Attention-Based Temporal-Frequency Aggregation for Speaker Verification.

Area of Science:

• Speech processing
• Machine learning
• Biometrics

Background:

• Convolutional neural networks (CNNs) are pivotal in speaker verification (SV) due to their feature learning capabilities.
• Utterance-level aggregation in CNN-based SV compresses frame-level features but often overlooks frequency domain information.
• Existing methods primarily aggregate features temporally, limiting the capture of speaker-specific frequency characteristics.

Purpose of the Study:

• To address the limitations of temporal aggregation in CNN-based SV systems.
• To propose novel attention-based frequency and temporal-frequency aggregation methods.
• To enhance the discriminability of speaker embeddings by incorporating frequency domain information.

Main Methods:

• Development of an attention-based frequency aggregation method to identify informative frequency bands.
• Introduction of two novel temporal-frequency aggregation methods combining time and frequency domain analysis.
• Implementation and evaluation of a CNN-based SV system utilizing the proposed aggregation techniques.

Main Results:

• The proposed temporal-frequency aggregation method significantly improves speaker embedding discriminability.
• The CNN-based SV system achieved a superior equal error rate (EER) of 5.96% on the Voxceleb dataset.
• Experimental validation on TIMIT and Voxceleb datasets demonstrates the effectiveness of the proposed methods over state-of-the-art baselines.

Conclusions:

• The novel temporal-frequency aggregation methods effectively capture speaker-dependent information from both time and frequency domains.
• The proposed approach enhances the performance of CNN-based speaker verification systems.
• This work offers a promising direction for improving the accuracy and robustness of speaker recognition technologies.