FAConformer: Frequency-Aware Convolutional Transformer for Auditory Attention Decoding

The paper introduces FAConformer, a frequency-aware CNN-Transformer architecture designed for auditory attention decoding (AAD), which enhances the extraction of frequency domain information from EEG signals. FAConformer employs independent CNN-Transformer encoders for multiple frequency bands and utilizes a frequency-aware attention module to model cross-band interactions, achieving a performance improvement of 4.9% over the previous state-of-the-art on two public AAD datasets. This framework's ability to effectively leverage band-specific information and prevent under-optimization through band-wise auxiliary supervision makes it a significant advancement for practitioners in neuro-steered hearing systems.