Vector quantization is a very important signal compression method with rich applications in wide fields such as image signal compression and speech signal compression. In the process of codebook design, traditional LBG algorithm owns the advantage of fast convergence, but it is easy to get the local optimal result and be influenced by initial codebook. Considering that genetic programming (GP) has the capability of getting global optimal result and the characteristic of changing the individual coding length, a new clustering algorithm GP-L which is based on GP and LBG to improve the quality of codebook. Concrete complementation of the GP-L algorithm as well as the experiments are given. Experiments show that GP-L algorithm can not only extract the speakers information, in favor of individuals model , but also improved the recognition rate of the system than the traditional LBG algorithm.