Frame-wise and overlap-robust speaker embeddings for meeting diarization
❓Question
- What is receptive field of ResNet34?
- Is it covered all duration of window?
Introduction
Speaker embedding extraction
Speaker Embedding in Diarization
⚠️ Problems
Related Solution: EEND
Related Solution: EEND-EDA
Contribution
Proposal
2. Teacher-Student Embedding
- "teacher"는 표준 speaker embedding system
- "student"는 TAP이 필요 없는 frame-wise embeddings을 생성하도록 설계
Objective
- diarization scenarios에 유용한 더 신뢰할 수 있고 높은 temporal resolution의 embeddings을 생성할 수 있는 시스템
2.1. Teacher speaker embedding