Researchers at Hugging Face have redesigned a high-performing magnetoencephalography (MEG)-to-speech retrieval decoder, achieving a 20-fold reduction in parameters while maintaining state-of-the-art accuracy. The model reaches 39.75% Top-1 accuracy among 1,005 speech candidates on the MEG-MASC benchmark, with about 20 times fewer decoder parameters than previous systems.
The new architecture replaces the spatial attention mechanism with spherical harmonics defined on the three-dimensional MEG helmet geometry, and reduces the subject-specific representation from 270 to 25 branches. Each branch includes a temporal filter to match neuronal sources in space and time, and the convolutional decoder is made shallower. Ocular and cardiac components are removed before training to prevent stimulus-locked shortcuts.
Critically, the redesigned model is interpretable: its weights map to source space, recovering generators consistent with the speech-perception network. Left-lateralized branches carry higher-frequency rhythmic components not evident on the right. Paired MEG occlusion experiments reveal that 15 of 19 stimulus features contribute to retrieval, with the largest effects for silence, sound intensity, vowels, and acoustic onsets.
Interestingly, random word lists behave oppositely: substituting narrative MEG into them improves retrieval, indicating that activity without narrative structure carries less recoverable information. The wav2vec target can be reduced to about twelve learned feature dimensions without loss of accuracy, whereas strong temporal compression causes a clear loss.