eDIAR: training-free streaming diarization removes fixed speaker-slot limits

kaist-ai · hf · 2026-11-16

KAIST AI researchers present eDIAR, a training-free open-speaker streaming diarization method. Instead of being capped by fixed output slots (a 4-slot model can't name a 5th speaker), it clusters speaker embeddings within slots via e-means once all slots are used, supporting unlimited speakers and efficient linking for long meetings and large archives.

Original post →

More from Research

Research channel →