Speaker
Description
Auditory attention decoding (AAD) aims to determine from neural recordings which sound source a listener is attending to. Here, we formulate spatial auditory attention decoding (SAAD) as a subject-independent end-to-end task, jointly modeling EEG signals and Gammatone-based audio representations to decode the attended spatial direction. While conventional approaches assume access to clean speech signals, real-world conditions provide only binaural mixtures. Therefore, we investigate the feasibility of decoding attention directly from mixture signals, in addition to ear-separated (clean) representations on the DTU and KUL datasets. Results show that decoding performance increases consistently with decision window length for all conditions. On the DTU dataset, binaural mixtures achieve performance comparable to ear-separated signals, with no statistically significant differences across decision windows. On the KUL dataset, the HRTF-based binaural condition outperforms the clean (dichotic) condition at longer decision windows, with significant improvements at 20 s and 40 s. These findings demonstrate that spatial auditory attention can be decoded directly from binaural mixtures in a subject-independent setting and highlight the potential of Gammatone-based representations for realistic AAD systems.