WhisperFlow-style voice dictation for Linux — Groq Whisper + LLM formatting, pastes into any focused window.
Press a keybind, speak, press again — the text appears in whatever window has focus. Transcription via Groq's Whisper Large V3, optional clean-up via a Groq LLM, history kept locally in SQLite.
[Press shortcut] → Recording…
[Speak] "das meeting ist um fünf uhr ähm nee doch nicht um fünf um sieben"
[Press shortcut] → Processing…
[Output] "Das Meeting ist um sieben Uhr."
- Fast transcription — Whisper Large V3 on Groq, sub-second for short clips
- Writes what you meant — resolves mid-sentence self-corrections ("um fünf, nee, um sieben"), drops filler and stutters, keeps your slang and your language
- Learns your vocabulary — mines your own history in the background for the names it keeps mangling ("SAP-Agenten" → "Subagenten") and feeds them back into the recogniser. A term must be proposed by several independent runs before it goes live, so guesses never reach your text
- Never swallows text — truncated or refusal-style LLM output is detected and replaced with the raw transcript
- Tiny upload — opusenc 16 kbps VOIP, ~4 KB per second of audio (~10× smaller than the previous ffmpeg/48k pipeline)
- Modern GUI — Tauri 2 + React 19 + Tailwind v4, dark UI, system tray, lives in
~/.local/share/plauder - Verlauf tab — virtualized history list with git-style word diff between Whisper raw and LLM formatted output, character-level change percentage, anomaly badges (refusal / truncated / inflated)
- Editable corrections — teach Plauder by adding "Whisper hears X → I meant Y" pairs; applied by meaning, so
Cloud → Claudefixes "yo Cloud, guck dir die Logs an" without touching "läuft in der Cloud" - Pinned vocabulary — your names and jargon in Settings, sent to Whisper as a decoding hint rather than patched up afterwards
- Live Logs tab — last 100 lines by default, scroll up to auto-load older
- API-key reveal — Eye/EyeOff toggle in Settings instead of a permanent password field
- 50+ languages — auto-detect or pin via the Settings/tray dropdown
- Keybind-driven — works with i3/sway/Hyprland/sxhkd/dwm
Every tagged release ships a Debian package and an AppImage on the Releases page.
# .deb
sudo apt install ./plauder_*_amd64.deb
# AppImage
chmod +x plauder_*.AppImage
./plauder_*.AppImagegit clone https://github.com/chukfinley/plauder.git
cd plauder
./install.shThe installer builds the Tauri + React GUI in release mode, copies everything to ~/.local/share/plauder/, drops a plauder symlink in ~/.local/bin/, sets up the systemd user service, and prompts for your Groq API key.
- Linux with X11 or Wayland
- PipeWire or PulseAudio (
pw-record) - Rust/Cargo + Node.js + pnpm (build from source only)
- WebKitGTK 4.1 dev libs (Tauri runtime)
- Groq API key — console.groq.com/keys
Required at runtime:
yad— tray iconxdotool— paste into focused windowxclip— clipboardopusenc(recommended) orffmpeg— audio compressionjqcurlsqlite3pw-record(PipeWire)
Ubuntu/Debian:
sudo apt install yad xdotool xclip opus-tools ffmpeg jq curl pipewire sqlite3 \
libwebkit2gtk-4.1-dev build-essential libssl-dev libayatana-appindicator3-dev librsvg2-dev
# build-from-source extras:
sudo apt install cargo nodejs && sudo npm install -g pnpmArch/Manjaro:
sudo pacman -S yad xdotool xclip opus-tools ffmpeg jq curl pipewire sqlite3 \
webkit2gtk-4.1 libappindicator-gtk3 librsvg
# build-from-source extras:
sudo pacman -S rust nodejs pnpmEdit ~/.local/share/plauder/.env or use the Settings tab in the GUI:
GROQ_API_KEY="gsk_…" # Required
LANGUAGE="de" # Optional: auto-detect if empty
MIC_SOURCE="" # Optional: device id from `pactl list sources short`
NOTIFICATIONS="true"
TRAY_ICON="true"
VOCABULARY="…" # Optional: your names/jargon, comma-separated
WHISPER_MODEL="whisper-large-v3"
SYSTEM_PROMPT="…" # Optional: override the LLM clean-up promptBind your WM/DE to run plauder:
| WM/DE | Config | Example |
|---|---|---|
| i3 / sway | ~/.config/i3/config |
bindsym $mod+r exec plauder |
| Regolith | ~/.config/regolith3/i3/config.d/91_custom |
bindsym $mod+r exec plauder |
| Hyprland | ~/.config/hypr/hyprland.conf |
bind = SUPER, R, exec, plauder |
| sxhkd | ~/.config/sxhkd/sxhkdrc |
super + r plauder |
| dwm | config.h |
{ MODKEY, XK_r, spawn, SHCMD("plauder") } |
systemctl --user start plauder(tray daemon)- Press your keybind → recording
- Speak
- Press keybind again → transcribe + paste
- Verlauf — every recording, expandable card with Whisper raw / LLM formatted / your correction, git-style word diff, anomaly badges, % changed, copy/delete
- Korrekturen — add and edit "Whisper hears X → I meant Y" rules used by every future LLM call
- Logs — live debug log, last 100 lines, scroll up for older
- Einstellungen — API key (with show/hide), mic, language, notifications, tray, system prompt
- Toggle Recording
- Einstellungen & Historie (opens GUI)
- Select Microphone
- Select Language
- Quit
Keybind → pw-record 16 kHz mono WAV → /tmp
Keybind → opusenc 16 kbps VOIP (~16 ms encode, ~4 KB/s upload)
→ POST /audio/transcriptions (Whisper Large V3 + vocabulary prompt)
→ POST /chat/completions (gpt-oss-120b, medium reasoning effort)
→ xdotool pastes into the focused window
→ SQLite history.db gets the row + timings
→ detached: the learner mines the history for new vocabulary
Anti-swallow safeguard: if the LLM finishes with finish_reason=length, returns less than 40 % of the input, or matches one of the known refusal patterns ("kann ich nicht", "I cannot", "as an AI", …), Plauder pastes the raw Whisper transcript instead. The threshold is 40 % rather than 60 % because removing filler legitimately shrinks a rambling passage by a third.
Whisper accepts a prompt — text prepended as decoding context. It is a bias, not a rule, but it is enough to stop "Playwright" coming back as "Play-Red". Plauder fills it with your vocabulary, so mishearings get fixed at the source instead of being patched afterwards.
The prompt has to read like speech. Whisper was trained to continue running text and copies the register it is primed with, so a bare word list actively makes things worse. Measured on one clip of "Yo Claude, guck dir bitte die Logs an":
| prompt | transcript |
|---|---|
| none | Klart, guck dir die Loks an. |
| word list | Claude, Cucke Dilox an — name saved, sentence wrecked |
| sentence | Klaude, guck dir die Logs an. — sentence intact |
So the terms ride inside a carrier sentence. That sentence is deliberately dull and unrelated to any real request, because a prompt can leak into the transcript of a near-silent clip — better it leaks something obviously wrong than a plausible instruction.
That list grows on its own. After a recording (detached, never in the dictation path, at most every LEARN_INTERVAL_MIN), a model reads your recent history and proposes terms — using your hand-edits as ground truth, then the repairs the formatter already made, then recurring jargon.
Left unchecked this loop is dangerous in two ways, and both are guarded:
- It learns its own errors. The history is full of mishearings nobody fixed, so "Hix Field" looks like a frequent proper noun. Feeding that back would teach Whisper to make the mistake forever. The learner is required to output the correction, and a term is only accepted if the term itself or one of its claimed mishearings actually occurs in your transcripts.
- It guesses. Faced with a garble it doesn't recognise, the model invents a plausible fix — "Hix Field" became Hugging Face one run and Higgsfield the next; "Cal AI" (a real app) became Claude. Asking a second model to review the claims proved useless: it rejected "Peison → Python" as "sounds different". What works is the model's own inconsistency — guesses land somewhere new every run, real mishearings come back every time. So a term must be proposed by
LEARN_CONFIRMATIONSindependent runs before it goes live.
Inspect it with voice-input.sh --learn (run now, print the table) or --vocabulary (show the hint the next recording will use). voice-input.sh --format "text" runs the formatting stage on text, for tuning the prompt without a microphone.
~/.local/share/plauder/
├── plauder-gui # Tauri release binary
├── voice-input.sh # toggle script (recording + transcribe + paste)
├── voice-input-daemon.sh # tray daemon
├── select-mic.sh / select-language.sh
├── .env # config (API key, language, mic)
├── history.db # SQLite history + corrections
└── icons/ # tray + window icons
~/.local/bin/plauder # symlink → voice-input.sh
~/.config/systemd/user/plauder.service
Tauri 2 app: React/Vite frontend in src/, Rust backend in src-tauri/.
pnpm install
pnpm tauri dev # hot-reload
pnpm tauri build --no-bundle # binary → src-tauri/target/release/plauder-gui
pnpm tauri build # full .deb + .rpm + AppImage bundleData lives in the same history.db and .env the shell scripts use, so the GUI and CLI stay in sync.
Push a vX.Y.Z tag and the GitHub Action builds .deb, .rpm, and .AppImage and attaches them to a new release automatically.
git tag v0.2.0
git push origin v0.2.0systemctl --user stop plauder && systemctl --user disable plauder
rm -rf ~/.local/share/plauder ~/.local/bin/plauder
rm ~/.config/systemd/user/plauder.service- No tray icon — install
yad, make sure your DE has a system tray - Empty / wrong transcript — pin the language in Settings (auto-detect can miss for short clips)
- Slow — mostly Groq upload + LLM round-trip; with opusenc 16 kbps a 10 s clip uploads in < 200 ms on WiFi
- Tray panic on launch — make sure
~/.local/share/plauder/icons/32x32.pngexists; the binary loads it viainclude_bytes!at build time
MIT — see LICENSE.
- Groq — fast Whisper + LLM inference
- OpenAI Whisper — transcription model
- Tauri — Rust + WebView shell