In this thesis work, various schemes in the residual-excited linear prediction (RELP) vocoders are studied. In the first part, a comparative study is done on the high frequency regenerating methods used in RELP vocoding systems. It is observed that the noisy quality in RELP output speech is introduced primarily in the course of its high frequency regenerating process. Nonlinear distortion introduces audible hoarseness, whereas spectral duplication introduced metallic tonal noise. It is found that to design the best RELP system for a given transmission rate, an optimization should be done in choosing the baseband width and the number of bits for quantization.
In the second part, two methods are proposed to improve the output speech quality of the RELP vocoding system. In one method, the RELP system has two basebands separately coded and transmitted. Three different schemes are used to reconstruct the spectrally flattened exciting signal. In another method proposed, two hybrid excitation RELP systems are discussed. They combine the spectral duplication method with that of other high frequency regeneration. It is found that they have less metallic noise and less roughness than other high frequency regeneration method alone.