02 · Neurotechnology

Decoding the Silent Mind.

Recent work has shown AI systems that interpret the vocalisations of animals. The far more consequential frontier is the inverse-facing one: people who have language but have lost the channel to express it.

wrist · HRV · skin-conductance
context · calm
EEG cap → signal → decoder → assembled words. Wrist biosignals run in parallel as a context channel.

The problem

Millions live with intact cognition but a broken output channel: locked-in syndrome, late-stage ALS, post-stroke aphasia, severe cerebral palsy, minimally-verbal autism. The bottleneck is not thought — it is the motor pathway between thought and speech. Eye-gaze keyboards and switch scanning are slow, fatiguing, and degrade as the condition progresses.

State of the art — and where it stops

Invasive BCIs (Neuralink; Stanford's Willett lab) reach the highest accuracy. Stanford has decoded inner speech — silent, unspoken language — from implanted motor-cortex arrays, hitting ~74% accuracy against a ~125,000-word vocabulary. Powerful, but it requires neurosurgery.

Non-invasive EEG decoding (DeWave; Meta's EEG-to-text work) skips surgery entirely: a snug sensor cap reads scalp signals, and a model paired with an LLM translates patterns into words.

Invasive
~74% accuracy
High signal · neurosurgery required · does not scale
Non-invasive · our focus
Lab-bound, but accessible
Snug cap · no surgery · accuracy fragile in the field

The gap is sharp: invasive wins on accuracy but cannot scale; non-invasive is accessible but remains lab-bound — fragile across people, sessions, and electrical environments. Almost nothing addresses messy real-world deployment.

Our distinct angle — robustness, not raw accuracy

We do not try to out-accuracy Neuralink. We make non-invasive brain-to-text actually work outside the lab.
  • Robustness across people and conditions. Domain adaptation so one model generalises across patient populations, cognitive states, electrical interference, and varying signal quality.
  • Multimodal fusion. EEG carries the what of intended language. A wrist-worn device adds a parallel context channel — heart-rate variability, skin conductance — that cannot read thought but can disambiguate intent and flag urgency or distress.
  • Real-time, low-latency decoding. Suitable for a bedside or care-home conversation, not a thirty-second-per-word crawl.

Why now

  • Demand. Ageing populations, rising stroke and ALS prevalence, and a global shortage of speech-language assistive infrastructure.
  • Enabler. Foundation models and cheap, portable EEG hardware have crossed a practicality threshold in the last eighteen months.
  • White space. The field is crowded with proof-of-concept papers and near-empty on clinically deployable, person-agnostic systems. That is the research contribution.

Next · 03 — Beyond the Interface