Performance study of stochastic speech coders

Abstract

Four speech coders in the CELP (Code Excited Linear Predictive) family are described. By replacing the long term prediction with a self excitation sequence (adaptive codebook) as well as substituting a stochastic multipulse, or sparse codebook for the commonly used codebook of Gaussian noise, the speech quality is improved. The coders are fully quantised at 7.0 and 5.0 kbit/s, interesting bit-rates for such possible applications as the half rate GSM system and the INMARSAT-M service. The performance of the coding schemes are evaluated by a formal listening test and presented by their Mean Opinion Scores (MOS). For the coder with maximum performance with respect to quality and complexity, a set of tolerable bit error rates (BER) are given. It is shown that the presence of acoustic background noise does not influence the coder quality, and that bit errors in this case will be partially masked by the background noise giving the coder a high degree of robustness. Considering the performance in the presence of bit errors and background noise, the coder seems to be suitable for use in a mobile communication system using satellite links.