Convergence and Numerical Sensitivity of Backward-Adaptive LPC Predictor
Juin-Hwey Chen, Richard Vandervoort Cox · 2005
Backward-adaptive LPC prediction is a useful and relatively new scheme for low-delay speech coders. However, if not performed carefully, it can lead to numerical sensitivity problems, decoder divergence, or other unexpected problem in this paper, we report four useful techniques: white noise correction, bandwidth expansion, window tuning, and spectral smoothing, which collectively are sufficient to avoid all such problems known to us. All except window tuning are also useful for reducing the numerical sensitivity of forward-adaptive LPC prediction that is commonly used in predictive coders such as CELP, MPLPC, and APC.