Target sample mining with modified activation residual network for speaker verification.

Area of Science:

• Speech processing and machine learning
• Biometric security systems

Background:

• Traditional Softmax methods in speaker verification suffer from training-verification discrepancies and imbalanced sample issues.
• Overlapping similarity scores between positive and negative samples reduce discriminability in speaker verification systems.
• Deep neural networks face challenges like gradient vanishing, explosion, and model degradation.

Purpose of the Study:

• To introduce an adaptive target function, SphereSpeaker, to overcome limitations of traditional Softmax in speaker verification.
• To enhance the stability and generalization ability of speaker verification models.
• To improve the discriminability between positive and negative speaker samples.

Main Methods:

• Developed SphereSpeaker, an adaptive target function with modified hyperparameters based on Softmax.
• Incorporated three distinct angular margins to refine network updates.
• Utilized a Residual Network PReLu (ResNet-P) architecture to address deep neural network training issues.

Main Results:

• SphereSpeaker demonstrated improved suitability for speaker verification tasks.
• The ResNet-P architecture enhanced network stability and generalization.
• Experimental results showed the lowest equal error rate compared to existing deep neural network methods.

Conclusions:

• SphereSpeaker effectively addresses training-verification discrepancies and sample imbalance in speaker verification.
• The proposed ResNet-P with SphereSpeaker significantly improves speaker verification system performance.
• This approach offers a more robust and accurate solution for speaker identification and verification.