Аннотация
This paper presents a scalable edge-based multilingual speech processing system designed for secure communication environments with limited or intermittent network access. The proposed architecture integrates speech enhancement, language identification, streaming automatic speech recognition, confidence-aware re-ranking, and secure local output delivery within a unified on-premise framework. Experimental evaluation was conducted using a multilingual corpus containing clean speech, command phrases, and degraded radio-channel recordings in Kazakh, Russian, English, and Turkish. Results of the experiments indicate that practical deployment performance depends not only on recognition accuracy, but also on latency stability, thermal resilience, queue management, and robustness to mixed-language noisy speech.