MEG speech decoding hits 39.75% accuracy with 20x fewer parameters

Researchers at the intersection of machine learning and neuroscience have made a major advance in decoding perceived speech from non-invasive brain recordings. The team, led by Ilia Semenkov, redesigned a CLIP-style MEG-to-audio retrieval architecture by replacing spatial attention on flattened sensor layouts with spherical harmonics on the 3D MEG helmet geometry. They reduced subject-specific branches from 270 to 25, added temporal filters to align each branch with neuronal sources, and made the convolutional decoder shallower. After removing ocular and cardiac artifacts, the model achieved 39.75% ± 0.34% Top-1 accuracy among 1005 candidates on the MEG-MASC benchmark—with roughly 20 times fewer decoder parameters than prior systems.

The model's weights map cleanly to source space, recovering generators consistent with the known speech-perception network. Notably, left-lateralized branches carry higher-frequency rhythmic components not present on the right. Paired MEG occlusion experiments revealed that 15 of 19 stimulus features contribute to retrieval, with silence, sound intensity, vowels, and acoustic onsets having the largest effects. Random word lists behaved differently—substituting narrative MEG data improved retrieval, suggesting coherent speech carries more recoverable information. The wav2vec target could be compressed to about 12 learned feature dimensions without accuracy loss, though strong temporal compression degraded performance. This work makes brain decoding more interpretable and efficient, paving the way for real-time speech neuroprostheses.