The X-Lance Speaker Diarization System for the Conversational Short-phrase Speaker Diarization Challenge 2022

This paper describes X-Lance Speaker Diarization System submitted to the Conversational Short-phrase Speaker Diarization Challenge. The system outputs the ensemble results of the four modules: self-attentive-based VAD, uniform segmentation, ECAPA-TDNN-based embedding extractor, and spectral clustering. We evaluated our system on the Conversational Short-phrase Speaker Diarization (CSSD) dataset, which is based on MagicData-RAMC and contains plenty of conversational short-phrase segments. Besides being different from other diarization challenges, the challenge proposes a metric called Conversational Diarization Error Rate (CDER), which focuses on evaluating short segments. In this paper, we will analyze this metric and conduct extensive experiments. Finally, our system achieves CDER of 13.2% and 8.0% in the CSSD_dev and unseen CSSD_eval set, respectively.