Predictive and residual encoding of speech

Abstract

This paper surveys recent developments in adaptive predictive coding (APC) of speech. Prominent among these developments are the use of a three‐point pitch predictor, a pitch‐adaptive quantizer, entropy coding of the residual, and adaptive shaping of the quantization‐noise spectrum. APC systems produce high quality speech at around 16 kbit/s; their quality diminishes rapidly at 9.6 kbit/s or less. For those lower data rates, some form of baseband coding system becomes desirable. In such systems, a low‐frequency baseband is transmitted. The high‐frequency regeneration of the excitation spectrum from the baseband is of special importance. Traditional regeneration techniques have used some form of nonlinear distortion (usually rectification) of the baseband, followed by spectral flattening. We introduce a new set of regeneration based on duplication of the baseband spectrum at high frequencies. The audible signal distortions in rectification and spectral folding are compared.