runs offline on your machine
Talk to your machine.
Nothing leaves it.
Hold a key, say the thing, and the words land where your cursor is. The audio is transcribed on your own machine. No upload, no account, no internet.
macOS · English · always free · MIT licence
the bar, on the web
Watch the words settle before you stop talking.
Most dictation waits for silence and then thinks. Swar finishes each phrase as you pause, so the only thing left when you let go is the last one.
Your cursor sits here. Press play.
the trade every other one makes
Good dictation usually means sending your voice to someone else's computer.
That is the deal on offer nearly everywhere. Your microphone goes to a server, the text comes back, and you take it on faith that the recording was discarded. For a grocery list, fine. For a salary conversation, a medical note, an unannounced product, or anything under NDA, it is a real decision you have to make every time you hold the key.
Swar does not ask you to make it. The model sits on your disk. The audio never has anywhere to go.
how it works
Three steps, and one of them is just talking.
Hold the key
A bar appears above whatever you are working in. It never takes focus, so your cursor stays exactly where you left it.
Keep talking
Swar transcribes as you go. Each phrase is settled the moment you pause, rather than piling up until you stop.
Let go
Only the last phrase is still being decoded, so the wait is roughly the same whether you spoke for ten seconds or two minutes.
measured, not claimed
These come from our own benchmark corpus, and the speech in it is synthesised rather than recorded from people. It is a fair way to compare one build against the next. It is not yet a promise about how Swar handles your particular voice, room, or microphone.
what it does and does not do
A short list, honestly kept.
works today
- EnglishThe only language Swar writes.
- macOSApple silicon and Intel.
- Any appNative, browser, or Electron.
- Your vocabularyNames and jargon you add, spelled right.
- Fully offlineWorks on a plane, in a basement, on a locked-down laptop.
does not, yet
- Hindi and HinglishRemoved for now. It was not good enough to ship.
- Choosing how much it editsThe Raw, Clean, and Intent switch is there, but all three currently produce the same text.
- Windows and LinuxNot started.
- iPhone and AndroidNot started.
- Sync between devicesThere is no server, so there is nothing to sync through.
the part people ask about
There is no server.
Not a private one. Not an encrypted one. None. The model file is on your disk, the transcription runs in a process on your machine, and the recording is overwritten in memory as soon as the words are written out.
Turn off your wifi and use it exactly as before. That is the whole test, and it is the one we care about passing.
run it yourself
There is no download yet. Clone it and run it.
No signed installer exists so far, so the honest answer is a build from source. It is MIT licensed, which means you can read every line that touches your microphone before you run it, and change any of it.
what you need
- macOS
- 10.15 or newer, Intel or Apple silicon
- Flutter
- 3.35 or newer
- Rust
- 1.88 or newer
- Bridge codegen
- flutter_rust_bridge_codegen 2.12.0
git clone https://github.com/Priyammm05/swar.git
cd swar
./scripts/generate_bridge.shcd apps/swar_desktop
flutter run -d macosflutter build macos --release
cd ../.. && ./scripts/sign_local_macos.shSigning with a stable local identity is what stops macOS asking for Accessibility permission again after every build.
- Parakeet
- 639 MB
- The speech model. This is the one doing the work.
- Whisper small
- 180 MB
- Fallback, for when the fast helper is unavailable.
- Qwen 3B
- 1.96 GB
- Optional cleanup model. Fetched automatically today, which is more than most people want.
2.8 GB in total. macOS will ask for Microphone and Accessibility permission on first run. After that Swar never touches the network again.
Stop typing what you could just say.
macOS · English · always free · MIT licence