Abstract:

In order to find the number of speaker roles and the corresponding speakers ’ speech in meeting speeches ， a clustering method for multiple speaker roles is proposed. Firstly ， features for speaker role clustering are defined. Secondly ， geodesic distance is used to measure the similarities among features. Then ， inner-class distance is used to control inter-class mergence to form the clustering method. Finally ， four different types of meeting speech corpora are used to validate the effectiveness of the proposed method. The results indicate that ，for the meeting speeches obtained by both manual and automatic segmentation ， the clustering performance using geodesic distance is superior to that using traditional distance when the same clustering algorithm is used in all cases ， and that the proposed method performs better than the traditional hierarchical clustering method when the same measuring distance is used.

Key words: speaker role, characteristic distance measure, role clustering, geodesic distance, unsupervised clustering