Lambda-vector modeling temporal and channel interactions for text-independent speaker verification.

Area of Science:

• Speech Processing
• Machine Learning
• Biometrics

Background:

• Current speaker verification models (ResNet-based, attention-based) suffer from high parameter counts and hardware demands.
• Existing deep structures often neglect shallow and channel-related features by relying solely on the last frame-level layer's output.
• This leads to suboptimal feature representation and limits the model's ability to capture comprehensive speaker characteristics.

Purpose of the Study:

• To develop a computationally efficient and high-performing speaker verification model.
• To address the limitations of existing models by incorporating both shallow and deep features effectively.
• To enhance feature representation by capturing long-distance dependencies and channel interactions.

Main Methods:

• Proposed a shallow speaker verification model utilizing Lambda-vectors as its core component.
• The model's architecture is built upon three Lambda-SE modules designed to extract long-distance dependencies and channel-related information.
• Implemented multi-layer feature aggregation to fuse features from different frame-level layers, enriching deep feature information.

Main Results:

• The proposed Lambda-vector based model demonstrated a more stable training speed compared to baseline models.
• Achieved a significant reduction in model parameters, leading to lower hardware requirements.
• Exhibited superior identification performance on the Voxceleb1 and Voxceleb2 datasets.

Conclusions:

• The novel shallow speaker verification model effectively overcomes the parameter and hardware limitations of current state-of-the-art methods.
• Multi-layer feature aggregation enhances the model's ability to represent complex information by utilizing both shallow and deep features.
• The Lambda-vector based approach offers a promising direction for developing efficient and accurate speaker verification systems.