Analysis of transition cost and model parameters in speaker diarization for meetings

Abstract There has been little work in the literature on speaker diarization of meetings with multiple distance microphones since publications 2012 related to last National Institute Standards (NIST) Rich Transcription Evaluation Campaign 2009 (RT09). Lately, Second DIHARD Challenge also covered at dinner party that include distant microphones. Dinner are somehow harder than office because their participants can move freely around room. In this paper, we studied some algorithms for NIST 2007 (RT07) and RT09 provide definite clear improvements. On one hand, or no care taken problem penalizing favoring transitions between speakers other proposing a minimum duration turn calculating speakers’ probabilities using Variational Bayes (VB). We have issue determined transition penalty term is needed should be independent both number active turns. determination method automatically select right parameters crucial developing good models speakers. Previous studies proposed dynamic selection based speaker’s speech mixed performance when tested microphone meetings. propose new takes into account determine parameters, question overfitting maximum them, taking computation time order reduce it. carried out experiments support our findings, able improve baseline error rate distant-microphone Both methods achieve improved over baseline. The first obtains 21.6% decrease relative development set 4.6% test second 46.47% 17.54% set. complement each other, they applied combination, obtain 47.2% 22.02% obtained proposal outstanding subsets such as RT07 among best simple modifications. Furthermore, algorithm gains without jeopardizing performance. Results different publicly available database, augmented multiparty interaction (AMI) 28.44% confirming validity methods. Preliminary single stream (mfcc) endorse findings. Comparisons an x-vector system deliver superior unseen data.