In the past recent years, it appeared that sparse code books in CELP coders provide significant computation reduction without degrading the quality of the synthetic speech. We present a method that reduces the calculation complexity of the cross-correlation between the original speech p and the synthetic speech pin case of ternary valued samples (-1, 0, +1) code book sequences. After describing the method, we apply it to the Federal Standard 1016 CELP coder. We show that the computation reduction for calculating the cross-correlation can be up to 5.