Year: 2026
Type: conference-paper
Abstract: Auditory attention decoding (AAD) identifies the attended speaker from neural activity in multi-talker environments.Conventional AAD methods often rely on laboratorygrade EEG, which limits their performance in dynamic, realworld settings.This study presents a subject-independent pipeline for mobile EEG using contrastive learning (CL) framework.We evaluate the framework on 44-channel mobile EEG data from 24 normal-hearing participants across three naturalistic conditions: sustained attention to one of two speakers, attention switching, and a two-talker conversation competing with a single speaker.The proposed architecture integrates a three-model deep learning pipeline with a contrastive objective to align EEG embeddings with speech representations.Results show that the CL-based approach outperforms both linear and nonlinear models that do not incorporate a contrastive objective.A leave-one-groupout training protocol verifies that the framework generalizes to unseen subjects.Evaluation across 2-30 s time windows results in classification accuracies between 61.4% and 79.1%.These findings indicate that CL-based discriminative mapping provides a reliable approach for subject-independent AAD in realistic environments.
Institutions: Oticon (Denmark), Aalborg University, Linköping University
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7162613118