Restoration of speech packet estimating residue of burg linear prediction

Restoring speech packets is an inevitable phenomenon in Voice over Internet Protocol (VoIP) communications. This study proposes an accurate speech packet reconstruction for packet loss by estimating and incorporating the residual error of linear prediction (LP) into the actual linear prediction approach. Residual error refers to the difference between the actual speech signal and its predicted counterpart. The speech signal is non-stationary due to changing statistical properties such as frequency, amplitude, and spectral envelope over time, resulting in mismatches in frame-by-frame estimation if residual error is ignored. Ignoring residual error in conventional LP-based packet loss concealment (PLC) degrades reconstructed speech quality because it influences the system to generate a model-restricted and overprocessed speech signal. The result may seem numerically accurate, but it feels muffled, unnatural, and distorted in perception because the error component significantly affects realism. We employed the Burg LP method along with residual error estimates rather than traditional LP-based PLC methods. The conventional LP approaches minimize only the forward prediction error, whereas the Burg method minimizes the forward and backward prediction errors by applying the Levinson-Recursion (L-R) algorithm. We deployed the Burg linear prediction in the forward and backward directions to reconstruct the lost packet with higher accuracy. The Burg method effectively addresses numerical problems, creates a stable model, and delivers an accurate LP estimation even with short data lengths. However, in numerical analysis, the autocorrelation (ACR) method for LP encounters challenges in precise LP estimation. This study estimates the residual error in the packet loss case and integrates it into conventional Burg linear prediction estimates. Incorporating residual error estimation into Burg LP enhances reconstructed speech quality over previous methods because it influences statistically compatible, realistic, and perceptually natural speech signal restorations. The restoration method prevents overly confident estimations by accurately modeling the estimated residual error. Error integration to the base method minimizes spectral discontinuities. In addition, speech becomes less artificial and easier to understand when residual errors are distributed properly. The error-awareness method is more effective because it includes uncertainty. In this study, the simulation results from the proposed PLC method demonstrate significant performance improvements in both perceptual and objective evaluations compared to current LP-based PLC approaches. • This study proposes an accurate speech packet restoration by estimating and integrating residual error into linear prediction. • This study emphasizes enhancing the performance of LP for PLC. • We estimated the residual error in the packet loss case and integrated it into traditional Burg linear prediction estimates. • We proposed the Burg method in forward and backward directions to conceal packet loss with better precision. • The Burg method addresses numerical issues, creates a stable model, and delivers an accurate LP estimation even with short data lengths.