Online End-to-End Neural Diarization with Speaker-Tracing Buffer
share
End-to-end speaker diarization using a fully supervised self-attention mechanism (SA-EEND) has achieved significant improvement from the state-of-art clustering-based methods, especially for the overlapping case. However, applications of original SA-EEND are limited since it has been developed based on offline self-attention algorithms. In this paper, we propose a novel speaker-tracing mechanism to extend SA-EEND to online speaker diarization for practical use. First, this paper demonstrates oracle experiments to show that a straightforward online extension, in which SA-EEND is performed independently for each chunked recording, results in degrading the diarization error rate (DER) due to the speaker permutation inconsistency across the chunk. To circumvent this inconsistency issue, our proposed method, called speaker-tracing buffer, maintains the speaker permutation information determined in previous chunks within the self-attention mechanism for correct speaker-tracing. Our experimental results show that the proposed online SA-EEND with speaker-tracing buffer achieved the DERs of 12.84 for Corpus of Spontaneous Japanese with 1s latency. These results are significantly better than the conventional online clustering method based on x-vector with 1.5s latency, which achieved the DERs of 26.90 respectively.
READ FULL TEXT
