Convert speech to text locally with a desktop app
Press a shortcut key and speak, and your words appear as text in the input field. All processing happens on your computer without an internet connection, so you can use it for private conversations or sentences containing confidential information. Handy is a desktop app that removes the barrier between speaking and typing.
A fully offline process
Handy refines recorded audio with Silero VAD and converts it to text using Whisper or Parakeet models. This entire process runs locally, so audio data is not sent to external servers. If a GPU is available, it uses acceleration; otherwise, it uses CPU-optimized models to maintain processing speed. It operates in the same way on Windows, macOS, and Linux.
Control recording and input with a single shortcut
Users can assign a key in the settings to start and stop recording. Recording continues while the key is held down, and when released, the text is immediately pasted into the input field of the currently focused app. A toggle mode is also supported, allowing you to start with one press and stop with another. Integration with Raycast allows for faster control of recording, log lookup, and model changes.
Platform-specific installation and considerations
macOS can be installed via Homebrew Cask, Windows via winget, and Linux requires installing the .deb file with APT to automatically resolve dependencies. Linux users must install one of xdotool, wtype, or dotool depending on their display server for text input to work correctly. In Wayland environments, wtype or dotool is essential, and using dotool requires adding the user to the input group. On macOS, using a Bluetooth headset microphone may degrade audio quality during recording, so using the built-in or an external microphone is recommended.
Model selection and custom support
You can choose from Whisper's Small, Medium, Turbo, and Large models, as well as the CPU-optimized Parakeet V3 model. Fine-tuned Whisper GGML models created by the community can also be used by placing them as .bin files in the models folder, where they are automatically recognized. Since the model name is inferred from the file name, you should name the file meaningfully.
Stability and license verification
Handy is released under the MIT license and can be used freely, but brand assets such as the name and logo are not open source, so you must use your own branding when redistributing. On Linux, startup may fail due to the absence of the gtk-layer-shell library, so you must install the package appropriate for your distribution. In Wayland environments, the overlay may steal focus and cause pasting to fail, so it is recommended to change the overlay position to 'None' in the settings. The project is under active development, and known issues and solutions are transparently disclosed in the README.
