Multi-target Extractor and Detector for Unknown-number Speaker Diarization

Strong representations of target speakers can help extract important information about and detect corresponding temporal regions in multi-speaker conversations. In this study, we propose a neural architecture that simultaneously extracts speaker consistent with the diarization objective detects presence each on frame-by-frame basis regardless number conversation. A representation (called z-vector) extractor time-speaker contextualizer, implemented by residual network processing data both dimensions, are integrated into unified framework. Tests CALLHOME corpus show our model outperforms most methods proposed so far. Evaluations more challenging case simultaneous ranging from 2 to 7 achieves 6.4% 30.9% relative error rate reductions over several typical baselines.