A study of speaker segmentation of dialogue speech with speech overlapped section

摘要

In this paper, speaker segmentation of dialogue speech was studied to develop a support system for dialogue analysis. The dialogue speech addressed in this study was spoken by two speakers, and had speech overlapped section. To train HMM for the overlapped section, two different speech databases were constructed: (1) same words spoken by the two speakers were added; (2) pairs of different words spoken by the two speakers were added. These two training approaches were compared using dialogue speech data. The experimental result showed the latter have better performance. In addition, removing redundant arcs on the HMM network was improved the performance.