open access

https://hdl.handle.net/2115/39813

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

英語

2009-10-04

Asia-Pacific Signal and Information Processing Association, 2009 Annual Summit and Conference, International Organizing Committee

Proceedings : APSIPA ASC 2009 : Asia-Pacific Signal and Information Processing Association, 2009 Annual Summit and Conference

817-820

APSIPA ASC 2009: Asia-Pacific Signal and Information Processing Association, 2009 Annual Summit and Conference. 4-7 October 2009. Sapporo, Japan. Poster session: Speech Processing (7 October 2009).

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

会議発表論文

Version of Record