The problem of extracting the target speaker speech from multiple speaker speech is researched. In order to improve the accuracy of multi-speaker speech segmentation and clustering, a speech segmentation and clustering algorithm based on Mel frequency cepstral coefficient (MFCC) and Gammatone frequency cepstral coefficient (GFCC) hybrid features is proposed, which can effectively avoid problems such as poor robustness of noisy speech segmentation and clustering. For the superimposed pink noise and factory noise speech, a comparative analysis is made based on the conventional algorithm and the improved segmentation clustering algorithm respectively. The results show that the proposed segmentation clustering algorithm based on mixed features is more accurate in extracting target human speech.