Your voice stays on your computer by default

We talk to our agents a lot. "Make the empty state friendlier, and the button on the left should match the one in settings" is faster to say than to type, and it's easier to say it while looking at the screen than while looking at the keyboard.

What you say while you work is often about client code, unreleased products and people. So we made a simple default: your voice is turned into text on your own computer. If you'd rather use a cloud model, that's your call, on your own key.

How it works#

Shuriken runs NVIDIA Parakeet TDT 0.6B v3, an open speech model, on your computer with the sherpa-onnx runtime. Press the mic, talk, and the words appear as you speak. With the local model the audio never goes to a speech service, to us, or to any other third party.

• Offline. Once the model is on your computer it needs no connection. Trains, planes and bad hotel Wi-Fi are fine.
• One download. The model comes down once, about 465 MB. You can see its size and remove it any time in Settings › Dictation.
• No key, no bill. There's nothing to sign up for and nothing to pay per minute. Voice is in Free and Pro alike.
• 25 languages. Parakeet v3 transcribes 25 European languages and works out which one you're speaking.

Your phone and the browser#

A phone or a browser can't run the model. When you dictate from the phone web app or a browser, your voice travels end-to-end encrypted to your own computer, which turns it into text and sends the words back. When that goes through our relay, the relay sees only sealed bytes it can't read. (More on what our cloud can and can't see.)

Or bring your own provider#

Some people already pay for a speech service, or want one of its languages. In Settings › Dictation you can pick ElevenLabs, AssemblyAI, Deepgram, Speechmatics or OpenAI instead, paste your own API key and press Test: Shuriken opens a short live session to check the provider takes it.

• Live only. Every provider streams: your words appear while you speak. Providers that only transcribe a finished recording aren't offered.
• Your account, your bill. Your computer streams the audio straight to that provider under your key. It never passes through a speech account of ours.
• Silence is never sent. A voice detector on your computer holds back the quiet between sentences, so the provider hears only speech.
• Your words. AssemblyAI, Deepgram, Speechmatics and OpenAI also get the words Shuriken has learned you use, so a project name stops coming out as something else.
• Your key stays encrypted. It's kept in your encrypted vault (our cloud keeps a key that opens it) or your computer's keychain, and it's never written to a log.

Switch back to Local model any time and nothing leaves your computer again.

Paste while you talk#

Our favourite trick works with every provider. Halfway through a sentence, paste an error message, a file path or a screenshot. The text lands exactly where you were in the sentence, not at the end, and images attach as you go. It's also what makes pointing at a page in the built-in browser and saying the change feel natural.

Everything else about voice is on the Voice page. To hear it for yourself, try Shuriken and press the mic.