Contrastive Adversarial Domain Adaptation Networks for Speaker Recognition

Area of Science:

• Machine Learning
• Deep Learning
• Computer Science

Background:

• Domain adaptation addresses discrepancies between source and target data domains.
• Domain Adversarial Networks (DAN) use adversarial learning for domain invariance but struggle with single feature extractors.
• A need exists for improved methods to extract domain-invariant features in deep learning.

Purpose of the Study:

• To propose a novel domain adaptation network, Contrastive Adversarial Domain Adaptation Network (CADAN).
• To enhance the creation of domain-invariant feature spaces by decoupling feature extraction branches.
• To improve the accuracy of tasks like speaker identification in varied conditions.

Main Methods:

• Splitting the feature extractor into two contrastive branches: one for class-dependence, one for domain-invariance.
• Sharing initial and final hidden layers while decoupling middle layers for specialized feature extraction.
• Adversarially training a label predictor to yield equal posterior probabilities for class-discriminative features.

Main Results:

• CADAN effectively creates domain-invariant and class-discriminative embedded features.
• Speaker identification experiments showed significant accuracy improvements with CADAN.
• CADAN achieved a 33% increase in speaker identification accuracy compared to conventional DAN.

Conclusions:

• The proposed CADAN architecture offers a superior approach to domain adaptation.
• Decoupled contrastive branches in feature extractors enhance domain invariance.
• CADAN demonstrates significant performance gains in real-world applications like speaker identification.