Title: A Channel-Centric Approach to Reconstructing Music from EEG Signals

Abstract:

While brain-computer interfaces have made strides in decoding naturalistic stimuli from neural data, current research has predominantly targeted vision and language tasks. This study investigates a more complex and underexplored domain: reconstructing music from electroencephalography (EEG) recordings. In this context, neural signals are characterized by their weakness, distributed nature, and high vulnerability to noise and variations across recording channels. Our primary discovery reveals that premature mixing of channel data erases subtle but critical EEG signals.

To mitigate this issue, we introduce a channel-oriented architecture comprising three essential elements. First, channel-wise tokenization processes each electrode as a distinct token, thereby preserving spatially specific neural evidence. Second, channel-wise multi-view self-distillation fosters consistency between temporal crops and random subsets of channels, facilitating the learning of robust, distributed representations. Third, channel-wise data augmentation employs structured channel dropout to enhance the model’s resilience against noise, artifacts, and missing electrode data. Collectively, these strategies maintain weak yet informative signals across channels, allowing for stable alignment with a semantic music representation space.

We embed this channel-oriented framework into an encoding-alignment-decoding pipeline designed for EEG-to-music reconstruction. Theoretically, we define the conditions under which maintaining channel-level structure enhances alignment accuracy. Empirically, our approach is benchmarked against various state-of-the-art baselines, demonstrating consistent and statistically significant improvements in performance.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC