LPC AND PRONY PARAMETRIC MODELING INTEGRATED INTO THE KALMAN FILTER FOR SPEECH NOISE REDUCTION: COMPARATIVE ANALYSIS AND STATISTICAL EVALUATION

Abstract

This work presents a comparative experimental study between two parametric modeling strategies integrated into the Kalman filter for speech denoising: autoregressive models (LPC-AR) and pole–zero models estimated via the Prony method (ARMA). The experimental protocol included controlled input segmental signal-to-noise ratio (segSNR_in ≈ 3 dB), multiple stochastic noise realizations, and two noise scenarios (white and colored AR(1)). Performance was evaluated using complementary metrics: output segmental signal-to-noise ratio (segSNR_out), Itakura–Saito divergence (IS), short-time objective intelligibility (STOI), and computational time. Results were reported as mean ± standard deviation and analyzed using the Friedman test with post-hoc comparisons. The findings indicate that the Kalman+LPC approach achieved superior overall performance and greater statistical consistency, particularly under white noise conditions, while also exhibiting lower computational cost. The Kalman+Prony formulation demonstrated adequate numerical stability, with poles predominantly located inside the unit circle; however, it did not provide systematic performance gains over the all-pole model. Pareto analysis revealed a trade-off between spectral distortion and residual energy, with no clear advantage of the ARMA model under the evaluated conditions. It is concluded that, for the tested scenarios, the LPC model integrated into the Kalman filter constitutes the most robust alternative in terms of average performance, numerical stability, and computational efficiency.