spiral@out:~/blog$ cat transmissions/handy.md

Handy: Dictation That Never Leaves Your Machine

[2026-11-01] ::speech-to-text, privacy, accessibility, desktop, tools

Handy is a free, open-source speech-to-text app for the desktop that works completely offline. Press a shortcut, talk, and your words appear in whatever text field you're in. The README's pitch is short: "Your voice stays on your computer."

It's on GitHub Trending today with about 36,400 stars as of 2026-11-01, and it's licensed under MIT.

What it does

Handy does one job, which is how its own site describes it: "Transcribe what you say and put it into a text box." No account, no subscription, no meeting notes or summaries. The project's case for existing is that dictation is accessibility tooling and "belongs in everyone's hands, not behind a paywall." Most good dictation apps charge monthly or send your audio to someone's server, and Handy does neither.

The flow has four steps:

  1. Press your shortcut. Hold it to record and release to stop, or tap it once to start and again to stop. Hold-only and toggle-only modes are there if you want one or the other.
  2. Speak.
  3. Release, and Handy transcribes the recording.
  4. Get the text pasted straight into the app you were using.

How it works

Everything happens locally. Silero's voice activity detection trims out the silence first. Then one of two families of models does the transcription:

  • Whisper, OpenAI's speech model, in Small, Medium, Turbo and Large sizes, with GPU acceleration when you have a GPU.
  • Parakeet V3, a CPU-optimised model with automatic language detection, which makes it the obvious choice for laptops without a decent GPU.

Handy downloads models on first use. If you're behind a strict firewall, the README explains how to fetch them by hand and drop them in the models folder. You can also use your own: any Whisper model in GGML .bin format placed there shows up under Custom Models after a restart, and Handy picks up models already in your Hugging Face cache too.

The app itself is built with Tauri, so it's mostly Rust with a web front end, and it runs on Windows, macOS and Linux. It also has a small command line for controlling a running copy, which matters more than it sounds. On Linux, especially under Wayland, apps can't always grab global hotkeys, so you bind a key in your desktop or window manager to run handy --toggle-transcription instead. The README gives ready-made snippets for GNOME, KDE, Sway, i3 and Hyprland. There's a Raycast extension on the Mac as well.

Trying it

Download a build from the releases page or handy.computer. There are also community-maintained packages, which the Handy developers don't maintain:

brew install --cask handy        # macOS
winget install cjpais.Handy      # Windows

On Debian or Ubuntu, install the .deb with APT rather than dpkg so its dependencies come along:

sudo apt install ./Handy_*.deb

Launch it and grant microphone and accessibility permissions; it needs the accessibility one to type into other apps. Then set a shortcut in Settings and start talking. On Linux, install a typing helper for your display server: xdotool on X11, wtype or dotool on Wayland. To start it quietly at login:

handy --start-hidden --no-tray

The fine print

  • Linux is the rough platform. Wayland support is described as limited. You need the right typing helper, and on Ubuntu 26.04 wtype doesn't work, so you'll need ydotool plus some systemd setup. The recording overlay is off by default because some compositors let it steal focus and paste into the wrong window.
  • Clipboard juggling. Handy pastes through the clipboard and then restores what was there. Under load, the receiving app can read the clipboard too late and paste your old content. There are delay settings and an experimental "Reliable Paste" in the debug menu (Ctrl+Shift+D on Windows and Linux, Cmd+Shift+D on macOS).
  • Mac quirks. Shortcuts using the fn/Globe key only work on Apple keyboards. A Bluetooth headset mic can make playback sound worse while you record, so use the built-in mic instead.
  • Your hardware sets the pace. Bigger Whisper models are more accurate but slower without a GPU. Parakeet is the CPU-friendly option.
  • The code is open, the name isn't. It's MIT licensed, but the Handy name and logo are not. Forks have to use their own branding.

Who it's for

Handy suits anyone who'd rather talk than type: people with RSI or other reasons to keep off the keyboard, writers drafting long emails, developers dictating prompts to their AI tools, and anyone uneasy about streaming their voice to a cloud service. On a Mac or Windows box it's close to install-and-go. On Linux, budget ten minutes for the hotkey and typing-helper setup.

The GitHub repo and handy.computer are the places to start, and it's in the resource index now.

End of transmission.

<- BACK TO LOG