Abstract
We consider here the problem of delayed decision source coding using integer and fractional rate tree codes. First, we investigate the performance of path map symbol release rules when used with exhaustive searching of the exponential metric tree. The average single-letter distortion of fixed length symbol release algorithms and two variable length symbol release algorithms are derived for shallow search depths and compared to simulation results. The incremental single symbol release algorithm is shown to outperform other fixed multiple symbol release algorithms and variable symbol release algorithms. Simulation results are presented which compare the performance of an average fixed predictor and eight backward adaptive predictors in an adaptive differential pulse code modulation (ADPCM) system code generator for tree coding of speech at 16 kbits/s. Adaptive predictors considered include the CCITT 32 kbits/s standard predictor, two pole-zero gradient transversal predictors, and five all-pole lattice predictors. The lattice predictors are shown to have better performance than the transversal predictors, and the eighth order least squares lattice predictor yields the best performance. The experimental results also indicate the relative benefits of high order prediction and adaptive prediction for multipath tree coding. We study a new delayed decoding approach, called smoothed DPCM, by using a minimum mean squared error fixed-lag smoother in a code generator based on DPCM. For synthetic sources, SDPCM system is shown to achieve a substantial objective improvement over conventional DPCM. Rate 2 bits/sample tree coding of speech with adaptive DPCM and SDPCM code generators also shows that SDPCM has slightly better objective performance. Finally, we introduce the concept of multi-tree source codes for fractional rates and demonstrate their advantage over classical, multiple symbol per branch, fractional rate trees for speech coding with deterministic code generators. With a frequency weighted distortion measure, the forward adaptive multi-tree coder produces near toll quality speech at 16 kbits/s, while the backward adaptive 9.6 kbits/s multi-tree coder substantially outperforms adaptive predictive coding and has an encoding delay less than 2 msec.
Chang, Wen-Whei (1989). Analysis and design of low rate tree codes. Texas A&M University. Texas A&M University. Libraries. Available electronically from
https : / /hdl .handle .net /1969 .1 /DISSERTATIONS -1017058.