Energy Efficient Speech Algorithms for Intelligent Terminals with Pruning and Compression

Energy consumption is a key barrier to the use of speech algorithms on intelligent terminals, as limited battery capacity restricts continuous operation. This study analyzed the main energy bottlenecks in speech processing and introduced a framework that combines pruning and compression to lower power use while keeping recognition quality stable. A dataset of 1,200 speech samples was collected under different devices and acoustic settings, and the optimized models were compared with uncompressed baselines. The results showed that average energy use dropped by 42%, while recognition accuracy decreased by less than 0.3%. The analysis confirmed that feature extraction and model inference caused most of the energy demand, and their optimization produced the largest savings. Compared with using pruning or compression alone, the combined approach provided a better balance between efficiency and recognition accuracy. These results demonstrate a practical method for energy-aware speech systems, with applications in mobile, smart home and medical devices, although further testing on larger and more varied datasets is required.