Source-linked AI summary
A multi-scenario EEG dataset for auditory attention decoding in naturalistic multi-talker environments
Shu Peng, Rui Liu, Yufei Zhang, Wenlong You, Zhige Chen, Jiachen Xi, Qiyuan Sun, Yan Liu, Kay Chen Tan, Jibin Wu
TL;DR
Many open EEG datasets for auditory attention use simplified single-competing-talker paradigms, leaving naturalistic multi-speaker listening less represented. SoundBubble-EEG records EEG during themed group conversations across three scenarios. Its speech-tracking advantage remained positive across all three scenarios.
Problem
Most open EEG datasets for auditory attention use idealized single-competing-talker paradigms, while filtering collective group conversations remains largely unexplored.
Source: Background · p. 1Method
The dataset records EEG during a Sound Bubble task in which listeners attend to a themed group of speakers amid competing speech and ambient noise.
Source: Background · p. 1, Stimuli · p. 2Results
Mean reconstruction advantage remained positive in Restaurant, TV, and Meeting, with mean Δr values of 0.0252, 0.0193, and 0.0092, respectively.
Source: Application Baseline · p. 10Takeaways & Limitations
The dataset provides a naturalistic resource for studying auditory attention and evaluating AAD across varied acoustic scenarios.
Source: Background · p. 1
Abstract
from arXiv · showhide
Understanding how the brain selectively follows relevant speech amid competing voices is a central challenge in auditory neuroscience and a key step toward neuro-steered hearing technologies. However, most open-source Electroencephalography (EEG) datasets for Auditory Attention Decoding (AAD) use idealized single-competing-talker paradigms that oversimplify the acoustic, spatial, and semantic structure of everyday communication. To capture this ecological complexity, we introduce the SoundBubble-EEG dataset: a high-density 128-channel EEG resource comprising more than 25 hours of recordings from 30 participants. The paradigm requires listeners to selectively attend to a dynamic target speaker group, a designated "sound bubble", amid competing multi-speaker distractor bubbles across three realistic scenarios: a restaurant, a home TV viewing, and a meeting discussion. By bridging the gap between constrained laboratory protocols and real-world auditory scenes, this dataset enables investigations of multi-talker speech comprehension, neural speech tracking, and cross-scenario generalization. It also provides a benchmark for AAD algorithms under realistic acoustic and semantic variability and may support auditory neuroscience and the development of neuro-steered hearing technologies.
Background
Existing open EEG datasets largely use idealized single-target, single-interferer paradigms, leaving multi-speaker group conversations underrepresented. SoundBubble-EEG introduces a high-density dataset spanning realistic listening scenarios to support auditory-attention research and algorithm evaluation.
Background
- EEG’s millisecond-level temporal resolution captures rapid speech-comprehension dynamics, connecting auditory-attention research with neuro-steered hearing technologies.The paper presents this capability as a foundation for AAD research.Source: Background · p. 1
- Most open-access EEG datasets use idealized single-target, single-interferer setups that do not capture the multi-speaker conversational interference common in daily life.The neural dynamics and cognitive load involved in filtering a group conversation remain largely unexplored in available datasets.Source: Background · p. 1
- The Sound Bubble paradigm replaces a single competing talker with themed groups of speakers, requiring attention to a target group amid distractor speech and environmental sound.The dataset covers meeting, restaurant, and living-room settings.Source: Background · pp. 1–2
- The dataset combines over 25 hours of EEG with behavioral logs and spatialized audio for studying neural speech tracking and evaluating AAD robustness across scenarios.The authors describe it as a resource for cognitive neuroscience, machine learning, and biomedical engineering.Source: Background · p. 2
- The dataset workflow spans listening scenarios, EEG processing and validation, and public release in a BIDS-compliant format with analysis code.Validation includes behavioral, signal-quality, and neural/application-level analyses.Source: Figure 1 · p. 2
Methods
The study collected high-density EEG during a Sound Bubble listening task involving themed speaker groups across three scenarios. The protocol combines controlled spatialized stimuli, comprehension checks, and a standardized preprocessing pipeline.
Data release
- The release package includes raw EEG, preprocessed derivatives, behavioral logs, spatial audio stimuli, metadata, and analysis code.The data follow BIDS organization.Source: Methods · p. 2
Stimuli
- A Sound Bubble is a cohesive group of at least two speakers discussing a theme; 25 spatial-audio segments covered restaurant, home-TV, and meeting scenarios.The scenes varied in noise and spatial configuration, using speech, environmental noise, and HRTF-based spatialization.Source: Stimuli · pp. 2, 4
Experimental procedure
- Participants completed 25 trials in a fixed Meeting–Restaurant–Home TV order, with 5, 10, and 10 trials respectively, and answered general and detail-comprehension questions after listening.Each trial included a 3-second baseline and a 120-second active-listening period; breaks were self-paced.Source: Figure 2 · p. 3, Experimental Procedure · p. 5
Participants and acquisition
- EEG was recorded from 30 healthy volunteers using a 128-channel HydroCel Geodesic net at 250 Hz.Participants were native Chinese speakers with self-reported normal hearing; online reference was Cz.Source: Participants · p. 2, Data Acquisition · p. 6
Data preprocessing
- Preprocessing identified and interpolated bad channels before common-average re-referencing, filtered data to 0.1–40 Hz, and used ICA for artifact removal.ICA was fitted on a 1-Hz high-pass copy, with its weights applied to the 0.1–40 Hz data; epochs separated the 3-second baseline from the 120-second task.Source: Preprocessing Pipeline · p. 6
Data Records
The public release provides the dataset files in BIDS-organized form, including raw and processed recordings, behavioral data, and stimulus materials.
Data Records
- The complete dataset is publicly available on Zenodo and organized according to BIDS, with raw EEG, preprocessed derivatives, behavioral logs, and stimulus materials.Spatial audio files for the three scenario categories are included in the BIDS stimuli directory.Source: Data Records · p. 6
- The directory structure follows BIDS; the illustrated subject/run includes EEG, metadata, tabular, and text files.Only sub-01/run-01 is fully expanded; other subjects and runs share the structure.Source: Figure 3 · p. 7
Technical Validation
The validation assesses participant engagement, signal quality, neurophysiological signatures, and application-level AAD performance. Results support attended-speech tracking across the three scenarios, alongside behavioral and signal-quality checks.
Behavioral performance
- Q1 accuracy was near ceiling across scenarios, while Q2 accuracy had greater variance and slightly lower medians; Restaurant median accuracy remained high.Q1 assessed general understanding; Q2 assessed specific details in the multi-speaker scenes.Source: Behavioral Performance · pp. 7–8, Figure 4 · p. 8
Signal quality assessment
- Median bad-channel counts were approximately 10–13 of 128 across scenarios, with similar distributions; occasional runs had up to 30 channels interpolated.The outliers were primarily attributed to hair thickness or localized cap tension and were recovered by spherical-spline interpolation.Source: Signal Quality Assessment · p. 8, Figure 4 · p. 8
- Bad-channel probability was elevated at some peripheral/facial and midline/vertex sensors; the map is a quality-control profile, not evidence of condition-specific degradation.The channels were interpolated before re-referencing.Source: Signal Quality Assessment · p. 8, Figure 4 · p. 8
- Median ICA rejection was stable at 3–4 components per run across scenarios, with no significant condition-specific shifts.Artifactual components with over 80% classification probability were automatically excluded.Source: Signal Quality Assessment · pp. 8–9, Figure 4 · p. 8
Neurophysiological sanity checks
- Task EEG showed a 1/f spectral decay and spatially heterogeneous alpha modulation, with temporal/frontal suppression and parieto-occipital enhancement.The posterior alpha increase is described as consistent with cross-modal visual inhibition during sustained auditory attention.Source: Neurophysiological Sanity Checks · p. 9, Figure 5 · p. 9
Application-level AAD validation
- p = 9.31×10−10: a scenario-pooled backward mTRF model showed significantly higher reconstructed-envelope correlation for attended than unattended speech.The model used leave-one-trial-out cross-validation across all 25 trials.Source: Figure 5 · p. 9, Application Baseline · p. 10
- None of 50,000 label-swapped permutations reached the observed Δr = 0.0196 or decoding accuracy = 0.717 (permutation p = 2.0×10−5).The authors report this control rules out nonspecific reconstruction bias or chance alignment as explanations for the decoding effect.Source: Signal Quality Assessment · p. 9, Application Baseline · p. 10
- Mean Δr stayed positive across Restaurant, TV, and Meeting at 0.0252, 0.0193, and 0.0092, respectively.Single-trial accuracy followed the same pattern: 0.767, 0.720, and 0.613; Meeting had five trials per subject versus ten in the other scenarios.Source: Signal Quality Assessment · p. 9, Application Baseline · p. 10
Usage Notes
The release offers raw and preprocessed EEG in BIDS-organized formats, with guidance for choosing between them. Its scenarios and paired behavioral measures support analyses of cross-condition decoding, brain–behavior relationships, and spatial attention.
Data access and analysis
- BIDS-organized release includes raw EEG in .mff, preprocessed derivatives in .fif, behavioral logs, stimuli, metadata, and analysis code.Raw data suit custom preprocessing and artifact-rejection validation; derivatives suit decoding or cognitive modeling.Source: Usage Notes · p. 10
- The authors recommend MNE-Python version 1.11.0 or later for the standardized derivatives and event dictionaries.Source: Usage Notes · p. 10
Research opportunities
- The three scenarios and paired behavioral measures permit cross-condition AAD studies, brain–behavior analyses of comprehension and listening effort, and continuous spatial-attention decoding.Source-position metadata also supports investigating dynamic functional connectivity under varying acoustic complexities.Source: Usage Notes · p. 10