Generated indicator system
Transparent PNG indicators and metadata-driven spritesheets make the desktop presence feel alive without becoming noisy.
Linux dictation, local by default
Speak into any app without handing your voice to the cloud.
Hold a hotkey, speak naturally, release. JuhSpeak transcribes locally, cleans the text, and pastes it into the focused window on Wayland or X11.
- Default engine
- faster-whisper
- Desktop paths
- Wayland + X11
- Privacy model
- Local first
Three commands
Install it like a desktop tool, not a research project.
JuhSpeak ships with a distro-aware installer, a start script that chooses the correct indicator path, and a doctor that checks the real desktop environment before you rely on dictation.
install.sh
git clone \
https://github.com/JuhLabs/JuhSpeak.git
cd JuhSpeak
./scripts/install.sh
./scripts/doctor.py
./scripts/start.sh
Speech-aware cleanup
Voice commands, snippets, dictionary corrections, backtrack handling, punctuation, capitalization, and optional LLM cleanup.
Hardware recommendations
NVIDIA, AMD, Intel, VRAM, RAM, CPU, distro, session, and paste backend detection guide setup and model choice.
Control surface
From hotkey to clean text in one local pipeline.
The generated product diagram maps the daily flow: capture audio, run the local model, apply cleanup rules, and paste into the active Linux app.
Privacy posture
Local first means local first.
Speech recognition runs locally by default. Optional cloud cleanup is bring-your-own-key and stores provider keys through the system keyring, not plaintext config.
- Audio is processed on your machine by default.
- Cloud cleanup is explicit opt-in.
- Runtime config, logs, histories, models, and caches are excluded from the public release.
v1.0.1
Native packages and hardened release checks.
The latest release includes source archives, DEB/RPM packages, release notes, security policy, CI, CodeQL, secret scanning, Dependabot, pip-audit, Bandit, and a GitHub Pages deployment workflow.