Dictation

Hold F5 (or Fn for more than 0.3 s), speak, and let go: the text is typed into the active window. It runs fully offline with Voxtype and NVIDIA's Parakeet model, on the CPU (about 1.4 s for 6 s of speech on an M2 Max).

Voxtype isn't packaged for Fedora on aarch64, so it is installed by hand:

  1. Download the aarch64 binaries from the Voxtype releases into ~/.local/bin and check their sha256: voxtype-onnx (Parakeet), plus voxtype-osd, voxtype-osd-quickshell and voxtype-audio-bridge for the on-screen recording card.
  2. Download the model: voxtype-onnx setup --download --model parakeet-tdt-0.6b-v3-int8 (into ~/.local/share/voxtype/models).
  3. Install the OSD's QML: voxtype setup quickshell --source <voxtype repo>/quickshell. link.sh then replaces its OsdSurface.qml with the one in config/voxtype/ (card at the top centre, no dimming).
  4. sudo dnf install wtype so the text can be typed into windows (without it, it goes to the clipboard).

The settings are in config/voxtype/config.toml, which link.sh links into ~/.config/voxtype/. Hyprland starts the daemon at login, and hypr/scripts/dictate handles the keys and the muted-mic check.

Timings for 6 s of speech on the CPU: Parakeet 1.4 s, Whisper small 4.3 s, medium 13.7 s, large-v3-turbo 24 s.

From docs/dictation.md in the Nightdrive repo