Whisper Dictate

Push-to-talk voice dictation for Ubuntu, in Greek and English, running Whisper locally on your GPU (or CPU). Nothing leaves your machine.

The interaction copies Wispr Flow:

Action	What happens
Hold Right Ctrl and speak, release	Your words are inserted at the cursor of whatever app is focused
Tap Right Ctrl twice	Hands-free mode: recording keeps going until you press Right Ctrl again
Esc while recording	Discards the recording
Tray icon	Language (Auto / Ελληνικά / English), hands-free toggle, copy last dictation, history, config, quit

A small dark pill appears at the bottom of the screen while you dictate: a live waveform while you talk, a shimmer while Whisper transcribes, then a short preview of the inserted text. It never steals keyboard focus. While idle, nothing is shown; set show_idle_pill = true in the [ui] section of the config to keep a thin bar on screen as a sign that the app runs.

Auto-edits, like Wispr Flow:

• language of every dictation is detected automatically between Greek and English
• filler words are removed ("um", "uh", "εε", "εμ" ...)
• "new line" / "new paragraph" / "νέα γραμμή" / "νέα παράγραφος" become line breaks
• spoken punctuation works when you want to be explicit: "question mark", "comma", "full stop", "exclamation mark", "ερωτηματικό", "κόμμα", "τελεία", "θαυμαστικό" (Whisper already punctuates from your intonation, so this is optional)
• first letter of the text and of each sentence is capitalized, a space is added after the text so dictations chain nicely
• known Whisper hallucinations on silence ("Υπότιτλοι AUTHORWAVE", "Thank you.") are dropped
• every dictation is saved to ~/.local/share/whisper-dictate/history.jsonl

Requirements

• Ubuntu 22.04 / 24.04 with an X11 session (echo $XDG_SESSION_TYPE must print x11). On Wayland, global hotkeys and text insertion are blocked by the compositor. Log out, click the gear on the login screen and choose Ubuntu on Xorg.
• An NVIDIA GPU with 4 GB+ VRAM for the default turbo model (about 1 s latency for a sentence). Without a GPU set name = "small" or "base" in the config and beam_size = 1.
• Python 3.10+ (system Python 3.12 on 24.04 works).

Install

git clone https://github.com/AlexJoom/whisper-dictate.git
cd whisper-dictate
./install.sh

The script installs the few missing system packages (asks for sudo only then), creates .venv in the repo, installs openai-whisper with PyTorch (CUDA) and the app, downloads the turbo model (about 1.6 GB), adds a launcher at ~/.local/bin/whisper-dictate, a desktop entry and an autostart entry, and starts the app.

Start it later from the app grid ("Whisper Dictate") or:

whisper-dictate

Configuration

~/.config/whisper-dictate/config.toml is created on first run with comments for every option. The most useful ones:

[hotkey]
key = "ctrl_r"            # or "alt_r", "pause", "f8", "<ctrl>+<super>" ...

[model]
name = "turbo"            # tiny | base | small | medium | large-v3 | turbo
languages = ["el", "en"]  # auto-detect among these; one entry pins the language
beam_size = 5             # 1 is faster on CPU

[inject]
method = "paste"          # paste | type | xdotool | clipboard

Restart the app (tray icon, Quit) after editing.

Debugging

whisper-dictate -v                      # verbose log
whisper-dictate --mic-check 4           # record 4 s and print the transcript
whisper-dictate --file some_audio.mp3   # transcribe a file through the same pipeline
pactl list sources short                # find your microphone name for [audio] source

Log file when started by the installer or autostart: ~/.cache/whisper-dictate.log.

How it works

Right Ctrl held ──► parec (16 kHz mono) ──► Whisper turbo (fp16, CUDA)
                                                  │ detect el/en, decode
                                                  ▼
                              clean-up (fillers, commands, caps) ──► clipboard + Ctrl+V

• whisper_dictate/hotkey.py global hotkey with pynput
• whisper_dictate/audio.py microphone capture through PulseAudio/PipeWire parec
• whisper_dictate/engine.py model loading (fp16 on GPU, fits in 4 GB) and transcription
• whisper_dictate/textproc.py auto-edits (unit tested in tests/)
• whisper_dictate/inject.py insertion into the focused window (Ctrl+Shift+V in terminals)
• whisper_dictate/overlay.py the bottom-of-screen pill (GTK 3 + cairo)
• whisper_dictate/tray.py AppIndicator menu

Run the tests with .venv/bin/python -m pytest tests.

License

MIT. Whisper itself is MIT licensed by OpenAI.