dicti

Local, offline, live dictation for Linux.

Tap key, talk, starts typing in your active window.

_{dicti is fully on-device dictation. No cloud or account needed.}

Hey and welcome

dicti transcribes your speech on your own machine with whisper.cpp and types it into the focused window as you talk. No cloud, no account, nothing leaves your laptop.

I built it because I missed the great dictation I had on the Mac, and Linux deserved one too. It's what I use every day. It's young and honest about its rough edges, and you're very welcome to kick the tires, file issues, or help push it further.

Status: v0.3.5 (alpha). Live streaming dictation that works well day to day. Tested on Debian + GNOME (X11). See the roadmap for what's next.

Calm by design

dicti stays out of your way. A single animated mark by your clock tells you everything at a glance: green at rest, pink while it listens, a green sweep while it transcribes. No floating windows, no notification on every word.

What it does

• Click and Talk. Live streaming: text shows up as you speak. Each pass re-transcribes the whole thing, so whisper always has full context (quality matches one-shot batch mode), and only words that have stabilised across passes get typed. It's append-only, so it never rewrites text behind your cursor. To switch to all-at-once: change mode = "batch".
• No halucinations on silence. whisper-server runs with voice-activity detection (padded to not clip the first word), and the stabilise-across-passes rule is a second filter, so the classic silence hallucinations never reach the screen.
• Speaks ~100 languages. It's Whisper underneath, so whatever Whisper can transcribe, dicti types. It auto-detects the language each pass, or you can pin one in config.
• Types anywhere. ASCII goes in via ydotool; any non-ASCII (accents, Polish ąęóśżźćń, CJK, emoji) is pasted via the clipboard, so it works across editors, IDEs and terminals.
• Not in your face No popup spam, no tray clutter, no notification on every word. The top-bar glyph (above) is all you normally see.
• Yours, offline. whisper.cpp medium model, GPU-accelerated via Vulkan. Long sessions (1-hour cap) with a smart silence auto-stop.

How it works

Two small user services (whisper-server, dictation) plus a GNOME Shell extension (dicti@local) for the indicator. The daemon mirrors its state to $XDG_RUNTIME_DIR/dictation.state so the indicator can follow along.

[your key] -> keyd -> Super+Shift+Alt+F12 -> GNOME shortcut -> dictate-toggle
   -> unix socket -> dictation daemon -> pw-record -> /tmp WAV
   -> HTTP -> whisper-server (Vulkan) -> transcript -> ydotool / clipboard -> focused window

Quick start

You'll need:

• A Debian/Ubuntu-family distro (apt) with PipeWire audio
• A Vulkan-capable GPU (integrated is fine; CPU works but is ~4-5x slower)
• GNOME Shell (tested on 48) for the top-bar indicator
• ~1GB disk for whisper.cpp and the quantized medium model (the ~1.5GB full model is downloaded once to quantize, then deleted). While running, the model stays hot in RAM (~0.5GB) for instant transcription.

git clone https://github.com/tksimson/dicti.git
cd dicti
bash install/install.sh

The guided installer runs the phases in order: system packages, the keyd key remap, building whisper.cpp and fetching the model, the user services, ydotool, the GNOME shortcut, and the indicator extension. You'll be asked to log out and back in once after the first phase so your input-group membership takes effect. Each phase under install/00..07 also runs on its own.

If your dictation key isn't a Copilot/AI key, run sudo evtest (or wev on Wayland) to find its KEY_* name and edit keyd/default.conf.

Using it

• Tap your bound key to start listening, tap again to transcribe and insert.
• Pause to think; it won't cut off until a few minutes of real silence.
• Click the indicator for its menu: Start / Stop dictation and the Translate-to-English switch.
• From a shell: dictate-toggle [START|STOP|TOGGLE|CANCEL|STATUS].
• Lost a good dictation? dictate-last reprints the last one (--copy to clipboard).
• Speak any language, type English: dictate-translate toggles Whisper's translate task (bind it to a key). One-directional, English is always the target (a Whisper limit).

Configuration

dicti runs fine with no config. To tweak, edit ~/.config/dicti/config.toml (the installer seeds one from config/config.toml.example), then systemctl --user restart dictation. Common knobs:

Tip: Whisper transcribes one language per pass. auto works well for mixed use now that streaming keeps full context, but pin language to your main one if a quiet voice gets detected wrong.

Compatibility

dicti is young, so this is honest about what's actually been tested. Help moving things up the list is very welcome.

Troubleshooting

Roadmap & contributing

Where it's headed (and what's deliberately not in scope) lives in ROADMAP.md. Issues and PRs are welcome, especially test reports from other desktops and distros. See CONTRIBUTING.md for the dev setup, branches, and how to send a PR.

License

MIT, see LICENSE. Uses OpenAI's Whisper model via whisper.cpp; please respect the model's license and terms.