Auriscribe

A lightweight, offline speech-to-text for Linux desktop use, built using whisper.cpp.

• Native application with no web tech or python
• Low memory usage, completely unloads the Whisper model when idle
• Vulkan backend by default for acceleration without pulling in lots of dependencies
• Optional voice activity animation that can be centered on the screen or over the window recieving input

Install

Debian/Ubuntu (.deb)

Download auriscribe_*.deb from the GitHub Releases page:

sudo dpkg -i ./auriscribe_*.deb
sudo apt-get -f install

Fedora/RHEL (.rpm)

Download auriscribe-*.rpm from the GitHub Releases page:

sudo dnf install ./auriscribe-*.rpm

Arch Linux (AUR)

• Package: https://aur.archlinux.org/packages/auriscribe

# Using an AUR helper
yay -S auriscribe

# Manual
git clone https://aur.archlinux.org/auriscribe.git
cd auriscribe
makepkg -si

Building instructions

# Install build dependencies (Debian/Ubuntu helper)
make deps

# Build
make

# Run
./auriscribe

make builds the vendored libs/whisper.cpp automatically (Vulkan-enabled when available). ./scripts/setup-whisper.sh is optional and only needed if you want to rebuild libs/whisper.cpp by itself.

Download a Model

Use the tray menu: Download Models... → choose a model → Download. Models are saved under ~/.local/share/auriscribe/models/ and the downloaded model is auto-selected.

Usage

• Use the tray menu Start/Stop Recording.
• Configure a global hotkey in Settings... (some desktop environments reserve Super shortcuts).
• On stop, Auriscribe transcribes and inserts the text into the window that was active when recording started (X11).

Autostart: enable Start Auriscribe on login in Settings.

Dependencies

• GTK3
• libayatana-appindicator3
• PulseAudio
• json-c
• libcurl
• X11 (for global hotkeys on X11)
• xdotool (for text input on X11)
• Optional (faster Whisper): Vulkan dev/runtime (e.g. libvulkan-dev)

Performance knobs

• AURISCRIBE_VULKAN=0 disables the Vulkan backend build (not recommended)
• AURISCRIBE_REQUIRE_VULKAN=0 allows CPU-only builds (not recommended)
• AURISCRIBE_NO_GPU=1 forces CPU at runtime
• AURISCRIBE_GPU_DEVICE=0 selects GPU device index
• AURISCRIBE_VULKAN_WARMUP=0 disables one-time Vulkan shader warmup on app startup
• AURISCRIBE_THREADS=8 sets Whisper CPU thread count
• AURISCRIBE_HF_REPO=ggerganov/whisper.cpp overrides the Hugging Face model repo
• AURISCRIBE_VK_ICD_FILENAMES=/path/to/icd.json limits Vulkan ICD probing (can reduce one-time RAM overhead)

Legacy env vars (XFCE_WHISPER_*) are still accepted for backwards compatibility.

Whisper initial prompt

In Settings... you can optionally set an Initial prompt (max 244 chars). This is passed to Whisper as an “initial prompt” to bias decoding (useful for names/jargon and consistent formatting).

Examples:

• Vocabulary / proper nouns: Auriscribe, whisper.cpp, Vulkan, GGML, PipeWire, PulseAudio
• Formatting: Use proper punctuation. Use sentence case. Prefer numerals for numbers.
• Domain bias: Meeting notes. Action items. Technical discussion about Linux and GPUs.

Other Useful Projects

• Try my AI panel plugin for XFCE XFCE Ask
• For a feature complete AI application try out ChatGTK
• A Markdown note taking application for the system tray TrayMD