Interpretable MEG Decoding of Perceived Speech: Cortical Sources and the Stimulus Features That Drive Retrieval
The paper makes a MEG speech-retrieval model smaller and more readable while keeping strong accuracy.
The authors replace the model’s sensor-layout attention with spherical harmonics tied to the MEG helmet geometry, then map learned branches back toward cortical sources. On MEG-MASC, it reports 39.75% Top-1 accuracy among 1005 candidates, using about 20 times fewer decoder parameters. The analysis points to silence, sound intensity, vowels, and acoustic onsets as major retrieval drivers. Random word lists produced weaker recoverable information than coherent narrative speech. HF Daily Papers' note
The authors replace the model’s sensor-layout attention with spherical harmonics tied to the MEG helmet geometry, then map learned branches back toward cortical sources. On MEG-MASC, it reports 39.75% Top-1 accuracy among 1005 candidates, using about 20 times fewer decoder parameters. The analysis points to silence, sound intensity, vowels, and acoustic onsets as major retrieval drivers. Random word lists produced weaker recoverable information than coherent narrative speech. HF Daily Papers' note
score 4