Speaker Representation Learning via Contrastive Loss with Maximal Speaker Separability

Keyphrases

• Learning Objectives 100%
• Separability 100%
• Contrastive Loss 100%
• Speaker Representation Learning 100%
• Utterance 80%
• Embedding Space 60%
• Contrastive Learning 40%
• Easy-to-implement 20%
• Training Data 20%
• New Learning 20%
• Design Learning 20%
• Label Information 20%
• Random Data 20%
• Deep Model 20%
• Speaker Embedding 20%
• Data Augmentation 20%
• Supervised Contrastive Learning 20%
• PyTorch 20%
• Additive Angular Margin Loss 20%
• Positive Pairs 20%
• ECAPA-TDNN 20%
• Speaker Discrimination 20%
• Unseen 20%
• Unseen Domain 20%

Computer Science

• Embedding Space 100%
• Contrastive Learning 100%
• Representation Learning 100%
• Contrastive Loss 100%
• learning design 33%
• Data Augmentation 33%
• Training Data 33%
• Experimental Result 33%