The Xiaozhi AI project adopts a separated architecture in which the device side handles audio capture and playback while the cloud side handles ASR/LLM/TTS computation, maintaining real-time communication with cloud services via WebSocket.