Reproduce “End-to-End Supervised Hierarchical Graph 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
Prachi Singh, Sriram Ganapathy
Researchers from the Indian Institute of Science developed E-SHARC, an end-to-end framework for speaker diarization that jointly optimizes speaker embedding extraction and a Graph Neural Network-based hierarchical clustering module. The system achieved relative Diarization Error Rate (DER) improvements of 11-20% over x-vector clustering baselines on AMI, Voxconverse, and DISPLACE datasets, and includes a method for handling overlapping speech.
Key insights
- Jointly optimizing speaker embedding extraction and GNN-based clustering in an end-to-end manner leads to more discriminative speaker representations and enhanced diarization accuracy.
- Graph Neural Networks effectively model complex speaker interactions and hierarchical relationships within conversational audio, outperforming traditional LSTM architectures for this task.
- A dedicated two-pass mechanism for handling overlapping speech significantly improves diarization performance by accurately identifying and assigning multiple speakers in concurrent speech regions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.