Speaker recognition system needs sufficient data to discriminate speaker well. In case of limited data, especially when the amount of available training and testing data were few seconds, the system performance decreased significantly. It proposed a discriminative weighted fuzzy kernel vector quantization method for speaker identification with limited data. By non-linear mapping, it quantized the input data in the high-dimensional feature space, and used the cluster centers to form the speaker's model. In the matching phase, it took into account the relationship between the reference models in feature space, and assigned the larger weights for code vectors with high discriminative power. Experimental results show that when the training data and testing data is limited, this method can provide good performance.