On the performance of CELP algorithms for low rate speech coding

The performance of a 4.6 kb/s CELP (code-excited linear prediction) speech coder has been investigated by using almost stationary speech sources. The results show that quantization of voiced phones produces much more audible noise than that of unvoiced ones. Investigations using spectral analysis lead to the conclusion that a considerable part of the noise is caused by degradation of fine structure information. The performance of the long-term prediction is severely influenced by the quantization of the short-term prediction parameters. Some strategies for improving coder performance are discussed