The invention discloses a vector quantization method based on the normal distribution law. In the codebook generation, the feature center of a training sample set is first used as an initial codeword, and then the initial codeword is divided. When the codeword is not divided for the first time, the width of each codeword and the number of training samples belonging to each codeword are multiplied. The L codewords with the largest product are selected as the codewords to be divided, and a new codeword mean value is obtained to realize the codeword division. Every time after the codeword division, a data partitioning stage starts. In the data partitioning, the codewords are described as normal distribution, the degree of membership of each training sample to each codeword is calculated, and the cell is divided to update the codeword. Whether the quantitative distortion is convergent is judged, and if not, the data partitioning continues. If the quantitative distortion is convergent, whether the total number of the codewords has reached a certain value is judged, if not, the codeword division continues, and if so, a final codebook is output. The method of the invention can improve the accuracy of codeword division and reduce the error of vector quantization.