Sotto

Free, private, local dictation for Windows & Linux — hold a hotkey, speak, and accurate text is typed wherever your cursor is.

A fully offline, open-source alternative to cloud dictation tools like Wispr Flow. No cloud. No account. No subscription. No telemetry. Your voice never leaves your PC.

   

   

Live dictation in English and Hindi — text appears at your cursor as you speak.

Paid dictation apps are fast and accurate, but they send your microphone to the cloud, cost a monthly fee, and stop working offline. Sotto gives you the same press-to-talk, type-anywhere workflow while keeping 100% of the audio on your own machine — for free, forever.

• 🎙️ Types into any app — terminals, Word, Excel, browsers, chat, IDEs, even elevated windows. Uses simulated keystrokes (works where paste is blocked) with a clipboard-paste fallback that saves and restores your clipboard.
• 🌍 English + Hindi + Hinglish — auto-detected per phrase, so you can dictate a mixed-language meeting. Hindi that Whisper mistakes for Urdu is auto-corrected back to proper left-to-right Devanagari. One-click language switch from the tray.
• ⚡ Live typing — in toggle mode, text streams onto the page as you speak — great for hands-free meeting notes.
• 🧠 Robust in the real world — a voice-activity gate ignores static/background noise so it never invents phantom text when you're not speaking.
• 🔒 Private & free — OpenAI Whisper running locally on Intel OpenVINO (int8, CPU). The only time it touches the network is a one-time model download you trigger yourself.
• 🪶 Runs on a normal laptop — no GPU required.

1. ⬇️ Download the latest release (Sotto-windows-x64.zip) and unzip it.
2. Run Sotto\Sotto.exe. Windows SmartScreen may warn about an unsigned app — click More info → Run anyway. (Optional: run install.ps1 for Start-menu shortcuts.)
3. On first launch, open Settings → Model & language → Download (one-time, ~300 MB, from the official OpenVINO mirror).
4. Hold Ctrl + Win, speak, release. Your words appear at the cursor. 🎉

Prefer to build it yourself? See Build from source. It's ~400 lines of readable Python.

Sotto runs natively on Linux (X11 and Wayland) with the same press-to-talk workflow.

git clone https://github.com/smirk-dev/sotto && cd sotto/packaging/arch
makepkg -si                       # or: publish to the AUR and `yay -S sotto`
sudo usermod -aG input "$USER"    # one-time: lets the hotkey read the keyboard
# log out / back in, then launch "Sotto" from your app menu

The global hotkey is read from /dev/input via evdev (hence the input group), and typing is delegated to xdotool (X11) or wtype (Wayland) so Unicode — including Hindi/Devanagari — is emitted correctly. All dependencies are pulled in by the package. See packaging/arch/ for details.

Good to know on Linux:

• Prefer the hold chord (default Ctrl+Super) or a function-key toggle (F10) — the toggle combo's trigger key isn't suppressed from the focused app.
• GNOME-Wayland lacks the virtual-keyboard protocol wtype needs: use an X11 session or install ydotool as a fallback.

Hold-to-talk — hold Ctrl + Win (either side), speak, release. A pill with a live waveform shows at the bottom of the screen and never steals focus. Text lands at your cursor, usually in under a second.

Toggle / live mode — press Ctrl + Alt + D to start, again to stop. Text is typed as you speak, so you can dictate for minutes or take meeting notes hands-free.

Spoken commands: “new line”, “new paragraph”, “comma”, “period”/“full stop”, “question mark”, “exclamation mark”, “colon”, “semicolon”, “open/close quote”. Emails and URLs are formatted automatically (“name at gmail dot com” → name@gmail.com), and filler words (um, uh) are removed.

The default is a multilingual model set to “Auto — English + Hindi.” Switch any time from the tray → Language menu:

• Auto — English + Hindi (default) — detects each phrase; best for mixed/Hinglish sessions.
• English only / Hindi only — force one language for the most reliable single-language results.
• Auto — all languages — unconstrained 99-language detection.
• For the fastest pure-English dictation, switch Model to base.en (sub-second).

Tip: if accuracy disappoints, pick a single language instead of Auto, keep the mic close, and reduce background noise. Sotto runs a compact model on your CPU, so it trades a little accuracy for being fully private and free.

• Hotkeys — configurable hold chord (Ctrl+Win, Alt+Win, Ctrl+Alt, F9) and toggle combo (Ctrl+Alt+D, Ctrl+Shift+Space, …).
• Microphone picker with a live level meter.
• Custom dictionary — teach it names and jargon; they're fuzzy-corrected in the output.
• Dictation history — searchable local log, one-click re-copy; disable or clear anytime.
• Session stats — words dictated and estimated time saved vs. typing.
• Start with Windows, single-instance, subtle start/stop sounds, dark minimal UI.

Settings → Model & language. Downloaded models show a ✓; others show a Download button (one-time, from the official OpenVINO mirrors on Hugging Face).

Data lives in %LOCALAPPDATA%\Sotto (models, config, history, log).

Run on defaults to Auto, which picks the device that is actually faster for the model you chose. That is the CPU for everything except large-v3-turbo, where an Intel iGPU takes it from ~4.6 s to ~1.8 s from hotkey release to text (i7-1360P + Iris Xe, measured through the app), with identical output. The reason is architectural: turbo is ~79% encoder, and Whisper's encoder is one big parallel pass over a fixed 30 s window — dense, parallel work an iGPU eats. Smaller models spend their time in the memory-bound decoder instead, where the iGPU is a slight loss, so Auto leaves them on the CPU. Pick CPU or GPU explicitly if you would rather not have it decide; if the GPU can't run the model, Sotto falls back to the CPU by itself. The first GPU run compiles the model (~15 s, one time) and caches ~900 MB under %LOCALAPPDATA%\Sotto\models\ov-cache.

Is my audio uploaded anywhere? No. Transcription runs entirely on your CPU. The only network request is the one-time model download you start yourself.

Does it need a GPU? No — every model runs on the CPU, and the default one is tuned for it. An Intel iGPU is only worth it for large-v3-turbo, where it cuts ~4.6 s per utterance to ~1.8 s; Auto uses it for that model if you have one, and nothing changes for anyone who doesn't.

The hotkey does nothing — or text isn't typed — in one specific app (often a terminal). That app is running as Administrator while Sotto isn't. Windows UIPI blocks a normal-privilege app from both typing into and seeing hotkeys destined for a higher-privilege window, so the overlay won't appear while it's focused and dictation lands nowhere (Sotto falls back to copying the text to your clipboard and shows "run Sotto as admin"). Everything at your normal level — Word, browsers, chat, Obsidian — keeps working. Fix: run Sotto elevated — see Dictating into apps that run as Administrator.

Windows says "unknown publisher." The app is unsigned (code-signing certificates cost money; Sotto is free). Click More info → Run anyway, or build it yourself.

Can it do languages other than English/Hindi? Yes — set Language to Auto — all languages or pick a specific one; Whisper supports ~99 languages.

If an app runs elevated (as Administrator) and Sotto doesn't, Windows UIPI stops Sotto from interacting with it: injected keystrokes are silently dropped and the global hotkey isn't seen while that window is focused (the overlay only reappears once you click a normal-privilege window). The most common case is an elevated terminal — e.g. Windows Terminal, or a VS Code integrated terminal when VS Code itself was started as Administrator. When it happens, Sotto shows "run Sotto as admin — text copied" and puts the transcript on your clipboard so nothing is lost.

Two fixes:

• Run Sotto elevated. Launch the "Sotto (administrator)" Start-menu shortcut (one UAC prompt per launch) so Sotto sits at the same integrity level and can type everywhere. To start it elevated automatically at logon without a UAC prompt, register a Task Scheduler task that runs it with highest privileges:

  $exe  = "$env:LOCALAPPDATA\Programs\Sotto\Sotto.exe"
  $user = "$env:USERDOMAIN\$env:USERNAME"
  Register-ScheduledTask -TaskName "Sotto (admin autostart)" -Force `
    -Action    (New-ScheduledTaskAction -Execute $exe) `
    -Trigger   (New-ScheduledTaskTrigger -AtLogOn -User $user) `
    -Principal (New-ScheduledTaskPrincipal -UserId $user -LogonType Interactive -RunLevel Highest) `
    -Settings  (New-ScheduledTaskSettingsSet -StartWhenAvailable -ExecutionTimeLimit ([TimeSpan]::Zero))

  If you do this, turn off "Start Sotto when Windows starts" in Settings (or delete …\Startup\Sotto.lnk) so the non-elevated copy doesn't race the task at logon.

• Or just don't run the other app as Administrator. If you don't specifically need it elevated, launch it normally and Sotto dictates into it like anywhere else — no elevated Sotto required.

git clone https://github.com/smirk-dev/sotto
cd sotto
py -3.13 -m venv .venv
.\.venv\Scripts\python -m pip install faster-whisper sounddevice PySide6-Essentials numpy pyinstaller pillow openvino-genai huggingface_hub
.\.venv\Scripts\python -m PyInstaller sotto.spec --noconfirm   # -> dist\Sotto\Sotto.exe

Run the app in dev with .\.venv\Scripts\python -m sotto. Tests are in tests/; PROGRESS.md documents the engineering decisions and on-machine benchmarks.

Issues, ideas, and PRs are very welcome — see CONTRIBUTING.md. If Sotto is useful to you, a ⭐ star helps others find it.

MIT for Sotto's own code. Bundled components keep their own licenses — see THIRD-PARTY-NOTICES.md.