On reducing the bit rate of a CELP-based speech coder

For a code excited linear prediction (CELP)-based coder, coarse quantization usually results in the degradation of speech quality. Spectrum and gain were found to have the least sensitivity to quantization errors and were efficiently quantized. The frame period and number of subframes per frame were also appropriately selected for the purpose of bit-rate reduction. The performance of four possible excitations, namely, stochastic excitation, binary pulse excitation, self-excitation and repeat excitation were investigated. Using signal-to-noise ratio as a quality measure, the coder at 3076 b/s was found to have a performance very close to that of the federal standard at 4800 b/s