Mixed Signal Design Using C++ and DSP Acceleration for Low Latency and Secure Speech Systems

Abstract Real-time speech systems must balance fast response with reliable protection against attacks. This study introduces a mixed-signal framework that links C++ modules with DSP units, combining efficient feature extraction and instruction parsing with secure sandboxing. Tests were carried out on mobile and smart home devices under quiet, office, and street-noise settings. The system reduced response latency by about 34% compared with software-only designs and improved recognition accuracy by 2.5–4.6 percentage points. It kept latency below 110 ms in all acoustic conditions and lowered replay and injection attack success rates by 40–60% through adaptive control. An ablation study showed that both sandboxing and adaptive checks are required to keep these gains. The results highlight the novelty of joining low-latency processing with strong protection, offering practical value for safety-related speech systems, while also noting limits in scaling to larger vocabularies and energy use on resource-constrained devices.