Speaker segmentation based on between-window correlation over speakers' characteristics

摘要: Speaker segmentation is widely applied in many domains such as multi-speaker detection and speaker tracking. However, the performance of conventional metric-based methods neither good enough nor stable due to stability between-window distance calculation. In order enhance hence improve performance, a new method based on correlation over speakers' characteristics proposed. this method, set reference models are trained which can represent whole model space. The likelihood vectors scores against these taken metric. gender information Peak Valley also used. Experiments NIST SRE 2002 Segmentation BNEWS SWBD Datasets show that better be achieved compared with BIC GLR methods. What's more, proposed achieve approximately best wider value range predefined thresholds than methods, reduces threshold sensitivity.