NEURAL NETWORK MODEL FOR ASSESSING THE SECURITY LEVEL OF HEAVILY NOISE-MASKED SPEECH INFORMATION BASED ON THE RII STRUCTURAL FRAMEWORK

Abstract

This study addresses the problem of determining the security level of speech information under complex acoustic and vibroacoustic interference, where traditional speech-quality metrics (SNR, STI, SII, PESQ, STOI) fail to reflect the actual capability of modern reconstruction algorithms (HMM, DNN, BayesianNN, GAN) to recover the semantic content of intercepted signals. Under low signal-to-noise ratios, and even after applying active vibroacoustic masking, a substantial portion of speech structures remains recoverable, creating a risk of information leakage. The absence of a quantitative criterion that unifies the physical, linguistic, and neural-network parameters of the signal and allows evaluation of the real possibility of semantic reconstruction defines the central scientific and technical challenge of this work. To address this, an integral model is proposed for estimating the residual speech information of heavily noise-masked speech, taking into account spectral structure, contextual linguistic dependencies, and the capabilities of modern reconstruction systems. The multi-level architecture of the model includes an analytical spectral description of the signal (Aᵢ, μᵢ, σᵢ*, Z₀(f), s(f)), a Bayesian-Markov triphone structure, and a multilayer emission neural network built using CNN, MLP, and BayesianNN components. The spectral level provides a formal description of energy peaks and adaptive smoothing; the linguistic level reflects probabilistic regularities of transitions between triphones; the neural-network level integrates all feature types and models uncertainty in emission probabilities. Based on the synthesis of these levels, the residual intelligibility criterion (RII) is formulated, quantitatively characterizing the ability of a potential interceptor to recover message content after interference and filtering. A threshold value RII* is defined, interpreted as a conditional boundary between information-dangerous and information-insufficient reconstruction regimes. The proposed model can be applied in technical information-protection systems, testing laboratories, and evaluation complexes for active vibroacoustic interference. The obtained results establish scientific and technical foundations for developing instrumental methods to determine the level of residual speech informativeness and enhance its protection.