Invertible generative speech hiding with normalizing flow for secure IoT voice

Speech-based control is widely used for remotely operating the Internet of Things (IoT) devices, but it risks eavesdropping and cyberattacks. Speech hiding enhances security by embedding secret speech in a cover speech to conceal communication behavior. However, existing methods are limited by the extracted secret speech’s poor intelligibility and the stego speech’s insufficient security. To address these challenges, we propose a novel invertible generative speech hiding framework that integrates the embedding process into the speech synthesis pipeline. Our method establishes a bijective mapping between secret speech inputs and stego speech outputs, conditioned on text-derived Mel-spectrograms. The embedding process employs a normalizing flow-based SecFlow module to map secret speech into Gaussian-distributed latent codes, which are subsequently synthesized into stego speech through a flow-based vocoder. Crucially, the invertibility of both SecFlow and the vocoder enables precise secret speech extraction during extraction. Extensive evaluation demonstrated the generated stego speech achieves high quality with a Perceived Evaluation of Speech Quality (PESQ) score of 3.40 and a Short-Term Objective Intelligibility (STOI) score of 0.96. Extracted secret speech exhibits high quality and intelligibility with a character error rate (CER) of 0.021. In addition, the latent codes of secret speech mapped and randomly sampled Gaussian noise are very close to each other, effectively guaranteeing security. The framework achieves real-time performance with 1.28s generation latency for 2.22s speech segment embedding(achieving a real-time factor (RTF) of 0.577), which ensures efficient covert communication for latency-sensitive IoT applications.