Analysis-by-synthesis speech coding with quantization noise modeling

In analysis-by-synthesis linear predictive coding (AbS-LPC) an LPC synthesis filter is combined with an analysis-by-synthesis search of the excitation signal. The synthesis filter is an estimator for the speech signal given the excitation. However, in most AbS-LPC algorithms this estimator has no explicit model of the quantization noise, which is present in the excitation signal. This paper describes quantization noise modeling in a vector AbS-LPC algorithm. Methods based on recursive Bayesian filtering and Kalman filtering are considered. Simulations indicate improved signal-to-noise ratios due to quantization noise modeling.