Ten native Mandarin speakers produced six isolated vowel rhymes, 100 repetitions per class, while OPM-MEG and EEG were recorded under the same task. The team compared six feature representations and five classifiers using time-resolved decoding, temporal generalization, pairwise classification and source-space searchlight analyses. Decoding stayed near chance before stimulus onset and rose afterward.
Averaged across classifiers over 150-500 ms, common spatial patterns (CSP) was the strongest feature representation, at 57.9% for OPM-MEG and 55.0% for EEG. Across features, a linear support vector machine (Linear SVM) reached the highest mean accuracy, 56.4% and 54.1% respectively. CSP with Linear SVM was the best combination, yielding mean accuracies of 60.3% for OPM-MEG and 56.7% for EEG, with late peak accuracies of 63.0% and 60.1%. In the pairwise analysis, CSP with Linear SVM averaged 60.1%, and the vowel contrast /i/ versus /u/ peaked at 61.2%.
Temporal-generalization matrices were dominated by a narrow main diagonal, with limited off-diagonal generalization between 50.7% and 52.1%, indicating predominantly time-specific discriminative information. Exploratory source-space searchlight analysis identified cortical parcels with above-chance local decoding from 100 to 450 ms, while no parcel reached significance at 0 or 50 ms. The authors note these patterns do not directly establish activation or coherent functional-network recruitment.
The authors present the results as support for OPM-MEG as a high-resolution non-invasive platform for time-resolved decoding, and as practical guidance for feature and classifier selection in speech-related neuroimaging and brain-computer interface studies. All accuracies come from 10 participants and six vowel-rhyme classes, and have not been validated on continuous sentences or larger vocabularies.