Svara v0.1.0 — first public release

Svara v0.1.0 — first public release

Private voice dictation that runs entirely on your own machine. Double-tap
Right Alt, speak, and your words stream to the cursor in any app. Nothing is
uploaded — transcription runs locally with faster-whisper.

Download

Svara.exe — one 108 MB file. Download and run.

It auto-detects your hardware:

• NVIDIA GPU? On first launch it downloads GPU support once (~1.3 GB, with a
  progress bar) and runs the best model (large-v3-turbo) on your GPU at ~1s,
  live-streaming. You don't pick anything — it just uses the GPU.
• No GPU? It stays on the CPU with a fast, CPU-sized model. No big download.

Prefer not to download at runtime (offline / locked-down machine)? Grab
Svara-Windows-GPU.zip — the same app with the GPU runtime already inside.

Use it

1. Run Svara.exe. A quick setup appears — confirm the model and read the
   3-step how-to. (On a GPU, first launch fetches GPU support + the model once.)
2. Double-tap Right Alt, speak, and words stream to the cursor live.
3. Tap Right Alt once to finish. The mic icon lives in your tray.

Everything after the one-time download runs fully offline.

Windows SmartScreen may warn (the app isn't code-signed yet). Click
More info → Run anyway. It's open source — read every line in the repo.

Features

• Runs on your machine — no server, no account, no telemetry. Works offline.
• Auto-GPU — uses your NVIDIA GPU if you have one, CPU if you don't. No config.
• 90+ languages — dictate in any language, auto-detect, or speak-to-translate into English.
• Shout to CAPITALISE — raise your voice and it lands IN CAPS.
• Gets out of your way — the pill reads the focused app's text caret and slides aside.
• Eight visualizers + pop-culture themes — Matrix, Cyberpunk, Sakura, Evangelion, Saiyan, Vaporwave, minimal.

Full docs: see the README. Licensed AGPL-3.0.