Vox

Voice input reimagined — speak in any language, type in any language.

A macOS menu-bar voice input app built with Makepad and Rust.

Press Option, speak, release — your words appear wherever the cursor is.

Platform: macOS (Apple Silicon) only. See Roadmap for cross-platform plans.

Features

Press-to-talk — Hold Option key to record, release to transcribe and inject text
High-quality ASR — Powered by Qwen3-ASR (CER 5.88 on Chinese, 30+ languages)
LLM Refinement — Optional post-processing to fix recognition errors, translate, or convert text style
Real-time translation — Speak Chinese, output English (or any supported language)
Classical Chinese mode — Speak modern Chinese, output 文言文
Transparent capsule UI — Floating status indicator with pulse animation, fully transparent background
Multi-language — Chinese, English, Japanese, Korean, Traditional Chinese, Classical Chinese
Menu bar app — Lives in your menu bar, no Dock icon (when bundled)
Privacy first — Audio processed locally via OminiX-MLX, LLM refinement configurable (local or cloud)

Architecture

┌─────────────────────────────────────┐
│            Vox (Makepad 2.0)        │
│                                     │
│  Menu Bar ←→ Capsule ←→ Settings   │
│       ↕           ↕         ↕       │
│   CGEvent    Audio I/O    Config    │
│     Tap      (16kHz)    (~/.config) │
└──────┬────────────┬─────────────────┘
       │            │
       ▼            ▼
  macos-sys     ominix-api
  (ObjC FFI)   (HTTP :18080)
                    │
              ┌─────┴─────┐
              │ Qwen3-ASR │  ← local MLX inference
              │ LLM (opt) │  ← local or cloud API
              └───────────┘

Crate	Purpose
`macos-sys`	macOS FFI — CGEvent tap, NSStatusBar, clipboard, input source, key simulation
`app` (vox)	Makepad 2.0 UI — capsule window, settings, audio capture, HTTP client

Quick Start

Prerequisites

macOS 14.0+ on Apple Silicon (M1/M2/M3/M4)
Rust 1.82+
OminiX-API running locally
Accessibility permission for the terminal (System Settings → Privacy → Accessibility)

1. Start the ASR service

cd /path/to/OminiX-API
PORT=18080 ASR_MODEL_DIR=~/.OminiX/models/qwen3-asr-0.6b cargo run --release

2. Run Vox

cd /path/to/vox
cargo run -p vox

3. Use it

Look for MIC in the menu bar
Hold left Option key and speak
Release — text appears at your cursor

With LLM Refinement (translation/correction)

MOONSHOT_API_KEY=sk-your-key cargo run -p vox

Or configure any OpenAI-compatible API in Settings (MIC → LLM Refinement → Settings).

Language Modes

Menu Selection	ASR Language	LLM Action	Example
简体中文	Chinese	Correct typos	配森 → Python
English	Chinese/English	Translate to English	你好 → Hello
繁體中文	Chinese	Convert to Traditional	语音输入 → 語音輸入
日本語	Japanese	Preserve Japanese	そのまま出力
한국어	Korean	Preserve Korean	그대로 출력
文言文	Chinese	Convert to Classical	今天天气好 → 今日天朗气清

Translation and style conversion require LLM Refinement enabled (needs API key).

Configuration

Config file: ~/.config/vox/config.json

Environment Variables

Variable	Description
`MOONSHOT_API_KEY`	Kimi API key (auto-configures URL and model)
`VOICE_INPUT_LLM_API_KEY`	Any OpenAI-compatible API key
`VOICE_INPUT_LLM_API_URL`	LLM API base URL
`VOICE_INPUT_LLM_MODEL`	LLM model name
`VOICE_INPUT_API_URL`	OminiX-API base URL (default: `http://localhost:18080`)

Build

# Development
cargo run -p vox

# Release
cargo build --release -p vox

# macOS .app bundle (with LSUIElement — no Dock icon)
make bundle
# Output: target/Vox.app

Tech Stack

UI Framework: Makepad 2.0 — GPU-accelerated native UI with Splash DSL
ASR Engine: Qwen3-ASR via OminiX-API — pure Rust MLX inference
LLM: Any OpenAI-compatible API (Kimi, DeepSeek, OpenAI, local models)
macOS Integration: Raw ObjC FFI via makepad_objc_sys — CGEvent tap, NSStatusBar, NSPasteboard, TIS input sources

Platform Support

Platform	Status	Notes
macOS (Apple Silicon)	Supported	Full functionality
macOS (Intel)	Untested	Makepad supports it, but OminiX-MLX requires Apple Silicon
Windows	Planned	Requires platform-sys crate for hotkey/tray/clipboard
Linux	Planned	Requires platform-sys crate + alternative ASR backend
Web/WASM	Not planned	Global hotkey and clipboard injection not possible

Roadmap

v0.2 — Polish & Stability

Hide capsule/settings windows on startup (proper window lifecycle)
Waveform animation driven by real-time audio RMS
Entrance/exit animations for capsule (spring + fade)
Menu bar icon using proper SF Symbol or SVG instead of text "MIC"
Clean up debug logging and clippy warnings
.app bundle with code signing

v0.3 — Cross-Platform Foundation

Extract macos-sys into platform-sys trait abstraction
- trait HotkeyMonitor — global hotkey detection
- trait SystemTray — tray icon and menu
- trait Clipboard — read/write clipboard
- trait KeyInjector — simulate keystrokes
- trait InputSourceManager — detect/switch input methods
Windows backend (windows-sys) — RegisterHotKey, Shell_NotifyIcon, SendInput
Linux backend (linux-sys) — XGrabKey, libappindicator, xdotool/wtype

v0.4 — Flexible ASR Backend

ASR backend abstraction (not tied to OminiX-MLX)
Remote API mode — use any OpenAI Whisper-compatible endpoint
Whisper.cpp local backend for non-Apple-Silicon platforms
Configurable ASR endpoint in Settings UI

v0.5 — Enhanced UX

Streaming transcription (show text as you speak)
Custom hotkey configuration UI
Auto-start on login
Update checker
Localized UI (English/Chinese/Japanese)

Future

Plugin system for custom text transformations
Voice commands (not just dictation)
Multi-model ASR (auto-select best model per language)

License

AGPL-3.0 — See LICENSE for details.

Credits

Makepad — The UI framework
OminiX-MLX — ML inference on Apple Silicon
OminiX-API — OpenAI-compatible API server

Vox