Abstract: In speech signal processing, the self-adaptation vector quantization is widely used. A code training method of self-adaptation vector quantization based on SFCM algorithm is proposed. Its feature is that the membership between training samples and codebook is readjusted and the least coding distortion is reached by the fuzzy clustering method. The codebook is more adaptable to new speaker. The calculation of this method is simple. The experiment resu1t of this method is that the coding average distortion is low.