Nonlinear filtering of the LPC residual for noise suppression and speech quality enhancement

We study enhancing the quality of coded speech by nonlinear filtering the linear-prediction residual to make it more periodic. By attenuating the low-magnitude samples, more of the residual energy can be concentrated in the regularly spaced major excitation pulses. Wideband additive noise in the speech signal is attenuated because its residual energy is spread in time, rather than being concentrated in narrow temporal regions.