoding carries the signal
For decoding, the speech envelope was extracted and EEG was filtered to 1-8 Hz, the low-frequency band in which cortical tracking of the envelope is typically observed. A backward model trained with time lags from 0 to 400 ms reconstructed the envelope, with ridge parameters selected by leave-one-trial-out cross-validation, and performance was scored as the Pearson correlation between reconstructed and true envelopes on held-out trials.
Across sliding windows, the attended-stream correlation stayed relatively stable as window size grew while the unattended correlation remained near chance, producing a clear separation at every window size tested. AAD accuracy, defined per trial as the proportion of windows with attended correlation exceeding unattended correlation, rose with window length and substantially exceeded the 0.5 chance line even at short windows. At the subject level, envelope features gave consistently positive attended-minus-unattended correlations and higher accuracy across participants.
Trajectory features add limited support
The paper also tested whether the moving position of a talker could itself be decoded. Trajectories were generated with a first-order Markov chain at 1° resolution, smoothed, upsampled to 10 Hz for rendering, and interpolated to 100 Hz to match the neural sampling rate. Two bands were examined: delta (0.05-2 Hz) and alpha (8-12 Hz).
Trajectory decoding was weaker than envelope decoding. The delta trajectory stayed near chance across window sizes, while the alpha trajectory performed consistently above delta and showed modest, systematic gains as windows lengthened. In a full-trial LDA classifier over reconstruction correlations, envelope features gave the highest accuracy across subjects, alpha trajectory features reached significant above-chance performance, and delta trajectory remained at chance. Combining envelope with trajectory features did not beat envelope-only decoding, consistent with the authors' reading that envelope features dominate AAD while alpha-band trajectory information is complementary but limited.
Where the decoding claim stops
The decoding evidence supports a specific priority: for moving naturalistic conversations, the attended speech envelope remains the dominant and most reliable AAD feature, and longer analysis windows improve accuracy. Spatial trajectory information, especially in the alpha band, is real but secondary, and the delta-band trajectory did not decode above chance. Anyone comparing feature sets or setting a window length should treat envelope and trajectory results as answering different questions rather than interchangeable measures.
The physiological sanity check reported feature distributions consistent with expected ranges from prior literature and device specifications, with a small number of outliers attributed to natural variability or minor recording artifacts. That check establishes signal quality, not attentional decoding. The dataset's contribution is therefore best read as a benchmark resource: it makes moving, multimodal listening testable, while the reported decoding results show which signals already carry attention and which remain candidates for future multimodal modeling.
Sources used in this answer
MOV-AAD: A Large-Scale Multimodal Dataset for Auditory Attention Decoding During Moving Conversations
Envelope-based EEG decoding separated attended from unattended speech across window sizes and rose with window length, while alpha-band trajectory features gave limited above-chance support and delta-band trajectory stayed near chance.