Silent speech interface · In development

Speak without making a sound.

Unspokn translates silent lip movements into text or natural-sounding speech — so you can be heard without using your voice.

Join the waitlist Hear the voice
02 The problem

Millions of people face barriers to spoken communication.

Speech disabilities. Temporary voice loss. Environments where speaking aloud isn't possible or desirable.

Existing speech-to-text tools require audible speech. Many assistive alternatives rely on typing, selecting predefined phrases, or specialized hardware — slow, conspicuous, and often unusable in an everyday conversation.

  • Speech-to-textrequires your voice
  • Phraseboardsrequires predicting what you'll say
  • Dedicated hardwarerequires having it with you
  • Unspoknrequires a camera
03 The voice

This is the voice we give back to you.

Confirmed words are spoken with Deepgram Aura. Choose a phrase or write your own, and hear it out loud.

Real Deepgram Aura speech, generated live. Lip-reading capture runs in the Unspokn app — not on this page.

Ready

Built with

  • Deepgram Aura
  • TypeSafe Jev
  • Cloudflare
  • Open-source VSR
05 How it works

A camera-first alternative, in four steps.

  1. 01

    Camera on. Microphone off.

    Separate, explicit permissions. Both states stay visible on screen, so you always know what's listening — and that nothing is.

  2. 02

    Your mouth movement is captured.

    Face detection crops the mouth region and segments an utterance from movement, or from a press-to-capture button when you'd rather be deliberate.

  3. 03

    You approve the text.

    Open-source visual speech recognition proposes a transcript. It is editable, and nothing is ever spoken before you confirm it.

  4. 04

    Your words, out loud.

    Confirmed text becomes natural speech through Deepgram Aura — or stays as text you can display, share, or send.

Inside step 03 · the confidence gate TypeSafe Jev

Before anything is spoken, the candidate transcript is scored. Low confidence never gets a voice — it gets a question. Try it: write what the model thinks it saw.

Not run yet.

06 Why it matters

Be heard without your voice.

Mouth the words you would have said. Unspokn carries them the rest of the way — as text on screen, or as speech other people can hear.

Nothing speaks until you say so.

Every transcript is editable and requires confirmation. Low-confidence guesses are flagged, never voiced on your behalf.

No special hardware.

A camera you already own. Built for web and mobile, so it's wherever the conversation is.

Privacy as a default, not a setting.

The microphone stays off in silent mode. Processing stays on-device where feasible, remote processing is disclosed, and raw video isn't retained by default.

07 Who it's for

Another way to express yourself
— without requiring you to speak aloud.

  • Speech disabilities

    Motor-speech conditions, ALS, apraxia, dysarthria, and other conditions that make reliable audible speech difficult.

  • Temporary voice loss

    Surgery and recovery, laryngitis, vocal strain — the weeks when your voice isn't available to you.

  • Situational silence

    A quiet carriage. A shared office. A room where someone is sleeping. A factory floor too loud to be heard in anyway.

08 What it can't do yet

We'd rather tell you the limits than oversell the promise.

This is an accessibility product in development. Overstating it would cost the people who need it most.

  • Lips are ambiguous.

    Many sounds look identical on the mouth. Open-vocabulary silent speech is not dependable for every user or every phrase.

  • Benchmarks aren't guarantees.

    Published visual-only error rates come from research datasets recorded under good conditions. They are not a promise about your room.

  • Conditions change everything.

    Face angle, lighting, occlusion, frame rate, camera quality, accent and individual articulation all move results, sometimes a lot.

  • It may not work for you yet.

    Models trained on typical speech can perform poorly for people with atypical articulation or limited mouth movement — exactly the people we most want to serve. We're testing with intended users, not around them.

09 Standing on

Unspokn is built on open visual speech recognition research.

We don't have testimonials yet — we have citations. Here's the work this is built on.

10 Questions
Does Unspokn listen to me?

No. In silent mode the microphone is off, and the interface shows you that it's off. Camera and microphone are separate permissions you control independently.

Do I need special hardware?

No. A standard camera on a phone or laptop. Unspokn is being built for web and mobile.

Is my video uploaded somewhere?

Processing happens on-device where feasible. Where remote processing is required, it's disclosed in the app, and raw video isn't retained by default.

Can I use it today?

Not yet — it's in active development. Join the waitlist and we'll tell you when there's something real to try, not before.

What languages does it support?

English first. Multilingual visual speech recognition models exist and are on the roadmap, but each language needs its own evaluation before we'd claim it works.

What if lip reading doesn't work for me?

Then you still need to communicate, so Unspokn keeps a no-camera path: typing, saved phrases, and one-tap responses. The camera is an option, not a requirement.

Your words.
Your way.

We're building Unspokn with privacy, accessibility and user control at its core. Be the first to know when it's ready to try.