Tree and trellis coding of speech and stationary speech-like signals

This paper deals with the application of tree and trellis coding concepts to the coding of stationary first-order auto-regressive sources and speech signals at a rate of 1 bit/sample. We show first that a probabilistically generated tree (or trellis) performs significantly better than a tree (or trellis) that is generated in a deterministic way. We show then that the segmental signal-to-noise-ratios of conventional speech coding schemes (including) those with adaptive quantization and/or prediction) can be increased by about 2 dB. Slightly higher values of 3 to 4 dB have been realized with schemes which employ backward-adaptive quantization.