Speaking is just easier.

Speak freely, type instantly — 100% local voice dictation for Linux with 25+ languages, translation, speaker diarization, and real-time visual feedback. Text appears right where your cursor is.

Installation • Configuration • Visual interfaces • Usage • Going further • Roadmap

---

dictee is a complete voice dictation system for Linux. Transcription is performed 100% locally — no audio data ever leaves your machine. Press a shortcut, speak, and the text is typed directly into the active application.

• 4 ASR backends: Parakeet-TDT (25 languages, native punctuation), Canary-1B (built-in translation, GPU), Vosk (lightweight, ~50 MB), faster-whisper (99 languages)
• Daemon mode: model loaded once, near-instant transcriptions (~0.8s on CPU)
• Translation: 4 backends — Google, Bing, LibreTranslate (local), ollama (local)
• Speaker diarization: who said what, up to 4 speakers via Sortformer (CLI only, not yet in voice dictation)
• 3 visual interfaces: KDE Plasma widget, notification area icon, fullscreen animation

---

Installation

Download the .deb from the Releases, then:

# GPU version (requires the NVIDIA CUDA repository — see "GPU dependencies" below)
sudo dpkg -i dictee-cuda_1.2.0_amd64.deb

# CPU version (any computer, no extra repository needed)
sudo dpkg -i dictee-cpu_1.2.0_amd64.deb

# Install missing dependencies
sudo apt-get install -f

Note: The GPU version requires cuDNN from the NVIDIA CUDA repository, which is not included in standard Ubuntu/Fedora repos. Without it, the GPU version will still work but in CPU mode only.

Fedora / openSUSE:

# GPU version (NVIDIA CUDA — see "GPU dependencies" below)
sudo dnf install ./dictee-cuda-1.2.0-1.x86_64.rpm

# CPU version (any computer)
sudo dnf install ./dictee-cpu-1.2.0-1.x86_64.rpm

Arch Linux (AUR):

A PKGBUILD is available in the repository root. It builds from source and includes all components (x86_64 and aarch64).

aarch64 (ARM64):

Pre-built packages are x86_64 only. On aarch64 (Raspberry Pi 5, Ampere, etc.), build from source — see below. CUDA is limited to NVIDIA Jetson on this architecture; most users will use CPU mode.

Other distributions (.tar.gz):

tar xzf dictee-1.2.0_amd64.tar.gz
cd dictee-1.2.0
sudo ./install.sh

From source:

tar xzf dictee-1.2.0-source.tar.gz
cd dictee-1.2.0-source
cargo build --release --features sortformer
sudo ./install.sh

For detailed build instructions and Cargo features, see docs/building.md.

GPU version: NVIDIA CUDA dependencies

The GPU version (dictee-cuda) requires cuDNN, which is not available in standard Ubuntu/Fedora repositories. You need the NVIDIA CUDA repository:

Ubuntu / Debian:

wget -qO - https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/3bf863cc.pub | \
  sudo gpg --dearmor -o /usr/share/keyrings/cuda-archive-keyring.gpg
echo "deb [signed-by=/usr/share/keyrings/cuda-archive-keyring.gpg] \
  https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/ /" | \
  sudo tee /etc/apt/sources.list.d/cuda-ubuntu2404-x86_64.list
sudo apt update
sudo apt install libcudnn9-cuda-12

Replace ubuntu2404 with your version (ubuntu2204, ubuntu2504, etc.). See NVIDIA CUDA repos.

Fedora:

sudo dnf config-manager addrepo --from-repofile=https://developer.download.nvidia.com/compute/cuda/repos/fedora41/x86_64/cuda-fedora41.repo
sudo dnf install libcudnn9-cuda-12

Without cuDNN, the GPU version falls back to CPU automatically. dictee-setup will detect this and guide you through the setup.

---

Configuration

On first launch, a setup wizard guides you through backend selection, model download, and keyboard shortcuts. You can reconfigure anytime from the application menu, tray icon, Plasma widget, or by running dictee --setup from the terminal:

ASR backend

Four mutually exclusive transcription backends, switchable from dictee --setup:

Backend	Languages	Model size	Warm latency	Type
Parakeet-TDT	25	~2.5 GB	~0.8s CPU · ~0.16s GPU	ONNX Runtime (Rust)
Canary-1B	4 (EN,ES,FR,DE)	~5 GB	~0.7s GPU	ONNX Runtime (Python, GPU recommended)
faster-whisper	99	~500 MB–3 GB	~0.3s	CTranslate2 (Python)
Vosk	9+	~50 MB	~1.5s	Python (lightweight)

Each backend runs as a systemd user service — same Unix socket protocol, fully transparent to the user.

Keyboard shortcuts

dictee --setup captures and registers shortcuts automatically (KDE Plasma / GNOME). Two separate shortcuts: one for dictation, one for dictation + translation.

For tiling WMs (Sway, i3, Hyprland…), the tool shows the command to add manually to your config.

Translation

Backend	Privacy	Speed	Quality	Setup
Canary-1B	100% local	Built-in	Best	Included with ASR backend
LibreTranslate	100% local	0.1–0.3s	Good	Guided from setup
ollama	100% local	2.3–3.4s	Best	Guided from setup
translate-shell (Google)	Online	0.2–0.7s	Good	Included
translate-shell (Bing)	Online	1.7–2.2s	Good	Included

Quick backend switching

Switch ASR or translation backend instantly from the command line, the tray icon menu, or the Plasma widget:

# Switch ASR backend
dictee-switch-backend asr canary

# Switch translation backend
dictee-switch-backend translate ollama

# Show current backends
dictee-switch-backend status
# → ASR: parakeet (dictee.service, active)
# → Translate: google (trans)

The tray icon and Plasma widget include sub-menus for switching backends without opening the configuration.

---

Visual interfaces

KDE Plasma widget

A native KDE Plasma 6 widget with real-time audio visualization during recording, daemon status, and quick controls (dictate, translate, cancel).

Five animation styles with Hanning envelope, per-style sensitivity, and optional color gradients:

Bars	Wave	Pulse	Dots	Waveform

All styles support color gradients, adjustable Hanning envelope (shape and center frequency), per-style sensitivity curve, and fine-tuning options (bar count, spacing, radius, speed…).

# Install (included in .deb, or manually)
kpackagetool6 -t Plasma/Applet -i /usr/share/dictee/dictee.plasmoid

Right-click on the panel → "Add Widgets…" → search for "Dictée".

For full widget settings documentation, see docs/plasmoid.md.

Notification area icon (dictee-tray)

dictee-tray is the alternative to the KDE Plasma widget for non-KDE desktops (GNOME, Xfce, Sway, Hyprland…). It displays a notification area icon reflecting the real-time state: idle, recording (green), transcribing (blue), daemon stopped (red).

• Left click → start dictation
• Middle click → cancel
• Context menu → all actions (dictation, translation, daemon, configuration)

# Launch manually
dictee-tray

# Enable at session startup
systemctl --user enable --now dictee-tray

The icon automatically adapts to light/dark themes.

Both the Plasma widget and the tray icon include:

• Backend selectors — switch ASR and translation backends without opening dictee-setup
• First-run detection — prompts to run the setup wizard if not yet configured
• Install detection (Plasma widget) — shows a clear message if dictee is not installed

animation-speech

animation-speech is a standalone project that provides a fullscreen visual animation during recording, with cancellation via Escape key. It works on any Wayland compositor supporting wlr-layer-shell (KDE Plasma, Sway, Hyprland…).

sudo dpkg -i animation-speech_1.2.0_all.deb

Download: animation-speech releases

Note: animation-speech is not compatible with GNOME (no wlr-layer-shell support). GNOME users can use dictee-tray for visual feedback. Contributions for a GNOME Shell extension are welcome — see the plasmoid source for reference architecture.

Without any visual interface, dictee works normally but without visual feedback during recording.

---

Usage

# Simple dictation — transcribe and type
dictee

# With translation (default: system language → English)
dictee --translate
dictee --translate --ollama    # 100% local translation via ollama

# Change translation languages
DICTEE_LANG_TARGET=es dictee --translate    # → Spanish

# Cancel an ongoing recording (via shortcut or Escape key)
dictee --cancel

# Test post-processing rules
dictee-test-rules                    # interactive mode
dictee-test-rules --loop             # continuous test loop
dictee-test-rules --wav file.wav     # test from audio file

# Switch backend from command line
dictee-switch-backend status         # show current backends
dictee-switch-backend asr canary     # switch to Canary
dictee-switch-backend translate bing # switch translation to Bing

---

Going further

Post-processing

dictee includes a configurable text transformation pipeline that runs after transcription:

• Custom rules — regex-based text replacements (e.g., voice commands like "new line", "comma")
• Dictionary — replace common ASR mistakes with correct words
• Continuation — detect incomplete sentences across multiple dictations
• Elisions — French grammar rules (e.g., "le arbre" → "l'arbre")
• Number conversion — spoken numbers to digits (e.g., "vingt-trois" → "23")
• Auto-capitalization — capitalize after sentence-ending punctuation
• LLM correction — optional grammar/spelling fix via Ollama before rules

Configure from dictee --setup → Post-processing tab, or test rules with dictee-test-rules.

Documentation	Description
docs/cli-programs.md	CLI binaries, direct usage, ONNX models
docs/building.md	Building from source, Cargo features, audio pipeline
docs/plasmoid.md	Widget settings, animation styles, configuration details
Post-processing	Text transformation pipeline: rules, dictionary, elisions, text2num, capitalization, LLM correction

---

Roadmap

v1.2.0 (current): 4 ASR backends (+ Canary), post-processing pipeline, quick backend switching, first-run wizard, dictee-test-rules

• (v1.3) Hotword boosting — bias ASR decoding toward custom names and terms without retraining (beam search + Aho-Corasick in Rust)
• Diarization from tray/plasmoid — select audio file, get speaker-labeled transcription
• CLI for speech-to-text (pipe audio, get text)
• dictee-ctl coordinator — single entry point, eliminates race conditions
• VAD (Voice Activity Detection) — hands-free dictation without push-to-talk
• Real-time streaming transcription with live text display
• Built-in visual overlay (replace external animation-speech)
• AppImage / Flatpak packaging
• COSMIC / GNOME applet (contributions welcome!)

Credits

The transcription engine is built on parakeet-rs by Enes Altun, which provides the Rust library for NVIDIA Parakeet model inference via ONNX Runtime. The Canary-1B backend uses onnx-asr by Ivan Stupakov for ONNX-based ASR inference.

License

This project is distributed under the GPL-3.0-or-later license (see LICENSE).

The original parakeet-rs code by Enes Altun is under the MIT license (see LICENSE-MIT).

dotool by geb is bundled for keyboard input simulation and is under the GPL-3.0 license.

The Parakeet ONNX models (downloaded separately from HuggingFace) are provided by NVIDIA. This project does not distribute the models.