Source-linked AI summary
A High-Density EEG Dataset for Stimulus-Driven Auditory Attention
Ruofan Yan, Na Lu, Shu Peng, Wenlong You, Zhige Chen, Yuxuan Yan, Yan Liu, Kay Chen Tan, Jibin Wu
TL;DR
Most open-access EEG attention datasets study explicitly instructed, goal-directed listening, although communication often involves complex environments and salient interruptions. This paper introduces an instruction-free dichotic-listening EEG dataset; validation analyses report behavioral consistency, high data quality, and robust decodability.
Problem
Most open-access EEG datasets focus on explicitly instructed, goal-directed selection, despite communication often occurring amid complex environments and salient interruptions.
Source: Background & Summary · p. 1Method
The SAAD-EEG dataset records 128-channel EEG from 23 participants reporting which side captured attention during instruction-free dichotic sound competition while they watched a silent movie.
Source: Background & Summary · p. 1, Experimental Procedure · p. 3Results
Validation analyses report high dataset quality, behavioral consistency, and robust decodability across multiple cross-validation schemes.
Source: Background & Summary · p. 1Takeaways & Limitations
Trial-level orienting labels, reaction times, subjective sound ratings, and standardized data and scripts support fine-grained analyses and reproducibility.
Source: Background & Summary · pp. 1–2The labels are self-reported behavioral measures of stimulus-driven orienting, not objective measures of attentional state.
Source: Usage Notes · p. 10
Abstract
from arXiv · showhide
Stimulus-driven auditory attention determines which sound gains priority when multiple sources compete without an explicit listening goal, yet most computational studies focus either on acoustic salience or on decoding predefined attended targets. This study investigates instruction-free auditory competition using the Stimulus-driven Auditory Attention (SAAD) paradigm and develops a neurophysiologically informed framework that integrates stimulus-derived sound priority with trial-specific EEG evidence. Behavioral analysis using a Bradley--Terry model showed that sound priority estimated from previous competitions generalized to unseen sound pairings, improving held-out prediction from an AUC of 0.577 to 0.718. EEG analysis further revealed mid-to-late centro-temporal lateralization associated with the reported selection side, with neural information remaining predictive beyond acoustic asymmetry. Guided by these findings, the proposed model first estimates a latent priority for each competing sound and forms relative stimulus evidence from their difference. A multi-scale EEG pathway with complementary signed and power-based readouts then extracts trial-specific neural evidence, which is incorporated through gated decision-level integration. The framework is evaluated using mirror-constrained and pairing-held-out protocols, together with representative acoustic, EEG, multimodal baselines, and systematic ablations. The results support a computational account in which spontaneous auditory selection reflects the interaction between generalizable stimulus priority and trial-specific neural variability.
Background & Summary
Most open-access EEG datasets address instructed, goal-directed listening, while resources for instruction-free competition between multiple sound streams remain scarce. SAAD-EEG provides a naturalistic dichotic-listening dataset with behavioral annotations and validated EEG for studying stimulus-driven auditory orienting.
Background & Summary
- Existing open-access EEG datasets largely target instructed listening, and resources for resolving instruction-free competition between multiple streams remain scarce.Available stimulus-driven resources have largely focused on temporal salience within a single stream.Source: Background & Summary · p. 1
- SAAD-EEG records 128-channel EEG from 23 participants who chose which of two simultaneously presented sounds captured attention, without explicit sound-attention instructions.Participants watched a silent movie during sound presentation to minimize goal-directed selection.Source: Background & Summary · p. 1
- Validation analyses confirm the dataset’s high quality, behavioral consistency, and robust decodability across multiple cross-validation schemes.The release also includes trial-level orienting labels, precise reaction times, and participant ratings of familiarity, urgency, and likability.Source: Background & Summary · p. 1
- The release provides raw and preprocessed EEG, stimulus manifests, and processing scripts in standardized formats to support accessibility and reproducibility.The annotations enable analyses of acoustic, semantic, and ecological properties in relation to neural processing and attentional capture.Source: Background & Summary · p. 2
- The dataset supports investigation of bottom-up attention and comparisons of how ecological validity and semantic urgency relate to orienting in naturalistic sound categories.The paper presents it as a foundation for developing attention-aware hearing devices for real-world acoustic environments.Source: Background & Summary · p. 2
Methods
The methods use naturalistic dichotic sound pairs in an instruction-free task, with EEG and behavioral responses collected and processed alongside participant ratings. The stimulus design balances left–right presentation and samples cross-category pairings.
Methods
- The study recorded 23 adult participants, all native Chinese speakers fluent in English, with normal hearing and no reported neurological disorders.The group included 14 males and 9 females and had a mean age of 27.0 ± 2.7 years.Source: Participants · p. 2
- Stimuli were organized into high- and low-ecological environmental sounds, speech, and music, each divided into two subcategories.Ecological validity was defined by everyday occurrence and attentional relevance.Source: Stimuli · p. 3
- The researchers selected 2-second clips and constructed 24 cross-category subcategory pairing types, each repeated 10 times with mirrored left–right versions, yielding 480 dichotic pairs.Clip amplitude was normalized using RMS amplitude from the loudest 1% of samples; within-category and self-pairs were excluded.Source: Stimuli · p. 3
- During each trial, participants heard sounds simultaneously for 2 seconds without ear-specific instructions, then reported the more attention-capturing side after a 1-second delay.A 3-second silent period preceded the sounds; participants watched a silent film and were asked not to deliberately monitor the dichotic sounds.Source: Experimental Procedure · pp. 3–4
- EEG was recorded with a 128-channel HydroCel system at 250 Hz, while responses and reaction times were logged and aligned with trial events.Participants also rated representative clips for familiarity, liking, and urgency on a 0–4 scale.Source: Data Acquisition · p. 4
- Preprocessing included noisy-channel interpolation, common-average re-referencing, filtering, and ICA-based artifact removal; 6-second epochs span baseline, stimulation, and post-stimulus intervals.Primary decoding and spectral summaries used the 0–2-second stimulation interval.Source: Data preprocessing · pp. 4, 6
Data Records
The publicly released SAAD-EEG dataset is organized into raw EEG, processed data, stimuli, and ratings, with metadata describing the dataset and its layout. Participant folders contain trial-level EEG and behavioral records, while stimulus and rating files support reuse of the experimental materials and annotations.
Data Records
- The Zenodo release is organized into raw_eeg, processed, stimuli, and ratings directories, with a consistent participant and session structure.The directory overview shows sub-01 in full; other participants and sessions follow the same naming and folder structure.Source: Data Records · p. 6
- Root-level files document the folder layout, technical dataset parameters, participant demographics, and changes between releases.The release states that participant data contain no personally identifiable information beyond demographics.Source: Root-level metadata · p. 6
- Each participant’s processed folder contains 480 trial-epoched EEG records and trial-wise orienting labels and reaction times; raw recordings are stored by session in EGI .mff format.The processed EEG array has shape (480, 128, 1500) at 250 Hz.Source: Processed folder · p. 6, Raw EEG folder · p. 6
- The stimuli folder contains reference clips and dichotic pair waveforms with a trial index, while the ratings folder holds scale descriptions and participant ratings across eight sound subcategories.The ratings cover familiarity, urgency, and liking on 0–4 scales.Source: Stimuli folder · p. 6, Ratings folder · p. 6
Technical Validation
Validation found robust above-chance EEG decoding across trial-random, mirror-constrained, and pairing-held-out splits, alongside behavioral and neurophysiological patterns consistent with stimulus-driven orienting. The analyses also characterize data quality and define the scope of the released behavioral labels.
Robustness across partitioning strategies
- 68.4%, 68.2%, and 68.3% BACC were obtained for random, mirror-constrained, and pairing-type held-out splits, respectively, with comparable decoding across schemes.The results indicate that decoding was not primarily driven by ear-of-presentation cues or memorized pairings.Source: Figure 4 · p. 7, Robustness across partitioning… · pp. 8–9
Signal Quality Assessment
- Across participants, 8–15 channels were interpolated (median = 11) and 3–9 ICA components removed (median = 5), while artifact removal reduced blink-related signals and preserved overall temporal characteristics.One participant required 30 interpolated channels.Source: Signal Quality Assessment · p. 7, Figure 4 · p. 7
Neurophysiological Plausibility Check
- Stimulation increased spectral power relative to baseline, strongest in theta and alpha; theta increases were most prominent frontally, while alpha and beta increases were most prominent parieto-occipitally.The topographies showed distinct spatial distributions across frequency bands.Source: Neurophysiological Plausibility Check · pp. 7–8, Figure 4 · p. 7
Baseline Decoding Results
- 68.38% (SD = 9.5%) group-level BACC shows EEGNet predicted self-reported response labels above chance under random-split cross-validation.Mean participant BACC ranged from approximately 57% to 88%, with relatively small fold-to-fold variability for most participants.Source: Table 3 · p. 8, Baseline Decoding Results · p. 8
Behavioral Data Validation
- Only 15 trials (0.14%) were excluded; median response time decreased from 554 ms in Session 1 to 440 ms in Session 3, while music had the highest liking and lowest urgency ratings and HE the opposite profile.Subjective ratings covered familiarity, perceived urgency, and liking across eight stimulus subcategories.Source: Behavioral Data Validation · pp. 9–10, Figure 5 · p. 9
- Speech (SR = 0.61) and HE (SR = 0.58) exceeded chance selection, while music (SR = 0.48) and LE (SR = 0.33) fell below it; acoustic-feature associations varied in direction.Loudness, spectral modulation, spectral centroid, and temporal modulation were positively associated with choices; spectral flatness and f0 were negatively associated.Source: Behavioral Data Validation · pp. 9–10, Figure 5 · p. 9
Interpretation and reuse
- The released labels measure self-reported stimulus-driven orienting outcomes, not objective attentional state; the dataset also supports analyses of acoustic, semantic, and ecological properties and multimodal models.Released raw and preprocessed EEG, stimulus manifests, and preprocessing scripts are provided in standardized formats.Source: Usage Notes · p. 10