Handy: Free Offline Speech-to-Text That Types What You Say

Handy is a free, open-source, and extensible speech-to-text application that works completely offline. Press a shortcut, speak, and your words appear in any text field, in any app. The transcription happens on your own computer, without sending any audio to the cloud.

The project states its case in four words: free, open source, private, and simple. Accessibility tooling belongs in everyone's hands, not behind a paywall. Open source means anyone can extend it and contribute to something bigger. Private means your voice stays on your computer. Simple means one tool, one job: transcribe what you say and put it into a text box.

It is a cross-platform desktop app built with Tauri and Rust, licensed under MIT, and the repository had 36,127 stars and 3,428 forks when we checked on November 1, 2026. The project site is handy.computer, and releases are published on GitHub. Note one branding detail from the license file: Handy itself is open source, but the Handy name, logo, icon, and brand assets are not, so unofficial forks and redistributions must use their own branding.

Using Handy takes four steps. Press a configurable keyboard shortcut: hold it to record and release to stop, or tap it to toggle recording on and off. Hold-only and toggle-only modes are also available. Speak your words while the shortcut is active. Release, and Handy processes your speech. Then the transcribed text is pasted directly into whatever app you were using.

The process is entirely local. Silence is filtered using voice activity detection with Silero, so pauses do not become garbled text. Transcription uses your choice of models: Whisper models in Small, Medium, Turbo, and Large sizes, with GPU acceleration when available, or Parakeet V3, a CPU-optimized model with excellent performance and automatic language detection.

That local pipeline is the whole privacy story. There is no account, no server round trip, and no audio leaving your machine. The trade-off is the one every offline tool makes: you download a model file once, and the bigger models take real disk space and a capable machine, while the smaller ones run almost anywhere.

Download the latest release from the GitHub releases page or from handy.computer. On macOS it is also available via Homebrew with brew install --cask handy, and on Windows via winget with winget install cjpais.Handy. Note that the Homebrew cask and winget package are not maintained by the Handy developers. On Debian or Ubuntu, install the downloaded .deb with apt so required dependencies are installed automatically, and do not use dpkg on its own unless the dependencies are already there.

After installing, launch Handy and grant the system permissions it needs, microphone and accessibility, then configure your preferred keyboard shortcut in Settings and start transcribing. If you want to control it from elsewhere, a Raycast extension can start and stop recording, browse transcript history, manage the dictionary, and switch models and languages.

Power users get more. Command-line flags can control a running instance: handy --toggle-transcription, --toggle-post-process, and --cancel, plus startup flags like --start-hidden, --no-tray, and --debug. An advanced debug mode, revealed with Cmd+Shift+D on macOS or Ctrl+Shift+D on Windows and Linux, exposes paste delay tuning and a Reliable Paste beta for apps where the standard clipboard paste misbehaves.

Handy ships a model picker, and the choice is a classic size versus quality trade. Whisper Small is 487 MB, Medium is 492 MB in quantized form, Turbo is 1600 MB, and Large is 1100 MB quantized. Parakeet Unified EN 0.6B arrives as a single 731 MB file and is the recommended option when you want strong quality on CPU alone.

If you are behind a proxy, firewall, or restricted network where Handy cannot download models automatically, you can install them by hand. Find your app data directory in Settings, About, create a models folder inside it, drop the model file in with its exact filename, and restart Handy. Do not rename the files. Handy also picks up models already present in the shared Hugging Face cache, so a copy downloaded by another tool works without being moved.

Custom Whisper models work too. Drop any valid Whisper GGML .bin file into the models directory, restart, and it appears under Custom Models, named from its filename. Community models are user-provided and may not receive troubleshooting assistance, but the door is open for fine-tuned or community models the default list does not include.

Handy works on Windows, macOS, and Linux, but each platform has quirks the project documents honestly. On Linux, reliable text input needs the right tool for your display server: xdotool on X11, wtype on Wayland, or dotool for both, which needs your user in the input group. Without them, Handy falls back to a method with limited compatibility, especially on Wayland. Wayland global shortcuts must also be configured through your desktop environment, using the CLI flags as the shortcut command.

Linux startup can also trip on a missing runtime library: if Handy fails with an error about libgtk-layer-shell.so.0, install libgtk-layer-shell0 on Ubuntu or Debian, gtk-layer-shell on Fedora, RHEL, or Arch. The recording overlay is disabled by default on Linux because some compositors treat it as the active window and steal focus, which breaks pasting back into the app that triggered transcription.

On macOS, a Bluetooth headset microphone may temporarily reduce playback quality while recording, because Bluetooth switches to bidirectional audio; keeping headphones as output and selecting a built-in or external mic avoids it. Shortcuts using the fn or Globe key only work on Apple keyboards, because fn is not part of the standard USB keyboard specification and third-party keyboards handle it in firmware. On Windows, Handy asks the Vulkan loader to skip implicit layers to avoid crashes from overlay and capture tools, which you can opt out of with an environment variable.

Install the release for your platform, grant microphone and accessibility permissions, and set a shortcut you will actually remember. A good first test is dictating into a plain text editor: hold the shortcut, say a sentence with a number and a name in it, release, and check what lands. If the text is right, try it in the app where you actually write, an email, a document, or a chat window, because pasting behavior can differ between apps.

Then pick your model deliberately. Start with a smaller Whisper model or Parakeet if you are on CPU, and move up only if accuracy disappoints you. If a previous clipboard item gets pasted instead of your transcription, open Debug mode and try Reliable Paste on macOS or Windows, or increase the paste delay, and test again in the app where it happened.

If you live on the command line or a tiling window manager, wire the CLI flags into your own shortcuts. On Sway or i3, bindsym $mod+o exec handy --toggle-transcription does the job, and on Hyprland the equivalent bind line works the same way. On Linux you can also trigger it with pkill -USR2, which simply delivers a signal and does not terminate the process.

Handy is actively developed, and the project is transparent about its rough edges, which is a good sign, but set expectations accordingly. The known-issues list covers the platform quirks above, plus overlay and pasting behavior on Linux, and the troubleshooting docs are genuinely detailed, including the exact debug settings to try when paste misbehaves.

Release artifacts are signed with Tauri's updater signature format, and the project documents how to verify a download manually with minisign, explicitly warning not to use gpg for those .sig files. That is a nice touch for a privacy tool: the verification story is written down, not assumed.

Finally, remember the trade you are making. Offline means private and free of subscriptions, but it also means you manage models, disk space, and platform quirks yourself. If you want dictation that just works everywhere with no setup, a cloud service will feel smoother. If you want your voice to stay on your computer, cost nothing, and be open source, Handy is the tool built for exactly that, and its 36,127 stars suggest a lot of people wanted exactly that.