The Design And Implication Of Speaker Clustering Method

Speaker clustering is of great importance in speaker diarization system.The task of speaker diarization system can be described as"who spoke when",which is in the case of no prior knowledge,a given multi-speaker audio data should be segmented into homogeneous parts and clustered into groups that each group contains speech data of only one speaker.The speaker diarization system mainly consists of two modules:speaker segmentation and speaker clustering.The quality of the speaker clustering directly influences the performance of the speaker diarization system.This paper focuses on the research of speaker clustering on the following aspects:(1)An estimation method was proposed to estimate the speaker numbers automatically,necessary experiments and analysis are carried out to evaluation the estimation method.(2)Implemented a speaker clustering module based on NJW spectral clustering algorithm,which improved the performance by the adaption of scale parameters.(3)A modified refinement algorithm based on the cross EM refinement is proposed after the research on the algorithms of purification and refinement.(4)Designed and implemented a speaker diarization system,a large number of experiments were conducted to compare the performance of the baseline system(SC2SD)and our modified system(MSC2SD).Experiment results showed that our modified system has a improvement on the baseline system.