Energy Measurement and Secure Low-Power Optimization for Speech Systems on Intelligent Terminals

Speech algorithms on intelligent terminals face strict limits of power, memory, and heat, but most earlier work has focused on model size or FLOPs rather than real energy. This study measured the energy use of automatic speech recognition (ASR) and keyword spotting (KWS) on smartphones, embedded boards, and wearables. In total, 108–120 runs were carried out under controlled and office conditions. Methods included quantization, structured pruning, frame-rate subsampling, and runtime scheduling with dynamic voltage and frequency scaling (DVFS) and NPU offloading. The results showed that energy per inference dropped by 26–33% on average, with maximum savings near 40%. Accuracy stayed stable, with WER changes ≤0.2 and KWS F1 changes ≤0.3. A regression test confirmed that parameter count and feature size were the main factors linked to energy cost. The study suggests that direct energy measurement combined with model simplification and scheduling gives a practical way to build secure and low-power speech systems. The limits are the small number of devices, short test time, and no far-field or multilingual data. These gaps should be addressed in future work.