Contrastive Speaker Representation Learning with Hard Negative Sampling for Speaker Recognition.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Speech Processing

Background:

• Speaker recognition is vital for security and authentication, requiring extraction of unique speaker features.
• Current methods often use classification or contrastive learning to learn speaker relationships.
• Extracting highly discriminative features is key to achieving high speaker recognition rates.

Purpose of the Study:

• To develop a robust speaker recognition framework using contrastive learning.
• To minimize the impact of hard negative samples in speaker recognition training.
• To improve the accuracy of identifying speakers in speech utterances.

Main Methods:

• A novel framework for robust speaker recognition using contrastive learning.
• Estimating hard negative samples within mini-batches during contrastive learning.
• Employing a cross-attention mechanism to determine speaker agreement between utterance pairs.

Main Results:

• The proposed method achieved an equal error rate (EER) of 0.98% on the voxceleb1-E dataset.
• The framework resulted in an EER of 1.84% on the voxceleb1-H dataset when trained on voxceleb2.
• Demonstrated superior performance compared to conventional loss functions in speaker recognition.

Conclusions:

• The proposed contrastive learning framework significantly enhances speaker recognition robustness.
• Effectively handling hard negative samples is crucial for improving speaker identification accuracy.
• The cross-attention mechanism aids in reliable speaker verification through utterance pairing.