A Hybrid EEG-Based Approach for Classifying Sound Imagery Using Reactive Auditory Signals

Speech brain-computer interfaces (BCIs) provide a direct communication pathway for patients with neurological disorders by decoding imagined speech or content. However, imagery-based signals are typically weak, making high-quality data acquisition difficult. Training reliable models often requires repetitive and demanding imagery tasks, which can cause fatigue, reduce user engagement, and hinder the widespread adoption of BCI technologies. To overcome these limitations, we propose a novel training strategy that leverages passive listening tasks rather than active imagery for training models in active BCIs. Specifically, by comparing the neural activity patterns underlying auditory perception and imagery, we assess whether features derived from passive listening can facilitate model learning for sound imagery. Our findings suggest that listening tasks can be an effective pre-training strategy for active BCIs, potentially reducing user fatigue and improving system usability.