Reproduce “Self-Tuning Spectral Clustering for Speaker Diarization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nikhil Raghav, Avisek Gupta, Md Sahidullah, Swagatam Das
Researchers from the Institute for Advancing Intelligence (IAI), TCG CREST, introduced SC-pNA, a self-tuning spectral clustering method for speaker diarization that adaptively prunes affinity matrices. This approach improved computational efficiency and achieved a Diarization Error Rate (DER) of 10.27% on the challenging DIHARD-III evaluation set, outperforming other unsupervised methods and demonstrating performance competitive with semi-supervised baselines.
Key insights
- Pairwise similarity scores for a given speaker embedding often form a bimodal distribution, allowing for unsupervised identification of 'within-speaker' versus 'between-speaker' similarities.
- Aggressively pruning the affinity matrix by retaining only a small, adaptively determined percentage of the strongest within-speaker connections for each node significantly improves clustering performance and robustness.
- By performing only a single eigendecomposition on the final refined affinity matrix, computational cost can be drastically reduced compared to iterative auto-tuning approaches.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.