Abstract
Tree coding is a promising way of obtaining good performance for medium-to-low rate speech coding. The key part of a tree coder is the code generator which consists of a short-term predictor and a long-term predictor. The best predictor designed for non-stationary speech should be adaptive, and forward adaptation and backward adaptation are widely studied in certain types of speech coding. In this research, based on the same adaptive long-term predictor, tree coders are implemented with different short-term code generators, forward adaptive and backward adaptive. To compare their advantages and disadvantages, the forward code generators use a 4-2 multi-tree structure at rate 13200 bits/s, while the backward code generators use a 4-2-4 multi-tree structure at rate 13333 bits/s. Interpolation is performed for forward code generators, while a powerful RLS algorithm is used in the backward code generators. A novel hybrid code generator which combines the forward and backward code generators in a system at rate 13200 bits/s is designed. The results indicate that the hybrid adaptive system performs better in terms of spectrograms than the forward adaptive systems and that the backward adaptive system at 13333 bits/s performs as well, even better than the hybrid adaptive system at 13200 bits/S. For each tree coder, reconstructed speech is evaluated using signal-to-noise ratios (SNRs), segmental SNRS, spectrograms, and informal listening tests.
Dong, Hui (1998). Adaptive code generators for tree coding of speech. Master's thesis, Texas A&M University. Available electronically from
https : / /hdl .handle .net /1969 .1 /ETD -TAMU -1998 -THESIS -D66.