Open-source voice dictation · 100% on-device
Talk. It types.
You edit nothing.
Hold one key, speak, release. Quobi transcribes and rewrites the mess: fillers gone, punctuation fixed, “at gmail dot com” turned into a real address. Then it drops clean text wherever your cursor is. All of it runs on your machine.
Free · Apache-2.0 · your audio never leaves the machine.
You said
um, so like can you send the report to uh, jordan at gmail dot com by friday, i think?
Quobi typed
How it works
Hold, talk, let go. That’s the whole interface.
Hold the key
Hold your hotkey (the ~ key, just left of 1) in any app or text field. Quobi listens for as long as you hold it.
Just talk
Say it however it comes out. Ramble, restart, spell an email aloud, say “new paragraph.” On-device Whisper catches every word.
Let go
Release the key. A local model trims the fillers, fixes punctuation, and types the clean result right where your cursor was.
Privacy by architecture
Your voice never
leaves the room.
There’s no cloud to leave. Speech-to-text and the Quill cleanup model both run locally, on your own CPU or GPU. Quobi works on a plane, behind any firewall, with the network unplugged. We can’t read your audio or your text, because they never touch a server. We don’t run servers.
- No account
- Install and go. There’s nothing to sign up for, ever.
- No tracking
- Nothing you say or type is logged, counted, or sent home.
- Never uploaded
- Your voice becomes text right on your device, then it’s gone.
- Works offline
- Turn off your Wi-Fi and it still works. The internet is optional.
The details
Small touches that make it feel effortless.
“Scratch that.”
Said the wrong thing? Just say “scratch that” and the last paste disappears. Your hands never leave what they were doing.
Three editing styles
Verbatim leaves your words alone. Tidy fixes grammar and fillers. Formatted adds paragraphs and lists. You pick how much it touches.
Types into any app
Quobi inserts text at the cursor through the OS, so it works in your editor, browser, terminal, Slack, or a text field on a website. If you can type there, you can talk there.
Knows spoken shorthand
“at gmail dot com,” “new paragraph,” “open paren.” It understands how people actually dictate and writes the real characters.
Fast, because it’s local
The compact Quill model cleans a sentence in ~80 ms on a GPU. No spinner, no datacenter round-trip, no rate limits.
Same on every platform
Linux, Windows, and Android run one shared cleanup engine and one prompt, so the output is identical wherever you dictate.
Open source
Open source,
top to bottom.
The app, the cleanup model, and the training recipes are all public under Apache-2.0. Privacy claims should be verifiable, not trusted, so read how it works, run it yourself, or fork it.
Get Quobi
Download and start talking.
free · no account · Apache-2.0
Android
On-device · in the works
Coming soonPrefer to build it yourself? Clone the repo. Everything you need is in the open.