Analysis of the smoothed residual driven algorithm for speech coders

The SRD (smoothed residual driven) algorithm is investigated to understand the effect of the smoothing approximation on the algorithm´s tracking capability for tone, AR (autoregressive), and speech inputs in the presence of channel errors. For the LMS (least mean square)-SRDB algorithm, which is for the transversal structure, and the GAL-SRD algorithm, which is for the lattice structure, the error-free performance loss due to the smoothing approximation is shown to be negligible since the quantization noise and leakage act as additive noise that pulls the predictor poles inside the unit circle. The LMS-SRDB and GAL-SRD algorithms can track faster or better than other algorithms for any types of input among their class. Therefore, it is clear that the LMS-SRDB and GAL-SRD algorithms provide an excellent tradeoff among tracking of nonvoice data, robustness, and error-free performance.<>