Analysis and design for robust, low delay tree coding of speech at 9.6 kbps

Tree coding is a promising way of obtaining good performance for medium-to-low rate speech coding, but when the system parameters are updated by backward adaptation methods to achieve low delay, the effect of channel errors has to be considered. The recursive structure in the code generator of a tree coder causes the output signal to have a long memory in which the effect of an erroneously received symbol can propagate. Several elements of the code generator, such as the pitch and formant predictors and the step size adaptation algorithm, have such recursive structures which need to be changed for reliable performance in the presence of channel errors. We develop a pitch predictor which has robust performance in noisy channels by examining each element of the pitch predictor such as the parameter tracking algorithm, voicing decision, and pitch detection algorithm. The step size adaptation algorithm is also improved by introducing the concept of a multiple leakage factor and a new 1-bit quantization scheme for the multitree structure. The choice of the input to the formant predictor coefficient adaptation algorithm should be judicious since the effective memory inherent in the adaptation input and its closeness to the reconstructed output seriously affect the noiseless and noisy channel performance. The inclusion of all-zero prediction into the code generator can compensate for the degradation of the existing residual driven adaptation algorithm while the robustness to channel errors is not impaired, when used with appropriate choices of the adaptation inputs. The 9.6 kbps tree coder with the new code generator shows a consistently improved performance over the one with the old code generator in the noiseless and noisy channels. Finally, various configurations for 9.6 kbps are examined and discussed.