Speak.It's written.
Svara floats over any app and writes down what you say, the instant you say it, on your own GPU. Nothing ever leaves your machine.
stringswoven strings · one of eight ways Svara draws your voice
Three moves, and your voice is on the page.
Double-tap Right Alt
The pill appears and Svara starts listening. No window to find, no button to hunt for.
Right Altx2Just talk
Words are written as you speak, punctuation and filler cleaned up on the fly. Raise your voice and it writes in caps.
~1s spoken to writtenTap to finish
The text lands at your cursor in whatever app you are using. Slack, VS Code, a browser, a terminal.
Works everywhereEvery colour recolors everything.
Everything it does.
Runs on your own machine
Audio is captured, written, and discarded locally. It cannot reach a server, because there is no server. Nothing to upload, ever.
Instant on your GPU
faster-whisper on CTranslate2, large-v3-turbo at int8 in ~1.5 GB of VRAM. Roughly a second from spoken to written.
Ninety-plus languages
Dictate in any language Whisper understands, or let it auto-detect what you speak each time. Not just English.
Speak-to-translate
Flip one switch and talk in any language: Svara writes clean English at your cursor. Your Spanish, its English.
Any Whisper model
From tiny to distil-large-v3 to large-v3-turbo. Trade speed for accuracy to fit whatever GPU you have.
Shout to CAPITALISE
Raise your voice on a word and it lands IN CAPS. Loudness-aware and shout-proof, so only real emphasis counts.
Cleaned up as you talk
Automatic punctuation, filler removal (um, uh), and self-corrections on the fly. Optional local-LLM polish via Ollama, still offline.
Writes into every app
System-wide injection places your words at the cursor anywhere you can type: Slack, VS Code, browsers, terminals.
Eight meters, your theme
Eight live visualizers and pop-culture themes: Matrix, Cyberpunk, Sakura, Evangelion, Saiyan, Vaporwave, and clean minimal.
It moves out of your way
The pill reads the text caret of the app you are in and slides aside the instant it would cover what you are typing, easing back when the coast is clear. Works even in browsers and Electron apps.
Never fights your typing
A poll-only hotkey watches one key with no global keyboard hook, so it never adds input lag. Lives in the tray and restarts itself if it ever hiccups.
Free and open source
No account, no subscription, no telemetry on your voice. The whole thing is yours to read, fork, and build.
Speak it.
Ship it. Free.
One 110 MB download that just runs. No installer, no unzip, no GPU required. Works on any Windows 10 or 11 PC; the model downloads once, then it runs fully offline.
~110 MB · Windows 10/11 · download and run · have an NVIDIA GPU? get the GPU pack ↗