Multistream Diarization Fusion Using the Minimum Variance Bayesian Information Criterion

Abstract

Speaker diarization is necessary with ubiquitous and individualized recorders. We focus on the specific task of speaker diarization from two information streams, two microphones, assigned to two participants of interest. In real scenarios, speakers may be co-located, in noisy environments with interfering speakers. Multistream diarization can exploit additional information and diarization fusion is necessary. In this work we first introduce a new database that realistically simulates a range of extremely challenging acoustic conditions; and propose a Minimum Variance of BIC (MVBIC) method to combine information from the various diarization streams. We use a 2-microphone subset of our proposed database and Root Mean Square Energy (RMSE) and Mel Frequency Cepstral Coefficients (MFCC) as our two diarization streams to validate the proposed method. We show that our proposed method exploits the complementarity of the individual diarization streams and outperforms static fusion mixing weights. We also demonstrate the robustness of the MVBIC method on RT-06S data.