Low-delay analysis-by-synthesis speech coding using lattice predictors

Results on low-delay vector excitation coding (LD-VXC) obtained by using adaptive lattice filters for implementing the short-term predictor and the perceptual weighting filter are discussed. The new codec, the lattice LD-VXC (LLD-VXC), is based on a backward adaptive analysis-by-synthesis configuration in which a least-mean-square (LMS) recursive algorithm is used for updating the lattice filters. The shape-only codebook and the gain-shape codebook are compared as possible candidates for the excitation codebook. The performance of the LLD-VXC codec versus the short-term predictor order is studied. It is shown that the performance increases for short-term predictor orders of up to 20-30 and then saturates. A LLD-VXC codec with a pitch predictor and a short-term predictor or order 20 achieves the same speech quality as a system without a pitch predictor and with a short-term predictor of order 50, and the LLD-VXC codec offers toll speech quality at 16 kb/s with moderate complexity and a total communications delay of under 2 ms