The Subvocal Company opened reservations on October 30, 2026, for glasses designed to turn deliberate, silent mouth movements into text. The Y Combinator-backed startup is collecting a $199 reservation deposit, with the founders targeting shipments in the first quarter of 2027. Buyers are reserving access to a future product, not ordering a finished wearable ready for delivery. (Y Combinator, launch announcement)
The pitch addresses a practical problem with AI voice tools: dictation becomes awkward in libraries, shared offices and public spaces. Founders Aarush and Arnav say their own frustration with using Wispr Flow around other people helped motivate the project.
Their proposed solution uses radio-frequency sensing, or RF, to detect articulation without requiring audible speech. The current glasses are a prototype, however, and the founders explicitly describe ongoing research and development needed to turn them into the product shown on the company’s website.
That distinction matters more now that money is involved. Subvocal has an interesting input-device concept and a demonstration described by its founders. It still has to establish how consistently that concept works across people, how much setup it requires and what customers will actually receive.
How Subvocal’s RF Sensing Works
Subvocal says antennas embedded in the glasses’ frames emit radio waves that pass through the wearer’s face. Movement of the tongue, lips and soft palate changes the signals. Machine-learning models then decode those changes into text. The company says the approach does not rely on cameras. (Founders’ technical explanation)
The useful distinction is between sensing articulation and recording sound. A microphone-based dictation tool needs an audible signal. Subvocal is trying to infer words from physical movements associated with producing them, allowing the wearer to articulate without speaking aloud.
The founders also clarify that visible lip movements are not always necessary. In an update to their announcement, they say the system can be fine-tuned for articulation with the lips closed, although learning to do that takes practice.
This is deliberate input, not a claim to read thoughts. Their explanation distinguishes the intended action from the automatic movements that may accompany thinking words to yourself. A wearer still has to perform an intentional physical action resembling mouthing words.
I’d treat that distinction as central to understanding the product. Silent speech could make dictation less disruptive, but users would still need to learn an input technique. The practical question is how comfortably and reliably they can repeat that technique during ordinary work.
The founders compare the emitted radio waves to Bluetooth. That comparison should not be read as a complete radio specification: their launch explanation does not supply the operating frequency, transmission power or other details needed to characterize the sensing hardware.
The Prototype’s Hard Problem Is Working Across Faces
The founders trace the project to Stanford’s TreeHacks in February, where they tried to build a silent-speech interface in 36 hours. Their first attempt used electromyography, or EMG, which measures electrical activity associated with muscle movement. They say that early setup worked imperfectly at the letter level. (Development account)
They subsequently abandoned that approach, citing sweat-related signal changes, electrical interference, placement variability and the inconvenience of attaching electrodes to a face. Those are their reasons for changing direction, not evidence that EMG cannot work in other silent-speech systems.
Switching to RF introduced another set of engineering problems. The founders describe building low-noise sensing electronics and positioning antennas so their fields overlapped with the parts of the mouth involved in articulation. Their first glasses-based attempt failed, they say, prompting experiments with taped-on antennas and a head rig before returning to antennas embedded in frames.
The next challenge was making the model work for different wearers. As the founders put it:
“Everyone’s facial anatomy, and hence RF signal patterns, are vastly different.”
Their response was to collect more training data. They describe setting up booths in their office and paying people to read prompts while wearing the glasses, then using those recordings to pretrain their models. They report that errors declined as the dataset grew.
This development history reveals an important dependency: the glasses’ usefulness rests on both hardware positioning and model generalization. Detecting a signal from one person is only part of the problem. The system must also map signals from different faces to the intended words.
That makes frame fit potentially consequential beyond comfort. Based on the founders’ account of antenna positioning, it is reasonable to ask whether changes in how the glasses sit affect decoding. The announcement does not quantify that sensitivity, so it remains a product question, not an established flaw.
What the Speed and Accuracy Claims Leave Open
Subvocal’s launch claims include continuous silent-speech decoding at more than 200 words per minute and high accuracy. Those are company-reported performance claims. The cited launch sources do not provide independent validation across users or real-world conditions, or a benchmark protocol detailed enough to establish typical customer performance. (Launch discussion)
A rate of 200 words per minute corresponds to about 3.3 words per second. That would be a substantial text-entry rate if the output remained accurate and the wearer could sustain it comfortably. But an input rate is not the same thing as finished-writing throughput.
For a practical comparison with typing or microphone dictation, several measurements matter:
- Error rate: How many words require correction, and how costly are those corrections?
- Latency: How long passes between articulation and usable text appearing?
- Calibration: What must a new wearer do before reaching the advertised performance?
- Consistency: Does performance hold across people and repeated sessions?
The founders say the person featured in their linked demonstration completed a 20-sentence calibration. That is a useful setup detail, but it describes that demonstration. It is not a published guarantee that every buyer will need the same calibration or achieve the same result.
Pretraining and calibration also solve different parts of the problem. Broad pretraining is intended to teach patterns across many people; calibration adapts the system to an individual wearer. The fact that the founders scaled their dataset does not, by itself, tell buyers how much individual setup the shipping product will require.
My view is that corrected text-entry speed would be the most useful eventual measure. A fast stream of output is promising, but the advantage shrinks if users repeatedly stop to fix misrecognized words. That is a measurement priority, not a conclusion about how this prototype performs.
What the $199 Reservation Actually Covers
The YC company listing describes a $199 reservation deposit toward a discounted early-batch price and identifies the product as under development. The founders’ announcement describes an initial batch of 500 units, a $200 discount from the final price and a Q1 2027 shipping target.
| Item | What the launch sources say |
|---|---|
| Payment now | $199 reservation deposit |
| Initial batch | 500 units |
| Early-batch offer | $200 off the final price |
| Shipping target | Q1 2027, meaning January through March |
| Product status |
Frequently Asked Questions
4 questions
1Do Subvocal Glasses Read Your Thoughts?
No. The founders describe decoding deliberate articulation, including movements made with the lips closed. They distinguish this from simply thinking words.
