Somach builds brain-computer interfaces that stay outside your head: no surgery, but a noisy signal. The bet is that knowing the context of your life helps a decoder cut through the noise.
Subvocalizing is talking with your mouth closed: you form a word without voicing it, and the muscles of your jaw and throat still fire, faintly. A forty-dollar skin sensor can tell a handful of words apart from that signal, as long as the electrode doesn't move.
A handful of words isn't a language, and a better electrode alone won't fix that. Context might, starting with what you were doing when the signal fired. That record stays private.
The priors come from a model of what you are likely to do next, so a decoder can favor the likelier words. Over 21 held-out days, a version trained on my own writing picked what I actually did out of five options 37.8% of the time, against 28.0% untrained and 20% for chance.
Honestly, it's early. The sensor works for one speaker on a small vocabulary and breaks if it moves a centimeter. So far, giving the model more context has made it slightly worse, which is the opposite of the bet and what I'm working on now.
Somach Systems is one person in San Francisco, building brain-computer interfaces (-ish). It's a research bet with no ship date. Leave an email if you work on BCIs, can't always rely on your voice, or just want to watch.
I am checking whether a summary of my day predicts my next move better than what is on my screen. The result stays sealed until the model is frozen. That way I cannot tune my way to the answer I want.