A Generalizable OPM-MEG Framework for Time-resolved Language Decoding During Natural Speech Production

Non-invasive decoding of rapidly evolving cognitive and motor states is limited by trade-offs between temporal precision, spatial fidelity and tolerance to natural movement. We present a modality-matched benchmarking framework that evaluates optically pumped magnetometer magnetoencephalography (OPM-MEG) against electroencephalography (EEG) during visually cued overt speech production. Ten native Mandarin speakers produced six isolated vowel rhymes (100 repetitions per class) while OPM-MEG and EEG were recorded under the same task. We compared six feature representations and five classifiers using time-resolved decoding, temporal generalization, pairwise classification, and source-space searchlight analyses. Decoding remained near chance before stimulus onset and increased after onset. In the principal common spatial patterns (CSP) analysis, exact paired cluster-mass permutation tests identified significant OPM-MEG > EEG clusters for all five classifiers within 150-500 ms. Averaged across classifiers over 150-500 ms, CSP was the strongest feature representation (57.9% for OPM-MEG and 55.0% for EEG). Across features, linear support vector machine (Linear SVM) achieved the highest mean accuracy (56.4% and 54.1%, respectively). CSP with Linear SVM was the best feature-classifier combination, yielding mean accuracies of 60.3% for OPM-MEG and 56.7% for EEG; late peak accuracies reached 63.0% and 60.1%, respectively. Temporal-generalization matrices were dominated by a narrow main diagonal, with limited off-diagonal generalization (50.7-52.1%), indicating predominantly time-specific discriminative information. In the pairwise analysis, CSP with Linear SVM achieved a mean accuracy of 60.1%, and /i/ versus /u/ reached a maximum of 61.2%. Exploratory source-space searchlight analysis identified time-evolving cortical parcels with above-chance local decoding from 100 to 450 ms, whereas no parcel reached significance at 0 or 50 ms; these patterns do not directly establish activation or coherent functional-network recruitment. These results support OPM-MEG as a high-resolution non-invasive platform for time-resolved decoding and provide practical guidance for feature and classifier selection in speech-related neuroimaging and brain-computer interface studies.