FAConformer Boosts Auditory Attention Decoding by 4.9% with Frequency-Aware AI

Researchers from the team led by Ziwei Wang have introduced FAConformer, a novel frequency-aware CNN-Transformer framework designed for auditory attention decoding (AAD) from EEG signals. AAD aims to identify which speaker a listener is focusing on in a multi-speaker environment, a key enabler for neuro-steered hearing aids. While prior models have made progress, they often relied on handcrafted feature extraction or simple cross-band concatenation, failing to fully exploit frequency-domain EEG information. FAConformer addresses this by first decomposing EEG into multiple frequency bands and assigning each band to an independent CNN-Transformer encoder for band-specific modeling. A novel frequency-aware attention (FAA) module then treats these band-wise features as tokens to adaptively model cross-band dependencies, while band-wise auxiliary supervision (BAS) prevents weaker branches from being under-optimized during joint training.

Extensive experiments on two public AAD datasets at three decision-window lengths showed that FAConformer consistently outperformed 12 competitive baselines, surpassing the previous state-of-the-art model by 4.9%. Further analyses confirmed the model's effectiveness, robustness, and interpretability, including band importance and ablation studies. The framework offers a more principled way to incorporate frequency information in EEG-based auditory attention decoding, with open-source code available for reproduction. This work represents a significant step toward practical neuro-steered hearing systems that can operate reliably in real-world noisy environments.