A local-first, Hebrew-first voice assistant for your desktop. Speak Hebrew, see it typed anywhere. Speak a command, watch it run. Local by default — cloud providers are opt-in, always.
Status:
v0.3.0— an early dictation preview. Hebrew speech-to-text with system-wide paste works, alongside a first-run tutorial and a settings hub (Hebrew/English interface, rebindable hotkey); command and chat modes are not built yet. Published as an unsigned pre-release. See the roadmap below.
Lashon (לָשׁוֹן — "tongue / language") turns speech into text and action on your own machine, in three modes:
- Dictation — hold a hotkey, speak Hebrew, release; the text appears in the focused app with correct right-to-left ordering.
- Command — speak a natural-language command; Lashon operates your PC.
- Chat — ask a question; Lashon answers, by voice.
Speech recognition, language models, and speech synthesis all run locally by default. Cloud providers exist only as opt-in adapters, each marked with a clear "cloud" badge. No transcripts, audio, or telemetry leave the machine without explicit consent.
Download Lashon-0.3.0-windows-x64-setup.exe (~66 MB) from the
Releases page and run it.
This pre-release is not code-signed. Windows SmartScreen will show "Windows protected your PC" on first run — click More info → Run anyway. A signed installer is the next release.
On first run Lashon downloads the ~1.6 GB Hebrew speech model; on an NVIDIA GPU it also fetches the CUDA runtime for faster transcription. The tongue shows a dim "preparing" state until it is ready; after that it works offline. Press Ctrl+Space, speak Hebrew, then pause — the text is pasted into the focused app. Press Ctrl+Shift+D for a debug surface that reports the STT sidecar.
Prerequisites: Rust 1.95, Node 20+, Python 3.11–3.12, and a WebView2 runtime (Windows; bundled by the OS on Windows 11).
# desktop app (Tauri 2 + SvelteKit 5)
cd apps/desktop
npm install
npm run tauri devTo build the installers yourself, see
docs/packaging-windows.md.
See CONTRIBUTING.md for the development workflow and
docs/architecture.md for the system design.
Lashon is built in three phases:
- Dictation — Hebrew speech-to-text with system-wide injection. (current)
- PC operation — voice-driven command mode, plus delegation to external coding agents.
- Voice response — Hebrew-perfect text-to-speech for confirmations and chat.
Milestones M0–M3 (bootstrap, STT pipeline, hotkey + injection, tongue UI) are
done; M4 (onboarding + settings) is in progress, and the v0.3.0 dictation
preview — with a first-run tutorial and a settings hub — is published. Next is
the rest of Phase 1.
The full roadmap — scope, all fourteen milestones, and per-phase workstreams —
lives in docs/roadmap.md. Active work is tracked as stories
in docs/stories/.
MIT © Lashon contributors.
Bundled and optional third-party components retain their own licenses; see
NOTICE. Only MIT/Apache-licensed models ship in the installer;
non-commercially-licensed models are surfaced as clearly-badged opt-in
downloads, never bundled.