Improving Transformer-based End-to-End Speaker Diarization by Assigning Auxiliary Losses to Attention Heads

Ye-Rin Jeoung, Joon-Young Yang, Jeong-Hwan Choi +1

Transformer-based end-to-end neural speaker diarization (EEND) models utilize the multi-head self-attention (SA) mechanism to enable accurate speaker label prediction in overlapped…