For the noise robustness problem in i-vector: Based on the theoretical principle of i-vector speaker recognition system, the extraction principle and scoring calculation method of i-vector and the process of channel compensation algorithm based on PLDA (Probabilistic Linear Discriminant Analysis) with PLDA model are studied. The matching principle is studied. A statistical averaging i-vector extraction algorithm based on speech fragmentation is proposed to extract more robust i-vector features by weakening the statistical parameters of bad speech fragments to improve the recognition performance of the system. After that, the i-vector system is designed to improve the recognition performance of the i-vector.l Then, a Quantum Particle Swarm Optimization is designed to optimize the parameters of the i-vector recognition system to avoid the degradation of the system performance caused by artificial empirical values. Experimental analysis shows that the proposed algorithm has improved performance over the traditional i-vector recognition algorithm, especially in the case of noise interference, and has better recognition performance