Online Neural Speaker Diarization with Spectral Clustering for Meeting Scenarios

This paper proposes an online speaker diarization system based on spectral clustering algorithm and a neural speaker diarization model. To efficiently process long-time recordings, we perform the online spectral clustering in a block-wise manner. We use two core samples, one of which contains fewer samples and is used for online clustering to meet the low latency requirement, and the other contains more samples, which are clustered of-fline to obtain more precise clustering results used to update the samples in the former. The neural speaker diarization model requires speaker masks and target speaker embeddings as inputs, so we can use results obtained from spectral clustering to obtain these components. Additionally, we evaluate our system on both the evaluation and test sets of the AliMeeting dataset and find that our online system achieves better results than previous work.