Reproduce “Neural Multichannel Distant Speaker Diarization with Heavy-tailed Source Separation Model” 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
Sicheng Mao, Baihan Li, Mathieu Fontaine, Anthony Larcher, Roland Badeau
A heavy-tailed extension of neural FCASA replaces its complex-Gaussian source likelihood with Generalized Gaussian and Student’s t models for joint multichannel source separation and speaker diarization. Student’s t variants consistently improve diarization on AMI and AliMeeting and substantially reduce false alarms on the challenging CHiME-6 corpus, although performance depends on the tail parameter and overlap results are less consistent.
Key insights
- Student’s t modeling is generally more reliable than the Gaussian baseline, while Generalized Gaussian variants show mixed results; heavier tails alone do not guarantee improvement.
- The largest practical gain on CHiME-6 comes from reducing false alarms, suggesting that heavy-tailed source modeling helps distinguish speaker activity from complex environmental interference.
- The optimal tail parameter is corpus-dependent, and cross-corpus transfer—especially involving CHiME-6—remains inconsistent.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.