Automatic speaker segmentation and classification produce generic labels rather than the true identity of the speakers. The proposed approach is based on the use of a semantic classification tree using lexical rules to extract the true identity of the speakers from the transcription. In this paper, experiments are carried out on French broadcast news from ESTER 2005 to evaluate this approach, focusing on the impact of the various combinations of automatic vs. manual transcription with automatic vs. manual speaker segmenta-tion/classification. We also study the errors generated by the system.