Arion
Local-first voice dictation and Pinyin input for macOS and Windows, built with MyGo native Go UI.
English · 简体中文 · 日本語
Overview
Cloud dictation tools require sending raw microphone audio and window context to remote servers, while built-in OS offline dictation struggles with mid-sentence Chinese/English code-switching, domain terminology, and application-specific formatting.
AstraFoundry built Arion as an open-source, local-first, self-hostable voice and Pinyin input method. Speech recognition, offline full-Pinyin typing, correction learning, and contextual polishing run on your device while keeping raw audio and encryption keys on your machine.
Features
-
Hold-to-Talk & Native System IME — Hold the global dictation shortcut (default right ⌥ on macOS, right Alt on Windows; record another key or modifier combination in Settings, see dictation shortcuts) in any focused editor to speak Mandarin, English, or mixed technical terms (e.g. "把这个 PR merge 到 main 之前，先跑一下 CI"), and release to insert once; tap once to keep listening until the next press (Esc cancels). Also includes native system IME hosts (macOS ArionInput.app and Windows ArionTSF.dll) with offline Simplified Chinese full-Pinyin typing and Shift language toggling.
-
On-Device Speech Recognition — Runs Qwen3-ASR 1.7B locally via Swift MLX (8-bit) on Apple silicon and speech.cpp (Q8_0, CPU/Vulkan) on Windows x64 (ADR 020). After the model bundle is downloaded and SHA-256 verified, dictation runs completely offline. Microphone audio stays in memory only during the active session, password fields are skipped automatically, and stepped software input gain (−6 dB to +12 dB) supports quiet dictation at shared desks.
-
Edit Learning & Draft Isolation — Streaming partial drafts stay inside Arion's floating pill and are committed to the editor only on release. Deterministic cleanup rules and user-approved vocabulary are applied locally; on macOS, manual edits made to inserted text in native editors are learned automatically with one-click undo (ADR 014).
-
Two-Stage Contextual Polishing — Optional polishing runs after key release (ADR 023) to strip spoken fillers, resolve mid-sentence self-corrections, disambiguate terms against visible window/cursor context (ADR 013), and format text for the active app (Match the app, ADR 012). An optional second shortcut dictates and translates into one of 12 target languages. Runs Qwen3.5-2B (Q4_K_M GGUF via llama.cpp) locally on macOS Apple silicon and Windows x64 (ADR 018, ADR 025), or connects to an explicit opt-in cloud model (ChatGPT sign-in or OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter API key) while keeping raw audio on-device (ADR 022).
-
Native Go Desktop Shell — Built with MyGo (v0.4.6) using Feature-Sliced Design in Go without Electron or a webview (ADR 001, ADR 007). Runs in the menu bar / system tray with a click-through floating pill that joins fullscreen Spaces without stealing focus, plus live-switched English and Simplified Chinese UI (i18n).
-
Zero-Knowledge Sync & Signed Updates — Personal vocabulary is encrypted locally with AES-256-GCM using keys stored in macOS Keychain or Windows DPAPI (ADR 005). The repository includes services/sync (a Docker self-hostable Go ciphertext relay that never holds decryption keys, ADR 015) and services/updates (Ed25519-signed stable and beta update channels, ADR 019).
-
Controlled Updates and Product Notices — Automatic checks and verified downloads have separate controls. Downloaded updates wait for an explicit safe restart. OS notification settings control reminders and automatic signed-in account notices, with running-app reception and next-launch catchup. See update lifecycle and product notifications.
Platform Support
| Platform | Architecture | Status |
|---|
| macOS | Apple silicon (arm64) | Working menu bar dictation app, native MLX ASR, local GGUF polishing, and ArionInput.app IMK host (Intel macOS is unsupported, ADR 021) |
| Windows | 64-bit (x64 / amd64) | Early preview with native speech.cpp ASR, local GGUF polishing, and ArionTSF.dll IME host |
See Release Status, Agent Handoff, and Validation Report for current milestone progress and open release gates.
Quick Start
Prerequisites
Build and Run
make install # install the pinned pnpm workspace (--frozen-lockfile)
make check # docs, JSON, Go layer boundaries, workflow lint, and core/sync/updates/i18n/public driver tests
make desktop-check # desktop tests, keyboard checks, and go vet (cgo)
make desktop-dev # stage native runtimes and launch Arion Dev in dev mode
make build desktop # package Arion.app + DMG (macOS) or NSIS installer (Windows) under apps/desktop/build/
make e2e # native window lifecycle E2E on a real desktop session
make e2e-macos # owned product scenarios against the packaged Arion.app
make site-check # marketing site (apps/site): i18n, types, lint, tests, static export
On first launch, make desktop-dev opens onboarding and provisions its own development profile. On macOS, configure a persistent signing certificate and quit the release before starting development. Use make desktop-status to check the actual running binary and make desktop-production to return to the installed release. Setup, profile paths, permission checks, and IME restrictions are in Local Development (docs/operations/local-dev.md).
Documentation
Repository ownership and internal evidence are documented in the two-repository guide.
Contributors
Star History
License
Arion is developed by AstraFoundry and licensed under AGPL-3.0 (ADR 009). Follow @AstraFoundry on X, join the community on Discord or Feishu, and report bugs in GitHub Issues.
