The Subvocal Company, founded by Aarush and Arnav, is developing a wearable system designed to solve a common friction point for voice-to-text users: the inability to use dictation in quiet or public spaces like libraries and offices. While traditional voice dictation requires audible speech, their technology aims to capture 'subvocal' input. This allows users to communicate with their computers by mouthing words silently or even by making micro-movements of the tongue and soft palate without opening their mouths. The goal is to maintain the natural flow of speech—which typically occurs at 120-200 words per minute (WPM)—while bypassing the physical friction of typing, which for many users averages between 40 and 100 WPM.
The technical journey to this product involved a significant pivot in sensing methods. The founders initially experimented with electromyography (EMG), which reads electrical signals from facial muscles. However, they found EMG impractical for a consumer product due to signal degradation from sweat, electrical interference, and the fact that users generally do not want to wear electrodes on their faces. They transitioned to a radio frequency (RF) approach, where the glasses emit waves similar to Bluetooth. As these waves pass through the face, they are modulated by the movement of facial articulators (the tongue, lips, and soft palate). The system uses foundation models trained on massive amounts of data to decode these perturbations into text. To achieve this, the team had to engineer a low-noise RF sensing stack and embed antennas directly into the glasses' frames to ensure maximal overlap with the mouth's internal structures.
A primary engineering challenge for the system is cross-user generalization, as every individual's facial anatomy produces a unique RF signal pattern. The founders addressed this by scaling their pretraining data, using a large dataset of people reading prompts while wearing the glasses to create a 'scaling curve' where error rates dropped significantly as the dataset grew. Despite this progress, the technology faces practical and social hurdles. Users have noted the potential 'weirdness' of mouthing words in public, and the reliance on RF waves has prompted questions regarding the possibility of the technology being adapted for espionage or remote sensing. The company plans to ship its first batch of units in early 2027, with the current prototype requiring a short calibration period to adapt to a specific user's anatomy.
While the founders report a successful scaling curve and compelling cross-user performance, the specific accuracy rates and the architecture of the foundation models remain internal proprietary details. Furthermore, the transition from a prototype to a polished consumer product involves significant hurdles in maintaining antenna alignment and minimizing environmental RF noise, which could affect reliability in different real-world settings.