In development · Swift, MLX, Kokoro, Whisper/Parakeet, Apple Foundation Models, MCP
VOX is a native macOS menu-bar app that runs a full voice assistant on-device: speech recognition, language model, and text-to-speech, all local by default. Nothing leaves the machine unless a web search is explicitly triggered.
Speech recognition can use Apple's built-in SpeechTranscriber, Parakeet via FluidAudio, or Whisper via WhisperKit. The language model layer supports MLX (any mlx-community model), Apple's Foundation Models, or any OpenAI-compatible API endpoint. Text-to-speech uses Kokoro by default, with Apple's AVSpeechSynthesizer as a fallback that needs no download.
The harder engineering problem is turn-taking — knowing when someone has actually finished speaking versus just paused. This combines voice-activity detection (Silero VAD), a semantic end-of-turn model, and echo cancellation, so the app doesn't mistake its own voice for the user's mid-reply.
It also handles file and image attachments (OCR for images, text extraction for PDFs and docs) and can connect to external tools over MCP (Model Context Protocol).
Still in active development, and the native Swift successor to an earlier Python prototype.