eDIAR extends fixed-slot streaming diarizers into an open-ended system with no limit on the number of speakers, without requiring retraining, and uses e-means to cluster speaker embeddings. The method is more accurate than Streaming Sortformer on meetings lasting up to one hour, while the cost of each update does not grow with meeting duration. Across multiple sets of recordings, e-means achieves clustering quality comparable to Lloyd’s k-means and runs several times faster.
Source:Hugging Face 社区论文精选 · Read original article
Leave a Reply