CELP coding with adaptive excitation codebooks

Abstract

Code-excited-linear-predictive (CELP) coding has led to speech coding schemes with a good speech quality at 8 kb/s and a fair quality at data rates of about 4.8 kb/s. In the present work, the authors propose a novel method for improving the speech quality of a CELP codec at low bit rates by adaptation of the excitation codebook (ACELP). The adaptive cookbook consists of a set of basic excitation vectors which are adapted to an analysis speech frame by calculating optimum amplitude values. The amplitudes are quantized and encoded at a very low bit rate using a gain shape vector quantizer. The results of simulation experiments show that the mean squared error by CELP with an adaptive codebook is substantially lower than that of CELP with a fixed stochastic codebook. In informal listening tests the quality of the speech processed by the adaptive CELP scheme was superior and less speaker-dependent compared to a conventional CELP scheme. >