Abstract
Residual-ecited linear prediction vocoding is known to be one of the best approaches to speech coding in the range of 4.8 to 9.6 kbits/s. One problem associated with the RELP vocoder is that it often produces some roughness and tonal noise as the transmission rate becomes lower. In this paper, we investigate three methods to improve its quality. These include the multiband spectral folding method, the method of using both the spectrally folded signal and the pulsed ecitation signal, and the method of using both the multiband spectrally folded signal and the pulsed ecitation signal. It has been found that, among the three methods, the last one yields the best performance. It produces no roughness and little tonal noise.