Attentive Statistics Pooling for Deep Speaker Embedding

 

 

 

 

 

 

 

Overview

• Speaker recognition should be able to get embedding that has
  • Small intra-speaker and
  • Large inter-speaker distance
• Evaluate most popular loss functions for speaker recognition on the VoxCeleb dataset
• Propose new metric learning objective function

 

Higher-order pooling with attention

• Statistics pooling
  • Calculate mean vector [1]
  • Calculate standared deviation vector over frame-level features h_t(t=1,...,T) [2]