eDIAR: Training-Free Open-Speaker Streaming Diarization for Long Meetings and Large Archives via e-means
eDIAR removes a streaming diarizer’s speaker-count cap without retraining the model.
Once all output slots are occupied, it uses e-means to cluster speaker embeddings within a slot while the existing model continues tracking speech activity and overlap. Streaming revisions stay within a bounded window, keeping the cost per update from growing with meeting length. The authors report substantially better accuracy than Streaming Sortformer on hour-long meetings. For linking speakers across archives, they report clustering quality matching Lloyd’s k-means at several times the speed.
HF Daily Papers' note
Once all output slots are occupied, it uses e-means to cluster speaker embeddings within a slot while the existing model continues tracking speech activity and overlap. Streaming revisions stay within a bounded window, keeping the cost per update from growing with meeting length. The authors report substantially better accuracy than Streaming Sortformer on hour-long meetings. For linking speakers across archives, they report clustering quality matching Lloyd’s k-means at several times the speed.
HF Daily Papers' note
score 5