Subvocal's AI glasses let you type 200 wpm silently

Talking to your AI assistant in a packed train makes you look unhinged, but typing on glass caps your brain at 40 words per minute. That friction has turned input bandwidth into the biggest bottleneck of the agent era.

A new startup called Subvocal thinks it has found the escape hatch. Founded by two Georgia Tech dropouts who met at a Stanford hackathon, the company just raised a $1.3 million pre-seed round led by Afore Capital, with backing from Y Combinator and angel investors, to turn silent speech into high-speed text.

Pre-orders are open at an early-bird price of $499, against a planned $699 retail tag, with shipments scheduled for the first half of 2027.

Radar on your face, zero cameras

Subvocal looks like an ordinary pair of glasses, but it completely skips cameras and microphones. It does not read your brain waves, and it does not spy on your surroundings.

Instead, the frames bounce low-power, radar-like radio frequency signals across your lower face. When you deliberately mouth words without vocalizing, speech muscles shift by micrometers. Those tiny movements alter the reflected RF waves.

An on-device neural network translates those signal shifts into phonetic tokens, and an LLM downstream fills in the blanks for flawless syntax. Subvocal claims its prototype clocks continuous transcription at over 200 words per minute, far eclipsing standard typing and easily matching natural conversation.

The AlterEgo pivot: Why now?

If this sounds familiar, that is because MIT Media Lab attempted something similar in 2018 with AlterEgo. That project used sticky electrodes on the jaw to read surface electromyography signals, achieving 92% accuracy.

The catch back then was brutal: it only worked on a minuscule, predefined vocabulary like numbers and basic commands. Expanding that to open-ended prose was impossible without today's AI stacks.

Eight years later, cheap RF sensors solve the sweaty skin-patch issue, while modern language models handle the heavy lifting of semantic error correction. If the radar drops a syllable, the language model predicts what you meant from context.

Why it matters

When AI software can write an entire app or research brief in seconds, human input speed is the primary drag on productivity. For developers directing autonomous agents, jumping from 40 words per minute on a keyboard to 200 words per minute silently changes the entire cadence of work.

We think Subvocal made a brilliant design call by ditching optical cameras. Smart glasses with lenses face massive social stigma and outright bans in private spaces; RF sensors sidestep privacy concerns entirely while remaining immune to street noise.

Still, we would advise caution before putting down the $199 deposit. A shipping window in early 2027 means this is still early-stage hardware. Every human face has different jaw geometry, fat distribution, and speech quirks, so training an RF model that works across diverse populations without exhausting calibration is an enormous hill to climb. But if they pull it off, tapping glass might soon feel as archaic as punch cards.