Subvocal: Type by "Silently Mouthing Words" Just by Wearing Glasses

Voice input is already quite usable, but in a library, open office, or on the subway, speaking to your computer is still an awkward thing to do. Two Georgia Tech dropouts wanted to make it possible for you to just move your lips—or even silently mouth words with your mouth closed—and have the computer turn it into text. Their product is called Subvocal, and it takes the form of a pair of smart glasses.

From facial electrodes to RF glasses: a complete technical pivot

Aarush and Arnav initially replicated MIT Media Lab's AlterEgo approach at Stanford's TreeHacks hackathon—attaching EMG (electromyography) electrodes to the face to read the electrical signals from facial muscles during speech. The prototype built in 36 hours barely worked at the letter level and didn't win an award, but it convinced them to go all in.

The EMG route quickly revealed serious flaws: sweating changes the signal, electrical interference is severe, and slight electrode position deviations produce different results. More realistically, nobody wants to stick electrodes on their face every day.

So the team pivoted to a radio-frequency (RF) approach: antennas embedded in the glasses frame emit Bluetooth-like radio waves that penetrate the face. As the waves pass through the tongue, lips, and soft palate, they are modulated by the user's silent mouth movements, and a foundation model trained on this data decodes them into text.

This pivot wasn't just a simple sensor swap. The team first built an extremely low-noise RF sensing chain, then repeatedly adjusted antenna placement to maximize the overlap between the electromagnetic field and the articulatory organs. The first version of the glasses failed completely; they had to tape the antennas back onto the face and built an odd but effective head rig to keep antenna position and angle consistent. Only after repeated refinement in simulation did they arrive at an antenna layout that could fit inside the frame.

EMG (electromyography) is a technique that uses surface electrodes to capture the faint electrical signals (typically microvolt to millivolt level) generated when muscles contract, in order to infer movement intent. In silent speech scenarios, even when a person makes no sound, the tongue, lips, and throat muscles still produce measurable electrical activity—projects like AlterEgo use this to convert "silent mouthing" into text. EMG's advantages are high temporal resolution and direct reflection of neuromuscular commands; its drawbacks are small signal amplitude, extreme susceptibility to sweat (which changes skin impedance), electrode displacement, and surrounding electromagnetic interference, plus the need for skin contact, which makes long-term wear unpleasant. This is precisely why Subvocal abandoned this route and turned to contactless RF sensing.

Works even with your mouth closed, but takes a little practice

Worth noting: Subvocal doesn't require you to actually move your lips or open your mouth. Because RF signals can "see" inside the oral cavity, the system can be fine-tuned for "inner speech"—not the involuntary tongue movements of silent reading, but mouthing words with your mouth closed. The author admits the motion feels a bit strange, but with practice it can be mastered quite well.

In the demo video, YC's Dave needed only about 20 sentences of calibration to start typing with the glasses. That amount of calibration is quite rare for a task like silent speech, because each person's facial anatomy and corresponding RF signal patterns differ enormously.

Silent speech refers to input where the user produces no audible sound and communicates only through articulatory organ movements—an independent research direction in voice interaction. Its key difference from traditional ASR (automatic speech recognition) is that there is no usable acoustic signal, so the system must rely on other modalities—EMG, ultrasound, radar, or RF—to capture articulation. Such tasks typically face two challenges: first, the gap in movement amplitude between "inner speech" and "overt articulation" is large, making signals weak; second, oral anatomy varies significantly across individuals, so the same sentence produces inconsistent sensing patterns. As a result, silent speech systems often require individual calibration, and calibration cost is an important metric of practicality.

The real difficulty isn't the signal—it's generalization

The team found that cross-user generalization was a thornier problem than hardware. Algorithmic workarounds quickly hit their limits; the only effective lever was massively scaling up pretraining data. They set up collection stations in the office and paid people to wear the glasses and read prompts aloud, using this data to pretrain the model.

The result was a beautiful scaling curve: as data volume increased, error rates dropped rapidly and cross-user performance became genuinely usable. This conclusion aligns with the general experience in today's foundation model field—for modalities like sensor data, data scale often solves problems better than clever modeling.

The authors' reflection on "first principles" is also valuable: they initially tried to solve the entire problem top-down and failed completely; only when they broke down every component—antennas, electronics, and machine learning—to the most basic level and optimized each one did the whole thing finally work.

Cross-user generalization refers to a model's ability to maintain performance on users it has never seen during training. In sensor-based tasks this is especially difficult, because each person's physiology, wearing style, and movement habits systematically alter the signal distribution, creating pronounced "domain shift." Common countermeasures include individual calibration, domain adaptation, and adversarial training, but results are often limited. The experience mentioned in the article—that scaling pretraining data is more effective than clever modeling—is consistent with the scaling law in the foundation model field: when data covers enough people and scenarios, the model can learn representations shared across individuals, reducing reliance on calibration for new users.

Product timeline and who needs it

Subvocal is currently at the prototype stage. The team plans to continue R&D to refine the prototype into the finished form shown on the website, with a target ship date of Q1 2027, while continuing to expand the dataset to improve performance in edge cases.

The first batch of 500 units is already available for pre-order, with a $200 discount off the final price. Judging by the post's popularity on Hacker News—262 points and 184 comments—the need to "use voice input without being overheard" clearly exists; power users of tools like Wispr Flow are exactly their target audience.

From a technical standpoint, RF silent speech's advantage is being completely contactless—no skin-contact electrodes needed—making it easier to turn into a daily wearable. The challenges are equally obvious: RF signals' sensitivity to head posture and environment, consistency over long-term wear, and adaptation to different languages and accents all still need data to validate. If this route works, its significance isn't just adding another input method—it's liberating voice interaction from "private spaces" to "any space."

One-line Summary

Georgia Tech dropouts use RF smart glasses to enable closed-mouth silent speech input, targeting a 2027 ship date.

Paragraph Summary

Subvocal is a pair of smart glasses with embedded RF antennas that emit radio waves through the face; the waves are modulated by silent tongue and lip movements and decoded into text—no skin-contact electrodes needed. The team first replicated MIT AlterEgo's EMG approach at a hackathon, then pivoted to RF because of sweat interference and the impracticality of face electrodes, going through repeated antenna-layout failures before arriving at a design that fits inside the frame. The real bottleneck was cross-user generalization, ultimately solved by large-scale paid data collection for pretraining; the scaling curve shows rapid error reduction. In the demo, only about 20 sentences of calibration are needed. The first 500 units are available for pre-order, with shipping planned for Q1 2027.