Secure Speech Communication via Syllable-Level Encryption and PUFBased Key Management in Edge Computing

Abstract

Recent studies indicate that over 60% of voice communications transmitted over public or wireless channels are vulnerable to interception, while speech transformation systems can reduce intelligibility leakage by more than 70% under noisy and compressed channels. Traditional secure speech systems rely on bitstream-level encryption or fixed cryptographic keys, which leads to high computational overhead, vulnerability to long-term key compromise, and poor robustness under channel noise and compression. Moreover, software pipelines introduce excessive latency and power consumption, limiting their feasibility for VLSI and embedded deployments. To address these challenges, this work proposes a syllable encryption–based secure speech architecture integrating on-chip neural accelerators, dynamic key rotation, and hardware-based Physical Unclonable Function (PUF) key generation. The system converts input speech into syllable sequences using an on-chip Whisper ASR accelerator, applies hardware-efficient syllable mapping, and performs encryption using dynamically rotated session keys derived from an intrinsic PUF, eliminating stored secrets and enhancing physical security. The encrypted syllables are resynthesized using a StyleTTS2 accelerator with reference speech to preserve speaker identity while ensuring unintelligibility. At the receiver, a symmetric PUF-enabled decryption pipeline reconstructs intelligible speech only on authorized hardware. This VLSI-oriented design achieves low latency, reduced power consumption, strong resistance to key compromise, and robustness to noisy voice channels, making it highly suitable for real-time secure speech communication in next-generation edge and embedded systems.