Silent speech isn't necessarily private: what Subvocal's cloud processing means

Subvocal's pitch is privacy you can hear, or rather, not hear. The startup's glasses are meant to let you dictate a message on a crowded train by mouthing the words, so nobody next to you knows what you said. Its own homepage uses the example sentence "I think it's time I looked for another job."

But a silent input is not the same thing as a private one. According to Subvocal's co-founder, the words you mouth are decoded in the cloud, through a companion app on your laptop or phone, and the product requires internet access. The person beside you cannot hear you. A server can read you.

That does not make the product unsafe. It means the privacy question moves from the train carriage to the data center, and as of November 9, 2026, Subvocal has not published any document that explains how that speech data will be handled. Here is what we could verify, what is only a statement from the founders, and what is still unknown.

What the prototype does

Subvocal is a three-person San Francisco company in Y Combinator's Fall 2026 batch, according to its YC profile. It is building glasses that turn silently mouthed words into text you can send to an AI assistant, a coding tool, email or chat. The company says the glasses do not use cameras and do not read lips from video.

Instead, the frames send low-power radio waves through the face and measure how they change as the lips, tongue and soft palate move, the company explains in its YC launch post. Machine-learning models then turn those changes into words. The same post describes the demonstrated hardware as "semi-portable" and rougher than the planned product, and the preorder page says the images show a design direction, with final specifications still in development.

The company also says the glasses cannot read thoughts: they decode only words you intentionally articulate, and they are not always listening. In the Hacker News launch thread, the co-founder added that the radio transmitter is active only while you dictate.

Where your silent words go

Asked how decoding works, co-founder Aarush Yadav wrote on Hacker News that "the decoding is done via a companion app on your laptop or phone in the cloud," with end-to-end latency "under 500ms usually" (comment, November 2, 2026). In a separate reply, he confirmed the product requires internet access.

So the data path, as the founder describes it, runs from your face to the glasses, to an app on your device, over the internet to Subvocal's servers, and back as text. The practical consequence is simple: no connection, no dictation. A commenter in the same thread pointed out that this rules out places where signal drops, such as many subway tunnels, which are exactly the crowded, quiet settings the product is marketed for.

Why the cloud? The company has not explained its reasoning in a document, but open-vocabulary decoding usually depends on large models that are hard to run on small, battery-powered hardware. The founder has said local processing is a hope, not a plan with a date: "At some point in the future we hope to launch local inference models too" (comment, November 3, 2026).

What Subvocal says about protecting that data

The founders have made three specific privacy statements, all in Hacker News replies rather than in a policy:

• Speech data will go to "cloud inference, but will be processed with encryption in hardware isolation" (source). That wording points to confidential computing, where data is processed inside a protected hardware enclave, but the company has not named the technology, the provider or any audit.
• Data will be "only retained and used for anonymized training with your consent" (source).
• The company has built safeguards "to prevent even us from accessing the raw data in the cloud" (source).

These are meaningful commitments if they are delivered. They are also, today, statements in a discussion forum. They are not in a privacy notice, terms of service, a security white paper or an independent audit.

What the privacy policy actually covers

Subvocal's Privacy Policy (effective October 30, 2026) is explicit about its limits. It applies to the website, the preorder and referral program, and related communications. It says a future product, app or service "that processes voice, sensor, device, or other product data will receive a separate or updated privacy notice before that processing begins."

In other words, the policy you accept when you reserve a pair says nothing about how your silent speech will be handled. It covers your name, contact and billing details, payment identifiers from Stripe, and website analytics. It lists the service providers used for that preorder data, including Stripe, MongoDB, Netlify, Vercel, Cloudflare and PostHog, and says the company does not sell personal information. None of those providers is described as handling speech or sensor data, because that processing has not started.

That is a reasonable stage for a company that has not shipped a product. It does mean buyers are reserving hardware before they can read the rules for its most sensitive data.

Update: a new addendum, about website visitors

On November 6, 2026, Subvocal added an addendum to the same policy. It says the company now uses a service called Letter to "identify the companies and business contacts who visit" its website, for sales and marketing, by matching IP addresses, cookies and pages viewed to professional profiles. The policy also lists "internal investor alerts in Slack" among the uses. According to the addendum, visitors in the EU, UK and Switzerland are identified only if they accept cookies, and anyone can opt out through Letter's opt-out page or the Global Privacy Control browser setting.

This addendum covers visits to the website, not speech data from the glasses. It still matters for buyers weighing trust: it shows the company is willing to identify people who browse its preorder page, and it was added within a week of launch. The promised notice for speech and sensor data had not appeared as of November 9.

Why this matters more for silent speech

People choose silent input precisely when the words are sensitive: a message to a doctor, a note about a colleague, a job search, a prompt containing company code. Subvocal markets the product for offices, cafes and transit for this reason. The more private the use case, the more the cloud side matters.

Cloud processing is not automatically a problem. Most voice assistants and dictation tools send audio to servers, and well-run services can protect it. The difference with any new device is that trust has to be earned in writing. The questions buyers should expect Subvocal's product privacy notice to answer:

• Is the uploaded data the raw radio signal, the decoded text, or both?
• Is anything stored by default, and for how long?
• Is training on your data off by default, and can you withdraw consent later?
• In which countries are the servers, and which cloud providers process the data?
• Can staff or contractors access recordings or transcripts under any circumstances?
• How does the company respond to law enforcement requests?
• Will an independent security audit be published before shipping?

Claims versus evidence on performance

Subvocal says the glasses decode open-ended silent speech at more than 200 words per minute. Its FAQ is more careful about accuracy, saying the system "can occasionally get something wrong," and its YC post describes a "reasonably low error rate" at small dataset scale. The founders said the demo user had completed a short 20-sentence calibration.

The only error figure so far came in a Hacker News reply on November 5. The co-founder put the current error rate at "8-15% on an unseen user," meaning someone the system was not trained on, and said it "will likely be 3-5%" by the time units ship (source). Those are the company's own numbers, with no published test method. No independent tester has measured the device.

Language support is narrow for now. The co-founder wrote that the product "supports just English for now," with Spanish as the second planned language. No other languages have been announced.

What buyers can actually reserve

Sources: preorder page and Preorder Terms, version 2026-10-30, both checked November 9, 2026. Some news reports gave different batch sizes and ship dates; we rely on the company's current pages.

The $499 may not be the whole cost. None of the preorder documents mention a subscription. But the co-founder wrote on Hacker News that there will be "a free tier with a cap on the amount of words you can dictate" and "a paid tier for ~$20/mo with unlimited dictation" (source). Because decoding runs in the cloud, an ongoing fee is plausible. The word cap and final price have not been published.

Outside the US, there is no date. Asked about Europe, the co-founder replied "Not immediately, no," citing separate regulatory approval for radio devices there (source).

The bigger picture

Subvocal is not alone. MIT Media Lab's AlterEgo headset read silent speech from facial muscle signals in 2018, and its spin-out still shows only a sign-up form for a pre-production prototype. On January 29, 2026, Apple confirmed it had acquired Israeli startup Q.ai, which works on machine learning for whispered speech; Reuters reported a value of about $1.6 billion, citing a source, and a Q.ai patent application on detecting silent speech from facial skin micromovements. Apple has not said how it will use the technology. Our tracker follows these efforts by stage.

Whoever ships first at scale will face the same question Subvocal faces now: where the words go after they leave your mouth.

What to know before reserving

• It is a reservation, not a product. Nothing has shipped, and nobody independent has tested it.
• Your words are processed in the cloud, according to the founder, and the glasses need internet to work.
• The product privacy notice does not exist yet. Today's policy covers only your order and website data, and since November 6 it also allows identifying business visitors to the site.
• Expect a possible subscription. The founder described a capped free tier and about $20 a month for unlimited dictation, on top of the hardware.
• You can get your $199 back any time before final payment, so you can wait for that notice before paying the rest.
• US addresses only, English only at launch, shipping estimated Q1 2027 and subject to FCC approval. No date for Europe.

We will update this story when Subvocal publishes its product privacy notice, security details, independent test results or FCC authorization.