Dita

Voice transcription overlay for Linux. Press a hotkey, speak, get text on clipboard.

Dita opens as a floating, frameless window over your screen. It records from your microphone, transcribes with faster-whisper, and automatically copies the result to the clipboard. No browser, no cloud, no subscription.

Features

• Global hotkey (configurable) — one key to open, record, stop, and close
• Local transcription via Whisper (runs fully offline after first setup)
• Auto-copy to clipboard on completion
• Glassmorphism UI, always-on-top, frameless
• Configurable model size, language, and auto-close delay

Requirements

• Ubuntu / Debian (x86_64)
• Python 3.11+
• System packages: python3-gi python3-gi-cairo gir1.2-webkit2-4.1 libportaudio2 xclip ffmpeg

Install

Download the latest .deb from the Releases page, then:

sudo dpkg -i dita_*.deb
sudo apt-get install -f   # install any missing system deps

The installer creates a Python venv under /opt/dita/ and downloads the Whisper model (~1.5 GB for medium). This takes a few minutes on the first install.

After that, launch Dita from your application menu or run:

dita

Usage

Action	Hotkey (default)
Open / show window	Ctrl+Shift+Space
Start recording	Ctrl+Shift+Space (window open + idle)
Stop & transcribe	Ctrl+Shift+Space (while recording)
Close window	Esc or hotkey when done

The transcribed text is copied to clipboard automatically. Paste with Ctrl+V.

Configuration

Dita reads ~/.config/dita/dita.toml on startup. The file is never written by the app — create it manually to override defaults.

# ~/.config/dita/dita.toml

model = "medium"                      # Whisper model size
language = "pt"                       # transcription language
hotkey = "<ctrl>+<shift>+<space>"     # pynput key string
auto_copy = true                      # copy result to clipboard
auto_close_after = 3                  # seconds to close after transcription (0 = never)
device = "cpu"                        # cpu | cuda
compute_type = "int8"                 # int8 | float16 | float32

[theme]
name = "dark"                         # dark | light
# Optional color overrides:
# background = "#1a1a1e"
# surface    = "#2a2a2e"
# text       = "#e2e2e2"
# subtext    = "#888888"
# accent     = "#e05252"
# border     = "rgba(255,255,255,0.08)"

Options

Key	Type	Default	Description
model	string	medium	Whisper model: tiny base small medium large-v3
language	string	pt	ISO language code or auto for detection
hotkey	string	<ctrl>+<shift>+<space>	Global hotkey in pynput format
auto_copy	bool	true	Copy transcription to clipboard automatically
auto_close_after	int	3	Seconds before window closes (0 = never)
device	string	cpu	Inference device: cpu or cuda
compute_type	string	int8	Quantization: int8 (cpu), float16 or float32 (cuda)

[theme] section

Key	Type	Default	Description
name	string	dark	Base theme: dark or light
background	string	(theme default)	Main window background color
surface	string	(theme default)	Card / panel background
text	string	(theme default)	Primary text color
subtext	string	(theme default)	Secondary / muted text color
accent	string	(theme default)	Accent / highlight color
border	string	(theme default)	Border color (supports rgba)

Color overrides in [theme] are optional — omit any key to use the theme default.

Development

# Prerequisites
sudo apt install python3-dev python3-gi python3-gi-cairo gir1.2-webkit2-4.1 libportaudio2 xclip

# Install deps (requires uv)
make install

# Run
make run

# Lint & format
make lint
make fmt

# Tests
make test

# Build .deb
bash build-deb.sh

uv install: curl -LsSf https://astral.sh/uv/install.sh | sh

License

MIT