chapter  17
10 Pages

Speaker Diarization of Large Corpora

Diarizing large collections of speech data is a task for which current systems quickly reach computational complexity limits. The few approaches in the literature addressing it unanimously split the task into speaker diarization (see Chap. 16) and linking stages. The role of the speaker linking stage is to label the segments output by the diarization system that are similar enough with the same identified speaker (speaker identifier). Speaker linking is often implemented as a hierarchical clustering algorithm, thus approaching the large scale diarization task as a two-step clustering process. Following an initial presentation of these stages (Sect. 17.1), we present in this chapter a method for speaker linking across meetings.