ODYSSEY 2004 - The Speaker and Language Recognition Workshop
May 31 - June 3, 2004
This paper presents the ELISA consortium activities in automatic speaker diarization (also known as speaker segmentation) during the NIST Rich Transcription (RT) 2003 evaluation. The experiments were achieved on real broadcast news data (HUB4), in the framework of the ELISA consortium. The paper firstly shows the interest of segmentation in acoustic macro classes (like gender or bandwidth) as a front-end processing for segmentation/diarization task. The impact of this prior acoustic segmentation is evaluated in terms of speaker diarization performance. Secondly, two different approaches from CLIPS and LIA laboratories are presented and different possibilities of combining them are investigated. The system submitted as ELISA primary obtained the second lower diarization error rate compared to the other RT03-participant primary systems. Another ELISA system submitted as secondary outperformed the best primary system (i.e. it obtained the lowest speaker diarization error rate).
Bibliographic reference. Moraru, Daniel / Meignier, Sylvain / Fredouille, Corinne / Besacier, Laurent / Bonastre, Jean-François (2004): "ELISA nist RT03 broadcast news speaker diarization experiments", In ODYS-2004, 23-28.