Audio Signal Processing and Coding

8.6 SUBBAND CODING WITH HYBRID FILTER BANK/CELP ALGORITHMS

While hybrid sinusoidal-DWT and sinusoidal-DWPT signal models seek to maximize robustness and basis flexibility, other hybrid signal models have been motivated by low-delay and low-complexity concerns. In this section, we consider, in particular, algorithms that combine a filter bank front end with subband-specific code-excited linear prediction (CELP) blocks for quantization and coding of the decimated subband sequences. The goal of these experimental hybrid coders is to achieve very low delay and/or low-complexity perceptual coding with reconstruction quality comparable to any state-of-the-art audio codec. Before considering these algorithms, however, we first define what is meant by “code-excited linear prediction.”

In the coding literature, the acronym “CELP” denotes an entire class of efficient, analysis-by-synthesis source coding techniques developed primarily for speech applications in which the analyzed signal is treated as the output of a source-system mechanism such as the human vocal apparatus. In the CELP scheme, excitation vectors corresponding to the lower vocal tract “source” contribution drive a slowly time-varying LP synthesis filter that corresponds to the upper vocal tract “system.” Parameters of the LP synthesis filter are usually estimated on a block basis, typically every 20 ms, while the excitation vectors are usually updated more frequently, typically every 5 ms. The LP parameters are most often ...