GNOME-first local dictation for Linux. Hold a key combo, speak, release — your words appear in the focused app. Everything runs locally, no cloud.

• Hold Alt+Super (configurable) to record, release to stop
• Streaming transcription via Nemotron 0.6B — text is ready the instant you stop talking
• LLM cleanup via Qwen 3.5 — fixes filler words, punctuation, capitalization, self-corrections
• Text insertion via uinput virtual keyboard — types directly into any focused app
• Window OCR context — captures screen text to help the cleanup model disambiguate names and terms
• Speaker diarization — filters out other voices (experimental)

On an Intel Core Ultra 7 155U (no GPU):

record=3.2s  transcribe=0.0s  cleanup=0.5s  insert=0.2s  total=0.7s

Transcription happens during recording (streaming). Cleanup uses a pre-warmed KV cache.

Ubuntu 25.04+ or Fedora 42+. GNOME 48+ on Wayland.

# Ubuntu
sudo apt install \
  build-essential cargo cmake \
  libadwaita-1-dev libatspi2.0-dev libgirepository1.0-dev \
  libglib2.0-dev libgtk-4-dev libgtk4-layer-shell-dev \
  libvulkan-dev libxkbcommon-dev \
  pkg-config tesseract-ocr

# Fedora
sudo dnf install \
  cargo cmake gcc gcc-c++ \
  at-spi2-core-devel glib2-devel gobject-introspection-devel \
  gtk4-devel libadwaita-devel libxkbcommon-devel vulkan-loader-devel \
  pkgconf-pkg-config tesseract

Your user must be in the input group for hotkey capture and text injection:

sudo usermod -aG input $USER
# Log out and back in

A udev rule is needed for the virtual keyboard:

echo 'KERNEL=="uinput", GROUP="input", MODE="0660"' | sudo tee /etc/udev/rules.d/99-pepper-x-uinput.rules
echo 'SUBSYSTEM=="input", ATTRS{name}=="Pepper X virtual keyboard", ENV{ID_INPUT_KEYBOARD}="1"' | sudo tee /etc/udev/rules.d/99-pepper-x-keyboard.rules
sudo udevadm control --reload-rules

cargo build --release
sudo install -m 755 target/release/pepper-x /usr/local/bin/
sudo mkdir -p /usr/libexec/pepper-x
sudo install -m 755 target/release/pepperx-uinput-helper /usr/libexec/pepper-x/
sudo install -m 755 target/release/pepperx-cleanup-helper /usr/libexec/pepper-x/
bash scripts/dev-install-extension.sh

Log out and back in for the GNOME extension to load.

Launch the app, go to the Models section, and click Download Missing Models. Or download manually:

• ASR: Nemotron 0.6B int8 (~850MB) — downloaded from HuggingFace on first run
• Cleanup: Qwen 3.5 0.8B Q4_K_M (~500MB) or 2B Q4_K_M (~1.3GB)

pepper-x

That's it. The app:

1. Starts the GNOME Shell extension (tray icon + status pill)
2. Pre-warms the cleanup model in the background
3. Listens for your trigger keys (Alt+Super by default)

The app window is organized into sections:

• Recording — Shortcut recorders (hold-to-record + toggle-to-record), mic picker, sound effects, speaker filtering, test dictation
• Cleanup — Enable/disable, window context toggle, prompt profile, custom prompt editor
• Corrections — Editable preferred transcriptions and commonly misheard replacements
• Models — ASR and cleanup model selection with download progress
• History — Transcription lab with per-stage model pickers, inline prompt editor, word-level diff, audio playback, diarization timeline
• General — Launch at login
• Diagnostics — Runtime status

# Transcribe a WAV file
pepper-x --transcribe-wav recording.wav

# Transcribe + cleanup
pepper-x --transcribe-wav-and-cleanup recording.wav

# Rerun an archived recording
pepper-x --rerun-archived-run <run-id>

• pepper-x — GTK4/libadwaita app, owns the recording pipeline, settings, history
• pepperx-cleanup-helper — Persistent subprocess running llama.cpp (llama-cpp-4) for Qwen 3.5 inference, isolated to avoid ONNX Runtime symbol collision with the ASR engine
• pepperx-uinput-helper — Persistent subprocess with XKB-aware virtual keyboard for text injection
• pepperx@obra GNOME extension — Tray icon, floating status pill overlay, D-Bus bridge

cargo test --workspace

See individual crate licenses.