This paper presents the embedded deployment of a fully offline voice assistant leveraging open-source C++ implementations of Large Language Models (LLMs). Our system integrates Whisper.cpp for automatic speech recognition, LLaMA.cpp for natural language understanding and dialog management, and Piper for speech synthesis. All components are optimized for embedded execution and deployed on a Raspberry Pi 5. We demonstrate that the proposed toolchain achieves an end-to-end runtime latency of under one second with a compact architecture, eliminating the need for separate modules or lookup tables. Benchmarks on a common command show that total latency remains under one second, with energy consumption suitable for mid-tier edge devices. Compared to prior multi-stage systems requiring high-end GPUs and Python-based stacks, our approach reduces initialization time by over 60% and achieves faster synthesis. Despite the LLM’s higher inference latency relative to rule-based dialog management systems, it enables more flexible and natural responses. This work highlights the feasibility of LLM-based conversational agents for embedded applications, offering enhanced privacy, ease of deployment, and enhanced generalization capabilities.