I built a local voice AI realtime music setup where my ESP32 microcontroller talks to my MacBook over WebSockets. The microcontroller is just a tiny Arduino-based device with a mic and speaker, and the MacBook M4 Pro runs Magenta Realtime 2 locally and streams the audio back to the device. The fun part is that it’s agentic and conversational. So I can tap the ESP32, speak into it, and it uses MLX Whisper to transcribe what I said. Then after detecting VAD, it sends that to a Qwen model, which ...