Speaker diarization of broadcast audio using automatic transcription, iVectors and cosine distance scoring

In this paper we present our system for speaker diarization of broadcast audio. In the system, segmentation of the processed spoken document utilizes an automatic transcription and speech segments determined by the speaker change point detector are represented by iVectors. Similarity of speech segments is evaluated using cosine distance scoring and linear discriminant analysis is applied to cope with intra-speaker variability. We demonstrate improvement of the performance over the baseline system employing methods based on the Bayesian Information Criterion (BIC). The presented speaker diarization system achieved 39.2% relative improvement of the diarization error rate over the baseline.