Vera

A Siri-like voice assistant for the Linux desktop. Press a hotkey, speak, and a local LLM turns your words into actions — open apps, set the volume, lock the screen, control media, set timers. Fully offline: whisper.cpp and llama.cpp are embedded — no ollama, no server.

See PLAN.md for the design, model evaluation strategy, and the build gotchas discovered along the way.

Install

scripts/install.sh              # CPU build
scripts/install.sh cuda         # GPU build (needs the CUDA toolkit / nvcc)

This:

• builds and installs verad, vera, vera-stt, vera-eval to ~/.local/bin
• registers D-Bus activation (the daemon auto-starts on first use)
• runs vera setup, which binds Super+Space to vera toggle and adds a login autostart entry so the daemon (and models) preload at login (hotkey override: VERA_HOTKEY='<Ctrl><Alt>space' scripts/install.sh; skip autostart with vera setup --no-autostart afterwards)

The overlay is placed by the compositor — Wayland apps can't position their own windows.

.deb package

scripts/package-deb.sh          # or: scripts/package-deb.sh cuda
sudo apt install ./target/deb/vera_*.deb
vera setup                      # per-user: hotkey, autostart

Then download the models once:

vera models download whisper-base.en
vera models download qwen3-1.7b
vera models list                # see all candidates + download status

Use

• Voice: press Super+Space, speak, stop talking. Esc (or the hotkey again) cancels. The first command after daemon start loads the models (a few seconds); after that it's warm.
• Conversation: the overlay stays open — after every answer or action the mic reopens for ~4 s, so you can chain commands ("open firefox" → "volume to 30" → "thanks, that's all"). It closes when you say goodbye ("that's all", "never mind", "bye"), stay silent, or press Esc. Configure with conversation_mode / followup_timeout_ms.
• Text (also handy for testing): vera run "set volume to 40 percent"
• Stop the daemon: vera quit (it auto-starts again on next use)

Development / running from the repo

cargo build && cargo test

# terminal 1: daemon with logs
RUST_LOG=info ./target/debug/verad

# terminal 2:
./target/debug/vera run "open firefox"     # text command
./target/debug/vera toggle                 # voice session

The daemon log prints each stage: capture device, transcript, chosen tool call.

Configuration

~/.config/vera/config.toml (created with defaults on first run):

Key	Default	Meaning
stt_model	whisper-base.en	registry name or absolute path to a ggml whisper model
llm_model	qwen3-1.7b	registry name or absolute path to a GGUF
device	auto	auto / cpu / cuda (cuda needs the cuda build)
endpoint_silence_ms	800	trailing silence that ends an utterance
leading_silence_timeout_ms	6000	give up if you never speak
max_utterance_ms	15000	hard cap per utterance
result_linger_ms	4000	how long results stay on screen
language	en	whisper language hint, auto to detect
temperature	0.2	LLM sampling temperature

Restart the daemon (vera quit) after editing.

What it can do (v1 tools)

Open apps and URLs, web search, set/change/mute volume, media play/pause/next/ previous (MPRIS), lock screen, screen brightness, do-not-disturb, screenshots, timers with notification, free-form Q&A fallback, and suspend / power off / reboot (each asks for on-screen confirmation first).

Model bake-off

The default model was picked by measurement, not vibes — and you can rerun it:

vera models download gemma-4-e4b            # + qwen3-4b-instruct, phi-4-mini…
vera-eval --models qwen3-1.7b,gemma-4-e4b --data data/eval.jsonl --verbose

Prints a markdown table of tool-selection accuracy, argument accuracy, and latency per model. Set the winner as llm_model in the config.

Bake-off results (81 cases, CPU-only, i7-10875H, models under load):

model	tool acc	args acc	avg ms	parse fails
qwen3-1.7b	91.4%	91.4%	2917	0
qwen3-4b-instruct (default)	98.8%	97.5%	3315	0
gemma-4-e4b	100%	98.8%	6223	0
phi-4-mini	97.5%	93.8%	2679	0

qwen3-4b-instruct is the CPU sweet spot and the shipped default; gemma-4-e4b is the accuracy pick once you have the CUDA build.

Troubleshooting

• "I didn't catch anything" — check the default input device: wpctl status (Sources section). Vera records from the default source.
• "model missing" error in the window — run the vera models download commands above.
• Slow first response — that's the one-time model load + system-prompt decode; subsequent commands reuse the KV cache and are much faster.
• GPU build fails — you need nvcc (sudo apt install nvidia-cuda-toolkit) and a working driver (nvidia-smi).

Layout

Crate	What it is
vera-core	config, model registry/downloader, audio capture, VAD endpointing, STT client, llama LLM with GBNF-constrained tool calls, tool executors
verad	the daemon: GTK4 floating window + GApplication D-Bus actions, pipeline thread
vera	CLI: toggle (bound to the hotkey), run <text>, models …, quit
vera-stt	whisper.cpp helper process (whisper and llama bundle incompatible ggml copies, so whisper runs out-of-process)
vera-eval	model bake-off: labeled utterances → accuracy/latency table

License

MIT — see LICENSE.

whisper.cpp and llama.cpp are pulled in via the whisper-rs and llama-cpp-2 crates and carry their own MIT licenses.