Skip to content

Installation

Jump to bottom
rcspam edited this page May 7, 2026 · 16 revisions

🌐 Language: English | Français

Installation

dictee ships as pre-built packages for the major Linux distributions, plus a portable tarball and a from-source path for everything else. Both CPU and CUDA variants are published for each format.

The recommended path is the one-liner installer: it auto-detects your distribution and GPU, adds the NVIDIA CUDA repository if needed, then pulls the right package from the latest GitHub release.

First-run setup wizard

Table of Contents

One-liner install

curl -fsSL https://raw.githubusercontent.com/rcspam/dictee/master/install.sh | bash

The installer:

  1. Detects your distribution (Ubuntu, Debian, Fedora, openSUSE, Arch) via /etc/os-release.
  2. Detects an NVIDIA GPU via nvidia-smi (or asks you if detection is ambiguous).
  3. Adds the NVIDIA CUDA APT/DNF repository if you have a GPU (needed for libcudnn9-cuda-12 / cuda-cudart-12-x).
  4. Installs the matching .deb or .rpm from the latest GitHub release.
  5. Offers to launch the setup wizard (dictee-setup) if a graphical session is available.

Installer options

Pass options after --:

# Force CPU variant (skip GPU detection)
curl -fsSL https://raw.githubusercontent.com/rcspam/dictee/master/install.sh | bash -s -- --cpu

# Force CUDA variant
curl -fsSL https://raw.githubusercontent.com/rcspam/dictee/master/install.sh | bash -s -- --gpu

# Pin a specific version
curl -fsSL https://raw.githubusercontent.com/rcspam/dictee/master/install.sh | bash -s -- --version 1.3.3

# Non-interactive (no prompts, auto-detect everything)
curl -fsSL https://raw.githubusercontent.com/rcspam/dictee/master/install.sh | bash -s -- --non-interactive

Supported distributions

Distribution Mode Package type Notes
Ubuntu 22.04 / 24.04 / 24.10 Native .deb Fully tested
Debian 12+ Native .deb Fully tested
Fedora 40+ Native .rpm Fully tested (VM Fedora CI)
openSUSE Tumbleweed Native .rpm Community tested
Arch Linux Native AUR (source) makepkg -si
Anything else Fallback .tar.gz Portable tarball

CPU vs CUDA

Two packages are published for each format:

  • dictee-cpu — runs inference on CPU. Recommended for laptops without NVIDIA GPU, older GPUs (compute capability < 6.0), or AMD/Intel GPUs.
  • dictee-cuda — runs inference on NVIDIA GPU via CUDA 12. Requires NVIDIA driver 535+ and a GPU with compute capability ≥ 6.0 (Pascal or newer).

The CUDA package bundles CUDA runtime libraries internally (via a pip venv since v1.3) so you only need the NVIDIA driver + libcudnn9 — no need to install the full CUDA toolkit. See GPU-Setup for distro-specific prerequisites.

Both packages can be installed, but only one at a time is active. The dictee-switch-backend tool activates the right systemd service.

Benchmark — Parakeet-TDT warm latency

Same utterance, same model, measured on a TUXEDO InfinityBook Pro Gen8 (MK2) — Intel Core i7-13700H vs RTX 4070 Laptop 8 GB, TUXEDO OS (kernel 6.17, NVIDIA 590.48.01):

Audio length CPU (i7-13700H) CUDA (RTX 4070) Speedup
3 s 0.4 s 0.09 s 4.4×
5 s 0.8 s 0.16 s 5.0×
30 s 4.7 s 0.9 s 5.2×
1 min 9.1 s 1.8 s 5.1×
5 min 43 s 8.9 s 4.8×

CUDA is ~5× faster across all audio durations. For sub-5-second dictations (typical interactive use), latency difference is ~600 ms — borderline perceptible. For long-form transcription (meetings, podcasts), CUDA is essentially mandatory.

Manual install per distribution

Download packages from the Releases page.

Ubuntu / Debian

CPU variant:

sudo apt install ./dictee-cpu_1.3.3_amd64.deb

CUDA variant — requires the NVIDIA CUDA APT repository first:

# 1) Add the NVIDIA CUDA repository (one-time setup; see GPU-Setup for other distro versions)
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt update

# 2) Install dictee (all dependencies resolved automatically)
sudo apt install ./dictee-cuda_1.3.3_amd64.deb

Replace ubuntu2404 with your distribution version (ubuntu2204, ubuntu2410…). Full matrix at developer.download.nvidia.com/compute/cuda/repos.

Fedora / openSUSE

CPU variant:

sudo dnf install ./dictee-cpu-1.3.3-1.x86_64.rpm

CUDA variant — add the NVIDIA CUDA DNF repository first:

# Fedora 41 example
sudo dnf config-manager addrepo --from-repofile=\
  https://developer.download.nvidia.com/compute/cuda/repos/fedora41/x86_64/cuda-fedora41.repo

sudo dnf install ./dictee-cuda-1.3.3-1.x86_64.rpm

Arch Linux (AUR)

Two PKGBUILD files are maintained at the repo root: PKGBUILD (CPU) and PKGBUILD-cuda (CUDA, with bundled ONNX Runtime + nvidia-*-cu12 wheels). Both work on x86_64 and aarch64.

CPU variant:

git clone https://github.com/rcspam/dictee.git
cd dictee
makepkg -si

CUDA variant:

git clone https://github.com/rcspam/dictee.git
cd dictee
makepkg -p PKGBUILD-cuda -si

The CUDA PKGBUILD-cuda declares the ONNX Runtime tarball in source=() (downloaded by makepkg), bundles libonnxruntime{,_providers_cuda,_providers_shared}.so, and runs pip install nvidia-*-cu12 into /opt/dictee/cuda-venv in post_install — symmetric pre_remove. You only need the NVIDIA driver on the host; no system-wide cuda / cudnn package is required.

Both PKGBUILDs ship options=('!lto') and this is REQUIRED on Arch. /etc/makepkg.conf enables LTO globally, and combined with onig_sys's static libonig.a the linker drops sections and fails with undefined reference to onig_*. Removing !lto will reproduce the bug.

Portable tarball

Works on any glibc-based Linux (Slackware, Void, Gentoo, NixOS via nix-ld, etc.):

tar xzf dictee-1.3.3_amd64.tar.gz
cd dictee-1.3.3
sudo ./install.sh

The tarball bundles the Rust binaries, the plasmoid, systemd user units, and all Python UI dependencies. It installs to /usr/bin/, /usr/share/dictee/, and /usr/lib/systemd/user/.

From source

git clone https://github.com/rcspam/dictee.git
cd dictee

# CPU-only
cargo build --release

# CUDA + diarization
cargo build --release --features "cuda,sortformer"

sudo ./install.sh

See Developer-Guide for the full list of Cargo features (coreml, tensorrt, openvino, webgpu…) and the build-deb.sh / build-rpm.sh scripts used to produce the official packages.

ARM64 / Jetson

No pre-built ARM64 package is shipped — build from source on the target device.

Platform Supported Notes
NVIDIA Jetson (Orin, Xavier, Nano) ✅ CUDA Use JetPack 6+; build with --features cuda,sortformer
Raspberry Pi 4/5 ⚠ CPU only Vosk recommended for low RAM (~50 MB model)
Apple Silicon (via Asahi Linux) ⚠ CPU only No GPU acceleration; coreml feature is macOS-only
Any other aarch64 ⚠ CPU only Portable tarball path

Post-install verification

After installation, verify the setup:

# Binary present
which dictee
# → /usr/bin/dictee

# Daemon service enabled (should show 'active')
systemctl --user status dictee

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

Launch the configuration

dictee doesn't auto-open on install. Your next step is to launch dictee-setup — on first run it opens the guided Setup Wizard that walks you through ASR backend, shortcuts, translation, microphone, etc., and writes ~/.config/dictee/dictee.conf.

Four ways to launch it

Method Command / Action
Terminal dictee-setup (or dictee-setup --wizard to force the wizard flow)
App launcher Search your desktop menu for « Dictee Configuration » (Settings · Accessibility)
Tray icon Right-click the dictee tray icon → Configure Dictée
Plasmoid widget Right-click the dictee plasmoid → Configure (KDE Plasma 6 only)

The first launch (no existing dictee.conf) opens the 8-step wizard automatically. Once the wizard is complete, the same command re-opens the classic sidebar UI — see the full tour: Configuration.

After completing the wizard, bind your global keyboard shortcut and trigger a test transcription:

# Test transcription (needs a working microphone, will pop a notification
# "dictee started" then transcribe for your default duration)
dictee

Uninstall

One-liner uninstall (recommended)

The universal uninstaller auto-detects how dictee was installed (.deb / .rpm / pacman / tarball) and removes each layer. User configs and models are preserved unless --purge is passed:

curl -fsSL https://raw.githubusercontent.com/rcspam/dictee/master/uninstall.sh | bash

Options (after --):

# Also remove configs + models (complete wipe)
curl -fsSL https://raw.githubusercontent.com/rcspam/dictee/master/uninstall.sh | bash -s -- --purge

# Skip the interactive model-removal prompt (keep models, remove app only)
curl -fsSL https://raw.githubusercontent.com/rcspam/dictee/master/uninstall.sh | bash -s -- --keep-models

# Non-interactive (assume defaults, no prompts)
curl -fsSL https://raw.githubusercontent.com/rcspam/dictee/master/uninstall.sh | bash -s -- --non-interactive

Manual uninstall per distribution

Debian / Ubuntu:

sudo apt remove --purge dictee-cpu dictee-cuda dictee-plasmoid

Fedora / openSUSE:

sudo dnf remove dictee-cpu dictee-cuda dictee-plasmoid

Arch Linux:

sudo pacman -Rns dictee

Tarball / from-source install:

sudo /usr/share/dictee/uninstall.sh

Package uninstall leaves your user configuration (~/.config/dictee/, custom rules, dictionary) and model cache (~/.cache/dictee/ or /usr/share/dictee/tdt/) intact. To remove them:

rm -rf ~/.config/dictee/ ~/.cache/dictee/
sudo rm -rf /usr/share/dictee/tdt/ /usr/share/dictee/canary/

Next steps

📖 dictee Wiki

🇬🇧 Home · 🇫🇷 Accueil

Getting started / Premiers pas

Speech recognition / ASR

Translation / Traduction

Post-processing / Post-traitement

CLI

Reference / Référence

🏠 Repo · 📦 Releases · 🐛 Issues

Clone this wiki locally