Could People Talk To AI Without Saying A Word? A Startup Is Making Glasses That Read Silent Speech.

A San Francisco startup is developing glasses that could detect the tiny facial movements people make when they silently form words and take them to a screen.

The Subvocal Company says its wearable can translate those movements into text in real time, allowing users to communicate with computers without making a sound. The startup has already opened preorders in the United States, with an early-bird price of $499, although the product is still in development.

Its founders claim their prototypes can process more than 200 words a minute, potentially offering a faster way to write and interact with AI tools. However, the technology's speed and accuracy have not been independently verified.

Unlike conventional voice assistants, which listen to spoken commands, Subvocal uses radio-frequency sensing to detect subtle changes in facial muscles as someone silently articulates words. Low-power radio waves pass through facial tissue, and the returning signals change as the muscles move, allowing AI models to interpret those signals and convert them into written text.

The glasses require users to deliberately form words, rather than simply think about what they want to say. According to the company, these movements can be subtle enough to go largely unnoticed by someone sitting nearby, making the technology potentially useful in offices, libraries, public transportation and other places where speaking aloud would be inconvenient.

Subvocal wants the glasses to function as a text-input method across existing applications, including email, messaging platforms, documents and AI assistants such as ChatGPT and Claude.

The company was founded in 2026 by Aarush Yadav and Arnav Kumar, who serve as chief executive and chief technology officer, respectively. Both are associated with Georgia Tech, according to their Y Combinator company profile.

In their Y Combinator launch announcement, the founders said they began developing the technology roughly four months earlier, starting with a prototype built at a hackathon. Their initial setup involved electronic components and antennas attached to their faces, but they have since developed a more portable version.

The hardware is still larger than the finished glasses they hope to sell, although the founders said their early experiments demonstrated that silent speech could be detected and decoded using a relatively small training dataset. Subvocal has since raised $1.3 million from investors including Y Combinator and Afore Capital, according to a public statement by Yadav announcing the product on October 30.

The startup's official preorder page lists a regular price of $699, reduced to $499 for early customers. Buyers can reserve a pair with a refundable $199 deposit, with the remaining $300 payable later. Shipping, taxes and optional prescription lenses will cost extra.

Deliveries are expected to begin during the first quarter of 2027, although the company has not confirmed an exact date. Shipping is also subject to approval from the Federal Communications Commission, and preorders are currently available only in the United States.

The glasses are entering a growing market for AI-powered wearables, where companies are experimenting with alternatives to keyboards and touchscreens. Meta's Ray-Ban smart glasses, for example, allow users to interact with an AI assistant through spoken commands, while several software companies have developed tools that turn speech into written text.

Subvocal is pursuing a different approach by offering the convenience of dictation without requiring users to speak or relying on camera-based lip-reading technology. But getting the system to work consistently outside controlled demonstrations remains a challenge.

In a Y Combinator job listing, the company acknowledged that a system working well for one person during a single session is not enough. Its AI models must recognize speech consistently when users return on different days, adjust the glasses or use the device for the first time.

To address these challenges, Subvocal plans to collect larger datasets and develop a calibration process that allows its technology to adapt to individual users. The company also says its system is designed to recognize intentional silent articulation rather than continuously recording everything a person does, although detailed independent assessments of its privacy protections have not yet been published.

As it prepares for its first production run, the startup is recruiting engineers specializing in machine learning, sensing and hardware development. Its current openings include a head of machine learning position in San Francisco offering a salary range of $150,000 to $200,000, along with equity, with the role focused on improving speech recognition across different users and changing conditions.