Spectral sensitivity analysis of PARCOR parameters for speech data compression

Spectral sensitivity analysis is introduced to measure the spectral deviation due to feature parameter perturbation in the linear prediction of speech. Using a 10 percent log sensitivity measure which is closely connected with the perception of spectral deviation, some factors influential on feature parameter sensitivity are evaluated. Three sensitivity reductions, 1) nonlinear transformation of PARCOR coefficients (reflection coefficients), 2) preemphasis, and 3) windowing, work effectively in reducing sensitivity. The combination of nonlinear transformation and lag windowing is the most useful to remove speaker differences in sensitivity and to achieve speech data compression. The information rate to transmit feature parameters in PARCOR analysis/synthesis is computed from the spectral sensitivity measure and the statistical distribution range of parameters. The required information rate per frame can be reduced below 40 bits in the worst case.