INTERPRETABLE MEG DECODING OF PERCEIVED SPEECH: CORTICAL SOURCES AND THE STIMULUS FEATURES THAT DRIVE RETRIEVAL

Modern deep learning models can retrieve speech segments from non-invasive magnetoencephalography (MEG) data with high accuracy, but these models typically function as black boxes. This study asks whether it is possible to build a high-performing decoder that is also interpretable, allowing researchers to identify the specific cortical sources and stimulus features that drive the model's decisions.

The researchers redesigned a standard MEG-to-audio retrieval architecture by incorporating physical and physiological constraints. Key modifications include replacing 2D Fourier spatial attention with spherical harmonics (which better match the geometry of the MEG helmet), reducing the latent representation to 25 branches, and adding trainable temporal filters to each branch. This design forces the network to learn representations that correspond to specific neural sources in space and time. The team also pre-processed the data to remove ocular and cardiac artifacts, which could otherwise act as stimulus-locked shortcuts.

The proposed model achieves a Top-1 retrieval accuracy of 39.75% on the MEG-MASC dataset while using 20 times fewer parameters than previous approaches. By mapping the learned weights back to source space, the authors identified generators consistent with the canonical speech-perception network, with left-lateralized branches exhibiting higher-frequency rhythmic components. Through paired MEG occlusion—a technique where specific stimulus features are replaced by 'donors'—the researchers found that silence, sound intensity, vowels, and acoustic onsets are the primary drivers of retrieval. Interestingly, the model performs better on coherent narrative speech than on randomly ordered word lists, suggesting that the brain's response to narrative structure is critical for recoverable information.