Recent progress in low-bit-rate speech coding

RELP schemes, which are based on a model of the vocal tract and the estimation of an optimum excitation function, are reviewed. In the RELP-S scheme the speech signal is interpreted as the result of filtering synthetic excitation functions by a filter which models the vocal tract. Two types of RELP-S schemes using different excitation generators are described. Refinements of this concept as vector sum excited linear prediction (VSELP) and post-filtering strategies lead to CELP codecs at rates between 8 kb/s and 4.8 kb/s able to achieve a speech quality which is scored between fair and good in subjective MOS-terms.<
>