A Modified Approach to Cluster Refinement for Speaker Diarization

Speaker clustering, in a speaker diarization system is of great importance since the result of speaker clustering impacts deeply on the final diarization result. However, errors can happen in every step in the clustering process, such as the estimation of the cluster number, the initialization of the cluster centers and so on. Therefore, it is necessary to modify the clustering result and improve the accuracy of the diarization system. In this paper, a modified clustering refinement approach based on "cross EM refinement" is presented to solve these issues. According to the experiment results, the performance of diarization result improved a lotwith our modified refinement, and can handle much more badly speaker clustering results than the original cross EM refinement method. The experiments are carried out three datasets of different types-meeting, broadcast news and talk-show.