ADAPTIVE PHONEME-SPECTRAL ENVELOPE RECONSTRUCTION UNDER SEVERE ACOUSTIC MASKING

Abstract

This article presents a phoneme-spectral methodology for restoring severely noise-corrupted speech while preserving parameter interpretability and a physically consistent spectral envelope under active acoustic masking. The approach unifies local peak parameters (amplitudes Aᵢ, frequencies Fᵢ, widths σᵢ) with the global energy centroid μ, which stabilizes formant-contour reconstruction at low SNR and maintains the energy balance between low- and high-frequency regions. The signal is first transformed into the spectral domain using the FFT, amplitudes are normalized, and the operating frequency range is defined. A Gaussian-mixture envelope is initialized from local energy maxima and iteratively refined by minimizing the discrepancy between the spectral envelope of the /ž/ phoneme and its physical amplitude spectrum. Noise suppression is performed using a discrete wavelet transform (sym8, L=3) with a threshold λ = κσₙ√(2lnN), which preserves peak shapes and minimizes reconstruction artifacts. Spline smoothness is governed by a frequency-dependent logarithmic law s(f)=a·log₁₀(1+b·f), which increases node density at high frequencies and enables accurate approximation of rapid spectral variations without overfitting noise. The combined σᵢ×f(μ) model adapts local widths using the global energy centroid, improving the consistency between low- and high-frequency regions of the envelope. The procedure is compatible with psychoacoustic frequency scales (Bark/Mel) and objective intelligibility metrics (STI), enabling analytical interpretation of parameter contributions to perceived quality. Experiments on the Ukrainian fricative phoneme /ž/ demonstrate RMSE ≈ 4% at SNR ≈ −6 dB in the physical amplitude domain, with stable μ values under threshold variation. The proposed methodology is applicable to room-acoustics assessment under active masking conditions and to security-oriented analyses of speech-channel informativeness. Transparent validation stages (reference-formant comparison, μ-balance inspection, and evaluation of σᵢ-, μ-, and combined models) support its transferability to other phoneme classes without loss of interpretability.