The Nist 2004 Spring Rich Transcription Evaluation: Two-axis Merging Strategy in the Context of Multiple Distant Microphone Based Meeting Speaker Segmentation

This paper presents the ELISA speaker segmentation approach applied on multiple audio channel meeting recordings in the framework of NIST RT’04s meeting (spring) evaluation campaign. As done for BN data speaker segmentation, the ELISA “meeting” system involves two speaker segmentation systems developed individually by the CLIPS and LIA laboratories. The main originality consists in a “two-axis” merging strategy, proposed to deal with both multiple expert segmentation outputs and multiple microphone segmentation outputs. While expert merging strategy did not really lead to an improvement of the performance, the individual microphone segmentation merging strategy allowed to provide a global segmentation output from several audio channels (microphones) with acceptable performance. The best system obtained 22.6% of diarization error rate during the NIST RT’04s meeting evaluation.