A magnetoencephalograph, a helmet-like device with hundreds of sensors, picks up the weak magnetic fields generated by active neurons. When a person listens to speech, these signals contain imprints of the heard words. But extracting them is tricky: signals from thousands of sources merge into a tangled mess.
The research team rebuilt the architecture of a neural network that learns to match magnetic patterns to audio tracks. Instead of treating the sensor array as a flat map, the network now views the helmet as a sphere and decomposes the signal into spherical harmonics — much like a prism splits white light into colors. The number of tunable channels was reduced from 270 to 25, but each now includes a temporal filter to capture the rhythm of individual neuron groups. Heartbeats and eye movements were cleaned out beforehand, so the network couldn't cheat by relying on extraneous rhythms.
The real surprise came from looking inside. The trained network's weights could be projected back into brain space, and they landed precisely in the regions that neurophysiologists associate with speech perception. The left hemisphere carried higher-frequency rhythms than the right. When the authors started "switching off" different sound properties one by one, they found that pauses, loudness, vowels, and sound onsets had the strongest impact on speech reconstruction. The brain encodes a random word list more sparsely than a living story — meaningful speech leaves a clearer trace in the magnetic field.
So a helmet that used to just listen to neuronal noise is becoming a tool that can not only guess words but also show where meaning arises in the brain.
A neuron's magnetic field is a hundred million times weaker than Earth's magnetic field — to catch it, you need a room shielded from all external fields.
Science fiction writer William Gibson described a similar technology: in his novels, neural interfaces directly read people's experiences, blurring the line between the inner world and the recording.