Reproduce “Encoder-Decoder Based Attractors for End-to-End Neural 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

Shota Horiguchi, Yusuke Fujita, Shinji Watanabe, Yawen Xue, Paola Garcia

Researchers from Hitachi, LINE, Carnegie Mellon, and Johns Hopkins introduced EEND-EDA, an end-to-end neural diarization system that accurately processes an unknown number of speakers in audio recordings. This approach consistently outperformed conventional and prior end-to-end methods on various real-world datasets, establishing new performance benchmarks for flexible-speaker diarization.

Key insights

  • Dynamically generated, adaptive speaker attractors are more effective for speaker representation than fixed classification layer weights, leading to superior diarization performance across varying speaker counts.
  • Shuffling the input sequence to the EDA's LSTM encoder improves results, suggesting that global, order-agnostic statistics of speaker embeddings are more crucial for attractor generation than strict temporal order.
  • Iterative inference, by processing 'silent' frames in successive passes and combining hypotheses with DOVER-Lap, effectively extends EEND models to handle more speakers than seen during training.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.