The glasses look like glasses. The demo is thirty seconds of someone mouthing words in an open-plan office and text appearing on a laptop, and it works well enough that the first reaction is suspicion. Radar through the cheek is not a thing most people have seen.
What kept me on the page was something else: the founder listing everything that did not work.
The list of dead ends
They tried sensors on the face first, then cameras, then muscle signals, then radar at 60 GHz. Each failed for a specific, physical reason, and the reasons are more informative than the success.
| Approach | Why it did not ship |
|---|---|
| IMU on the face | Measures aggregate motion; five words at best |
| Lipreading (camera) | 40 phonemes collapse to 12–15 visemes; P and B look identical |
| EMG sensors | Needs skin contact, awkward form factor |
| mmWave radar | Bounces off the mouth; microns of motion but 5 cm of spatial resolution |
| RF through the cheek | Their current approach |
The lipreading line is the one I cannot stop thinking about. Word error rate bottoms out around 17%, which is good enough for a demo and far from good enough for a product. Not because the models are weak — because the mouth does not carry enough information. You can add a language model to guess, and it will guess wrong on exactly the words that matter: names, numbers, the thing you were about to say.
What they are actually selling
Not the ability to type silently. The ability to type without a keyboard in public, which is a social problem more than a technical one. Voice input exists and is fine; nobody wants to dictate their email on a train. If the constraint is "don't look strange", glasses are a strange choice — people will assume you are recording them — but so is every other answer. The honest pitch is that there is no good answer yet, and this one has a working sensor.
The part I would worry about
Inference happens in the cloud. The founder says the unoptimised latency today is about 500 ms, and that local models are on the roadmap. Half a second per word is already too slow to feel like thinking; over a mobile connection it would be worse, and on a subway it would not work at all.
And this is a device that reads your mouth. Every sentence you form, including the ones you decide not to say out loud, would leave the room. They say they have safeguards so that even they cannot see the raw data. That is a promise, not an architecture, and the difference matters when the product is the input method for everything else you do.
I do not think that is a reason to dismiss the work. It is a reason to be precise about what is being sold: a sensor that works, wrapped in a service model that has not been solved yet.
https://subvocal.company/
Comments
Loading…