missuo/koe

What it solves

Koe is a lightweight, background-first voice input tool for macOS that replaces clunky or bloated dictation apps. It allows users to speak and have their words automatically corrected by an LLM and pasted directly into any active application, minimizing the friction between speaking and typing.

How it works

1. Trigger: The user activates recording via a configurable hotkey (e.g., Fn key) using hold, toggle, or double-tap modes.
2. Transcription: Audio is streamed in real-time to an ASR (Automatic Speech Recognition) provider. Supported providers include cloud services (Doubao, Qwen) and local options (Apple Speech, MLX, sherpa-onnx).
3. Correction: The resulting transcript is sent to an LLM (such as OpenAI, Anthropic, or local MLX models) to fix capitalization, spacing, and terminology based on a customizable system prompt and a user-defined dictionary.
4. Output: The final corrected text is automatically pasted into the active input field using macOS accessibility APIs.

Who it’s for

• macOS users (Apple Silicon) who want a fast, minimal voice-to-text workflow.
• Power users who prefer plain-text configuration files over complex GUIs.
• Developers or technical users who need high accuracy for domain-specific terms via a custom dictionary.

Highlights

• Minimal Footprint: Stays under 15 MB on disk and typically uses around 20 MB of RAM.
• Hybrid ASR/LLM: Supports both cloud-based and fully offline on-device inference via MLX.
• Lighweight Architecture: Built with Rust for performance and Objective-C for native macOS integration, avoiding the "Chromium tax" of Electron apps.
• Plain-Text Config: All settings, including the dictionary and prompts, are stored in ~/.koe/ for easy editing and scripting.