The brain's frequency atlas

The raw EEG trace looks like noise because it is a mixture: a dozen processes at different speeds, all running at once, summed into one wiggle. The first act of any pipeline is unmixing by frequency. The standard atlas, built over a century of electroencephalography, gives each band a name and, more usefully, a meaning.

The bands and what they are doing

Delta, below four hertz, belongs to deep sleep and slow repair. Theta, four to eight, marks memory work and drowsiness. Alpha, eight to twelve, is the idle rhythm of a relaxed visual cortex, prominent at the back and famously blocked the moment you open your eyes. Beta, roughly twelve to thirty, accompanies active engagement and holds the motor strip's working rhythm, which dips when a movement is planned, a phenomenon called event-related desynchronization. Above thirty hertz begins gamma, and at its upper edge, roughly seventy to one-seventy, sits high-gamma: broadband power that tracks local cortical population activity more faithfully than any slower band. On electrocorticography, where electrodes sit on the exposed cortex, high-gamma amplitude is the single best correlate of speech production, rising and falling almost syllable by syllable. That is the prize band, and note four already explained why the skull taxes it hardest.

Slower rhythms tell you the state of the brain. The fast bands tell you what it is saying.

Two kinds of structure, both useful

Rhythms are one kind of structure; the other is the event-related potential, the small consistent wiggle time-locked to a specific moment, exposed by averaging many trials so that random activity cancels. The classic BCI demonstration, the P300 speller of 1988, used exactly this: a spike of positive voltage about three hundred milliseconds after the row containing the letter you attend to flashes. It spelled at a few characters a minute by watching for attention itself. Modern decoders replaced averaging with learned features, but the discipline underneath is unchanged: time-lock to events, compare against a baseline, and let repetition expose the signal.

From scalp to samples

Digitization obeys Nyquist: to capture a frequency you must sample at least twice as fast, so a hundred-and-seventy-hertz band wants several hundred samples a second, which is why EEG systems record at 250 to 1000 hertz rather than more or less at random. Before analysis, the pipeline band-passes the signal to keep the interesting range, notches out the fifty or sixty hertz power-line hum, and cuts the recording into epochs time-locked to word onsets with a baseline before each. Then comes the unglamorous hero: artifact handling. Independent component analysis decomposes the recording into statistically independent sources and lets you drop the eye-blink and jaw-muscle components while keeping the neural ones. Every competent lab runs this liturgy, and the open-source MNE-Python package is the standard implementation, which is worth knowing because it means the entire preprocessing stack our study needs is free, inspectable, and already battle-tested by thousands of papers.

The atlas closes the loop with our plan: the speech payload concentrates in bands the skull attenuates, so the pipeline is designed to preserve and denoise exactly those, and the recordings are structured so every epoch carries its word label. Frequency by frequency, this is what Phase 1 buys.