End-to-End Neural Speaker Diarization with Non-Autoregressive Attractors

Explore the latest advancements in end-to-end neural speaker diarization (EEND) systems through this 37-minute conference talk by Magdalena Rybicka from the Center for Language & Speech Processing at Johns Hopkins University. Delve into the challenges of making diarization robust and effective in real-life scenarios, and learn about the EEND with Non-Autoregressive Attractors (EEND-NAA) approach. Discover how this innovative system handles recordings containing speech from a variable and unknown number of speakers, utilizing a clustering approach within the EEND-EDA framework. Gain insights into the explainable process of attractor generation and understand the advantages of replacing the autoregressive LSTM-based backend with non-autoregressive attractor estimation. Benefit from Rybicka's expertise in speaker recognition and machine learning applications as she shares her research findings and discusses the potential impact of these developments on the field of speaker diarization.