EEND-DEMUX: End-to-End Neural Speaker Diarization via...

Review Writer: Dongkeon Park (2024.01.29)

👁️‍🗨️ 1. Introudction

🗼2. Proposed EEND-DEMUX Framework

A. Model Architecutre

• EEND-DEMUX는 4개의 module로 구성
  • MixtureEncoder, Demultiplexer, AttractorDecoder, SpeakerEncoder

🖼️ Figure

• MixtureEncoder
• Demultiplexer
• AttractorDecoder
• Posterior probability estimation & speaker existence
• SpeakerEncoder

B. Training Objective Functions

(1) Diarization loss ($\mathcal{L}{\textnormal{diar}}, \mathcal{L}{\textnormal{ext}}$)

(2) Demultiplexing loss ($\mathcal{L}{\textnormal{dis}}, \mathcal{L}{\textnormal{ort}}, \mathcal{L}_{\textnormal{spa}}$)

🧪 3. Experiments