Nature Neuroscience· 2024
High-performance non-invasive neural speech decoding from magnetoencephalography and high-density EEG
D. Défossez, C. Caucheteux, J. Rapin, O. Orien, J.R. King
Significance
For: researcher
The problem
Low signal-to-noise ratio in non-invasive neural recordings previously made continuous phrase decoding inaccurate without surgery.
Method
Contrastive neural language representations aligned with multi-channel MEG and 128-channel EEG spectrograms using causal dilated convolutions.
Result
Achieved top-10 accuracy of up to 73% in 3-second speech segment identification from non-invasive sensors.
Why it matters
Proves that deep language models can extract rich linguistic structures from scalp recordings without surgical electrode placement.
Limitations
Requires high trial counts and subject-specific calibration; performance drops substantially with sparse consumer headsets.
Neural DecodingMEGEEGSpeech BCI