Abstract
Cross-dataset training for auditory attention decoding: A comparative study of canonical correlation analysis and convolutional neural network-based localization
The Journal of the Acoustical Society of America, Vol.159(4_Supplement), pp.A152-A153
08/01/2026
DOI: 10.1121/10.0045384
Abstract
Electroencephalography (EEG)-based auditory attention decoding (AAD) is a vital component of neuro-steered hearing devices designed to address the “cocktail party” problem. A significant challenge remains the effective utilization of data across diverse recording protocols and datasets. In this work, we evaluate and compare two baseline AAD approaches, canonical correlation analysis (CCA) and a convolutional neural network-based localization (CNN-Loc) method, using the public KU Leuven and DTU datasets. Notably, the CNN-Loc architecture employed in this study incorporates both spatial and temporal processing layers, differing from its original formulation. To investigate the feasibility of cross-dataset training, we constructed a combined dataset using a unified preprocessing pipeline. This included multi-channel Wiener filtering and shared EEG channel alignment. Our findings indicate that models trained on the combined data perform comparably to or better than, those trained on individual datasets. This suggests that aggregating heterogeneous data enhance model robustness. The CNN-Loc approach also showed strong results on the DTU dataset, which is surprising given earlier reports about the limited generalization of nonlinear methods in AAD. Our results suggest that combining diverse datasets is a beneficial strategy for the field. Future work will continue exploring correlation-based decoding to improve performance across diverse recordings.
Details
- Title: Subtitle
- Cross-dataset training for auditory attention decoding: A comparative study of canonical correlation analysis and convolutional neural network-based localization
- Creators
- Pooja Yakkala - Rochester Institute of TechnologySungyoung Kim - Daejeon UniversityInyong Choi - University of IowaHwan Shim - Rochester Institute of Technology
- Resource Type
- Abstract
- Publication Details
- The Journal of the Acoustical Society of America, Vol.159(4_Supplement), pp.A152-A153
- DOI
- 10.1121/10.0045384
- ISSN
- 1520-8524
- eISSN
- 1520-8524
- Language
- English
- Date published
- 08/01/2026
- Academic Unit
- Communication Sciences and Disorders; Iowa Neuroscience Institute
- Record Identifier
- 9985220839602771
Metrics
1 Record Views