My first attempt at building a fully local voice assistant, everything runs on your own machine. Say a wake word, talk to it, and it talks back. Simple as that.
It uses Faster Whisper for speech-to-text, a small local language model (Qwen3-1.7B) to generate responses, and Kokoro-82M to speak them out loud. The whole pipeline stays on-device, so nothing leaves your computer.
How It Works
Wake word detection — always listening for the trigger phrase, then starts recording
Voice activity detection — knows when you've stopped talking so it can respond
Streaming responses — starts generating and speaking before the full response is ready, so it feels snappy
Conversation memory — remembers what you said earlier in the conversation
Auto sleep — goes quiet after a period of inactivity
This was mostly a learning project to explore how far you can push local-only AI.
