Bidirectional Attention for Text-Dependent Speaker Verification.

Area of Science:

• Biometric authentication
• Speech processing
• Deep learning

Background:

• Automatic speaker verification is crucial for secure authentication.
• Existing deep learning methods have limitations in utilizing joint utterance information.
• Fixed speaker representations and ignored inter-utterance data hinder performance.

Purpose of the Study:

• To improve speaker verification accuracy using a single enrollment utterance.
• To leverage joint information between enrollment and evaluation utterances.
• To develop a more robust speaker discriminative neural network.

Main Methods:

• Combining Convolutional Neural Network (CNN)-based feature learning with a bidirectional attention mechanism.
• Exploiting evaluation-enrollment joint information for interactive features via bidirectional attention.
• Introducing an individual cost function to identify phonetic content for attention score calculation.

Main Results:

• Achieved a competitive equal error rate (EER) of 6.26% on the "DAN DAN NI HAO" benchmark dataset.
• Outperformed traditional i-vector/PLDA and deep learning baselines like d-vector, self-attention, and sequence-to-sequence attention models.
• Demonstrated the effectiveness of interactive features complementing constant speaker representations.

Conclusions:

• The proposed method significantly enhances speaker verification performance.
• Joint analysis of utterances and phonetic content identification are key to improved accuracy.
• The approach offers a promising direction for single-enrollment speaker authentication systems.