Selective GMM EM for telephone diarization

In this paper, we present a novel approach to optimizing the Gaussian-mixture-model (GMM) training phase for a Speaker Diarization System based on hidden distortion models (HDMs). Generally with HDMs, emission models and the transition costs adapts to decrease the overall cost at each iteration. In the current work we use GMM as an emission model. However, in speaker diarization, the real goal is improving the diarization error rate (DER). We used an existing HDM based speaker diarization system and change it by adapting the GMM-EM algorithm to be more selective. Instead of maximizing the likelihood, the new algorithm will emphasize different events by selectively focusing on the K highest posteriors for likelihoods estimation. A proper choice of K leads to a relative improvement of - 5.5%.