An efficient encoding method of excitation codes using a partial algebraic codebook (PAC) is proposed. Since the conventional algebraic code excited linear prediction (ACELP) encodes the positions and signs of all excitation pulses separately, the bits required for encoding excitation codes take a large portion of the total bit rate. Vector quantization (VQ) of the positions and signs of the excitation pulses results in a PAC. Using PAC instead of the full set of algebraic codes, we can reduce the bits required to encode the excitation codes while maintaining the output speech quality. An iterative training algorithm is proposed to obtain the suboptimal PAC by modifying the Lloyd algorithm. Simulation results show that considerable bit savings can be obtained with only a small amount of degradation in the segmental signal to noise ratio (SEGSNR).