A Low-Power, Self-Sustained Acoustic Node for Always-On Voice-Command Detection and Recognition Based on Triboelectric Nanogenerators and Edge Artificial Intelligence

Abstract Always-on voice-command recognition is important for smart-home control and human–machine interaction, but continuous acoustic sensing, edge inference, and wireless transmission can rapidly drain batteries. Here, we report a low-power, self-sustained acoustic node for always-on voice-command detection and recognition that integrates triboelectric acoustic wake-up and sensing, hierarchical edge artificial intelligence (edge AI), on-demand Bluetooth transmission, photovoltaic energy harvesting and energy storage. Rather than keeping all modules continuously active, the system reduces energy consumption through module-level power optimization and system-level event-driven control of module operating time. It remains in near-zero-power monitoring until acoustic activity crosses a preset threshold, then enters a low-power stage for wake-word screening. Command recognition and Bluetooth transmission are activated only after wake-word confirmation. The system achieved real-time accuracies of 96.0% for wake-word screening and 95.64% for 11-class command-word recognition on the microcontroller. The system consumed 8.04 J over 24 h, corresponding to an average power of 93 μW. During a separate 24 h test of the complete system with direct photovoltaic charging, approximately 1.72 kJ of net electrical energy was delivered to the battery terminals. This architecture integrates always-on event capture, on-device recognition, and local energy supply, and may support the development of long-term unattended voice-interaction nodes.