Speaker Diarization in Overlapped Speech

Abstract

In this paper, the effect of the overlapping speech on speaker diarization performances developed in our work, using NIST evaluation metrics (DER: diarization error rate, MR: misclassification rate and ACP: average cluster purity) on NIT2005 database is studied. Using cepstral features MFCCs for a speaker diarization system, its evaluation is done by including the overlapped speech regions. If an overlapped speech region is assigned to only one speaker, this region is considered as Missed speech for the second speaker. Through our experiences, the effects of this strategy using various overlapping speech scenarios show that the speaker diarization performances (in terms of DER, MR and ACP) are drastically degraded in all situations (DER=21.08%, MR=23.83% and ACP=80.14).