Coding Speech at Very Low Rates Using Temporal Decomposition-Based Spectral Interpolation and Mixed Excitation in the LPC Model

摘要

This paper presents a new method for speech coding at rates around 1700 bps based on the linear predictive coding (LPC) model. For encoding the spectral information, hierarchical temporal decomposition (HTD) is used, which reduces the information rate through transformation of a given matrix of spectral parameters from the parameter space to event space. Both long-and short-delay HTD-based coders are implemented and a method for designing the short-delay HTD is introduced. The LPC excitation is approximated using an adaptive three-band model, which combines periodic and random signal components. This is performed by characterising the LPC residual signal spectrum using an instantaneous frequency (IF) estimation-like technique, which gives accurate information about the pitch and its harmonics over the frequency domain. Preliminary results, obtained using a nonuniform scalar quantisation method, show that the performance of the proposed coding method is close to that of the MELP-2400 standard, particularly in adverse conditions.