In this paper, several algorithms for the calculation of ternary excitation signals for CELP coders are analyzed. A new suboptimal algorithm for placing non-zero components in a sparse ternary excitation vector is proposed. This algorithm consists of successive minimizations of the angle between the perceptual speech vector and its model. It has been compared with the optimal algorithm, in which each permissible excitation vector is tested. The new algorithm offers a substantial complexity reduction with a very small decrease of speech quality. It can be also used in a CELP coder with orthogonalization of the ternary excitation vector relative to the long term prediction vector.