What is vui?

Vui, pronounced "vooey", is a real-time voice assistant you run on your own computer. You speak into a microphone, it transcribes your speech, sends the text to a local language model, and streams an audio reply back, all without sending your conversation to a cloud service. The pipeline works in three steps. First, automatic speech recognition converts your voice to text. Second, a local language model generates a response. Third, text-to-speech converts that response back into audio. Each step is pluggable, so you can swap out different software for each stage. The default speech recognition uses faster-whisper or Moonshine, the language model connects to Ollama or other compatible local LLM software, and the text-to-speech uses Vui Nano, a 300 million parameter model trained on conversational speech that includes natural sounds like breaths and hesitations. Key features include barge-in, which lets you interrupt the assistant mid-reply and it stops to listen, voice cloning by uploading a sample audio clip to make the assistant sound like a specific voice, persistent memory across sessions, and a set of built-in tools like web search and timers. There is also an optional sidecar process that handles longer tasks like reading email or calendar events. The API is compatible with OpenAI's Realtime API format, so existing apps built for that standard can point to Vui instead. It runs on a local GPU and is designed for developers who want a self-hosted, privacy-preserving voice assistant they fully control. A one-line installer sets up the full stack, and a Docker compose file can bring up the streaming server, an optional bundled Ollama, and the optional task server together. The full README is longer than what was shown.