An investigation of mixed‐source excitation models for linear predictive vocoders

Abstract

Two methods were investigated for modeling the excitation for linear predictive vocoders as a mixture of pulses (periodic component) and noise (nonperiodic component). In one method, linear filters were used to separate the periodic and nonperiodic components of the residual signal. In another method, the two components were separated by locating the areas of peak energy in the residual and labeling the portion of the residual within each pitch period peak as the periodic component and the remainder of the signal as the nonperiodic component. Thus, in the first method, a frequency domain separation of the periodic and nonperiodic component was assumed whereas, in the second method, a time‐domain separation of the two components was assumed. The time‐domain method appeared to be more feasible. Speech was synthesized using a number of parameters to represent the two components of the excitation. Although some improvements in speech quality could be obtained by using very detailed representations of the pulses, less‐detailed representations resulted in poorer quality speech than that obtained using a standard impulse excitation.