Reproduce “Discriminative Neural Clustering for Speaker Diarisation” 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
Qiujia Li, Florian L. Kreyssig, Chao Zhang, Philip C. Woodland
Researchers at Cambridge University Engineering Department developed Discriminative Neural Clustering (DNC), a supervised sequence-to-sequence approach using a Transformer model for speaker diarisation. The method, enhanced by novel data augmentation and curriculum learning, achieved a 29.4% relative reduction in Speaker Error Rate compared to spectral clustering on the challenging AMI meeting corpus.
Key insights
- A discriminative, supervised approach using a Transformer can directly learn complex clustering patterns, effectively handling overlapping speaker clusters that challenge unsupervised methods.
- Novel data augmentation strategies, particularly Diaconis Augmentation, are critical for overcoming data scarcity and enabling robust training of sequence-to-sequence models on limited long audio data.
- Curriculum learning is an indispensable training strategy for DNC, allowing the model to generalize effectively from shorter to full-length, complex multi-speaker conversations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.