Improving Latency and Stability in Edge-Based Voice Assistants Through Memory and Scheduling Optimization

Intelligent voice assistants are now widely used on smartphones and embedded boards, where short response time and stable operation are essential. Heavy computation and limited hardware, however, constrain efficiency. This study tested a dual-path method that applied fine-grained memory control together with asynchronous scheduling. A total of 110 trials were run in both laboratory and office conditions. Results showed that median latency fell by 37.3% and 95th percentile latency by 39.8%. Jitter was reduced by 24.6%, and timeout events dropped by 74% compared with baseline runs. Accuracy remained stable, with word error rate changes not exceeding 0.2 and F1 score changes not exceeding 0.3. The results indicate that combining algorithm-level and system-level methods gives stronger benefits than using them alone. The study also reports jitter and timeout metrics, which are often not considered in related work. These findings suggest that dual-path optimization can support efficient and reliable deployment of voice assistants on edge devices. The main limits are the small number of device types, short test periods, and the use of only English speech. Future work should extend to multilingual datasets, longer trials, and secure execution tests.