A simplified method for RPE coding with improved performance

The regular-pulse-excitation (RPE) technique for speech coding gives good speech quality, but requires a heavy computational effort. A method for the excitation sequence search is proposed in which the pulse amplitudes are computed one at a time, rather than simultaneously in block, by solving a linear system of reduced dimension. The algorithm also provides an appreciable improvement in the performance, because the noise energy is minimized taking into account both quantization errors and the effect of pulses in the next block. The algorithm is implemented in a 8-kb/s codec, using vector quantization for the vocal tract all-pole model parameters. Computer simulations over a large database of male and female speech signals (about 80s of speech) gave good results in terms of segmental signal-to-noise ratio, even in the absence of a patch predictor. >