Subvocal launches RF-enabled silent‑speech smart glasses, opens preorders for 500 units

Aarush and Arnav, founders of The Subvocal Company, a YC F26 startup, are developing smart glasses that allow users to communicate with their computers without audible speech. The device functions by decoding silent mouthing movements into text at the speed of speech. The founders, who were studying computer science at Georgia Tech at the beginning of the year, initially became interested in the concept after reading about AlterEgo, a silent speech system from MIT’s Media Lab. While they had no prior experience in silent speech or sensing, they were power users of Wispr Flow and sought a solution for using dictation tools in quiet environments such as libraries, offices, or public spaces.

During Stanford’s TreeHacks in February, the pair attempted to recreate a silent speech system within 36 hours. Their initial prototype used EMG electrodes attached to the face to read muscle electrical signals, which achieved limited success at the letter level. However, they determined that EMG was not viable for a product due to signal changes from sweat, electrical interference, and placement variability. Consequently, they pivoted to using radio frequency (RF) sensing, where radio waves similar to Bluetooth pass through the face and are modulated by the tongue, lips, and soft palate during silent mouthing.

The engineering process involved significant challenges in building a low-noise RF sensing stack and positioning antennas to maximize field overlap with facial articulators. Early attempts to integrate antennas into glasses failed, leading the team to temporarily tape antennas to their faces using a custom head rig to maintain consistent placement. After validating the signal through these makeshift methods, the team used simulations to design a final product with antennas embedded within the frames of the smart glasses.

A major technical hurdle was achieving cross-user generalization, as individual facial anatomy creates vastly different RF signal patterns. The founders concluded that algorithmic workarounds were insufficient and instead scaled their pretraining data by setting up booths in their office where paid participants read prompts while wearing the glasses. This approach produced a significant reduction in errors and improved cross-user performance, allowing for a short 20-sentence calibration period for new users. The team noted that breaking down components such as antennas, electronics, and machine learning to their fundamentals was essential to solving these engineering problems.

The Subvocal Company plans to continue research and development to transform the prototype into a refined product, with a target shipping date in Q1 2027. They are also expanding their dataset to improve performance on edge cases. Preorders are currently open for the first batch of 500 units, which includes a $200 discount off the final price. The company has also clarified that users do not need to move their lips or open their mouth, as the system can be tuned to detect "internal speech" or closed-lip mouthing.