H-VECTORS: Improving the robustness in utterance-level speaker embeddings using a hierarchical attention model.

Area of Science:

• Speech processing
• Machine learning
• Biometrics

Background:

• Speaker identification and verification are crucial for security and access control.
• Existing methods may struggle with varying acoustic conditions and utterance lengths.
• Robust utterance-level embeddings are needed for improved speaker recognition.

Purpose of the Study:

• To propose a novel hierarchical attention network for generating utterance-level speaker embeddings (H-vectors).
• To enhance speaker identification and verification accuracy by capturing local and global speaker-related information.
• To evaluate the effectiveness of H-vectors on diverse benchmark datasets.

Main Methods:

• A hierarchical attention network architecture is employed.
• Frame-level encoder and attention are applied to utterance segments, generating segment vectors.
• Segment-level attention aggregates segment vectors into a final utterance representation (H-vector).

Main Results:

• The proposed H-vectors demonstrate superior performance in speaker identification and verification tasks.
• The approach achieves better results compared to strong baseline methods.
• Effective performance is observed across various acoustic conditions and benchmark datasets.

Conclusions:

• Hierarchical attention networks effectively generate robust utterance-level speaker embeddings.
• H-vectors significantly improve speaker identification and verification accuracy.
• The proposed method offers a promising solution for real-world speaker recognition challenges.