Important

This repository is archived. The code moved to Crown-OS/crownOs.

It is now crates/crowndictator in the CrownOS workspace, with its history intact. The crate name on crates.io is unchanged -- cargo add crowndictator is unaffected by where the source lives.

The nine desktop crates were merged because a change crossing two of them used to be two pull requests in two repositories that nothing built together. A schema change once broke the compositor for eight days before anyone noticed. One workspace, one lockfile, one CI run.

Issues and pull requests: https://github.com/Crown-OS/crownOs

This repository stays read-only so existing links keep resolving. Everything below describes the state at the time of the merge.

crowndictator

Push-to-talk voice dictation for Wayland.

Hold the shortcut, speak, release: the recording is transcribed locally with NVIDIA Parakeet TDT 0.6B v2 (ONNX Runtime, CUDA when available) and typed into the focused window. A small waveform pill sits at the bottom of the screen while it listens and thinks.

Nothing is sent to a server.

Status: Early. v0, but internally coherent and usable. The prerequisites are substantial — read them before starting.

How it works

hold hotkey ──► evdev reads /dev/input/event*   (bypasses the compositor)
                       │
                       ▼
               cpal captures 16 kHz mono f32
                       │
                       ▼
     ONNX Runtime: mel → conformer encoder → greedy TDT decode → detokenise
                       │
                       ▼
     wtype ──fallback──► ydotool ──fallback──► wl-copy + notify-send

The ASR engine loads on demand and drops the model after 300 s idle to free VRAM, so the first transcription after a pause is slower.

Both the hotkey and the text injection deliberately bypass the compositor, so it works under any Wayland compositor rather than only crownpositor.

Prerequisites

System packages (Arch):

sudo pacman -S --needed base-devel pkgconf alsa-lib libevdev \
  wayland wayland-protocols libxkbcommon \
  vulkan-icd-loader mesa libglvnd fontconfig dbus \
  wtype wl-clipboard libnotify

Debian/Ubuntu equivalents: Prerequisites.

/dev/input read access. The global hotkey is detected by reading /dev/input/event* directly, so your user must be in the input group:

sudo usermod -aG input "$USER"    # log out and back in

ONNX Runtime with CUDA. ort is pinned exactly (= "2.0.0-rc.12") and the cuda feature is not optional, so it is a hard build dependency even without NVIDIA hardware. The runtime falls back to CPU.

A large first-run download. Model weights come from Hugging Face: roughly 700 MB for the int8 CPU model, 2.5 GB for fp32 on GPU.

The dev overlay. crownshell = "0.3" and crownos-config = "0.2" are ordinary crates.io dependencies, but neither version is published — crates.io has only crownshell 0.1.0 and 0.2.0, and no crownos-config at all. A fresh clone fails at cargo metadata until Cargo is pointed at local checkouts. crownos-setup's ./bootstrap.sh --dev clones the repos side by side and writes a [patch.crates-io] overlay into a .cargo/config.toml one directory above them:

~/crownos/
├── .cargo/config.toml   # [patch.crates-io] crownshell, crownos-config
├── crowndictator/
├── crownshell/
└── crownos-config/

Cargo walks up from the working directory to find that file, and the paths in it are relative to the file's own directory. No particular layout inside a repo is required.

A wlr-layer-shell compositor for the overlay.

Build and run

cargo run -- --demo               # cycle the overlay states with fake audio.
                                  # No model download, no microphone, no input
                                  # group needed. Use this for UI work.
cargo run                         # the daemon
cargo run -- --transcribe f.wav   # one-shot, 16 kHz wav
cargo run -- --cpu                # skip CUDA

--demo is the contributor-friendly path — it exercises every visual state without touching a model or /dev/input.

Configuration

~/.config/crownos/input.ron, followed live: turning dictation off releases the keyboard grab, and changing the shortcut re-arms it without a restart.

(
    dictation_enabled: true,
    dictation_microphone: "Blue Yeti Analog Stereo",
    dictation_hotkey: "Super+Alt+D",
    dictation_gpu: true,
)

Field	Default	Effect
dictation_enabled	true	Off keeps the daemon resident but idle
dictation_microphone	None	None follows the system default
dictation_hotkey	"Super+Space"	Held, not struck
dictation_gpu	true	The same switch as --cpu

A microphone name that no longer matches any device falls back to the default rather than failing to record.

The default hotkey collides with the compositor. crownpositor binds Super+Space to cycle-layout. Because crowndictator reads the keyboard directly, both fire. Rebind one of them.

Design note

controller.rs is worth reading. Hotkey edges and settings changes are both Events on one channel, which is what makes "the user switched dictation off while holding the chord" an ordinary sequence of messages rather than a race.

settings.rs is 40 lines and is the model to copy when bringing another CrownOS component onto the shared config convention.

Tests

cargo test

Eleven unit tests: ten in src/hotkey.rs covering chord parsing and matching, one in src/controller.rs. They need no model, no microphone and no /dev/input access. There are no integration tests, examples or benches.

Known limitations

• CUDA is not optional at build time, and ort is declared with download-binaries, so cargo build fetches a prebuilt ONNX Runtime over the network. That also means the docs.rs build fails permanently — docs.rs blocks network access — so this crate will not have rendered documentation until the ASR backend sits behind an optional feature.
• ort is declared default-features = false with tls-rustls rather than the default tls-native. That is deliberate: the default drags in native-tls and therefore system OpenSSL, which no other CrownOS crate needs, and it was the only reason this crate would not compile. Do not restore the defaults.
• It pulls crownos-config with default features, which includes xilem — a headless daemon dragging in a GUI toolkit. default-features = false is probably correct.
• First run needs network access and several gigabytes of disk.

Contributing

See the organization-wide contribution guide. Default branch here is main.

License

Licensed under the MIT License.