Abstract: With the recording voice library in the laboratory, we extract Mel frequency cepstral coefficient, its delta-cepstral coefficient and delta-spectral cepstral coefficient, and fuse them with change component method effectively. And a speaker recognition model is established based on VQ model using LGB algorithm to design different codebook capacity for the experiment and we evaluate the hybrid feature parameters in noise environments. Experimental results show that by applying the fusion hybrid feature parameters the recognition system rate and robustness are obviously improved compared with other parameters.