Current cross-show diarization systems are mainly based on an overall clustering process which handles all the shows within a collection simultaneously. This approach has already been studied in various situations and seems to be the best way so far to achieve low error rates. However, this process has limits in realistic applicative contexts where large and dynamically increasing collections have to be processed. In this paper we investigate the use of an incremental clustering cross-show speaker diarization architecture to iteratively process new shows within an existing collection. The new shows to be inserted are processed one after another, according to the chronological order of their broadcasting dates. Experiments were conducted on the data distributed for the ETAPE and the REPERE French evaluation campaigns. It consist of 142 hours of data collected from 310 shows, from a period from Sept. 2010 to Oct. 2012.
Bibliographic reference. Dupuy, Grégor / Meignier, Sylvain / Estève, Yannick (2014): "Is incremental cross-show speaker diarization efficient for processing large volumes of data?", In INTERSPEECH-2014, 587-591.