Post-processing techniques for a speaker diarization system

This paper presents the post-processing techniques designed to improve the results of a speaker diarization system. Three different techniques are proposed: refinement of speech vs. non speech segmentation, assimilation of short speech segments and fusion of clusters from the same speaker. These techniques have been implemented in a post-processing module that improves the result of the baseline system by 22.3 %. The same module has been applied to another speaker diarization system with a similar architecture to that of the baseline system with a DER improvement of 21 % and to another one with a very different architecture where no improvement has been achieved. It has also been used with another database with an improvement of 17 %. These experiments prove the validity of the techniques developed.