A normalization knowledge distillation method for a speaker verification task
By decomposing the distillation loss into target and non-target class losses using the normalized knowledge distillation method and introducing dynamic weight adjustment, the decoupling problem between target and non-target classes in the speaker verification task is solved, thereby improving the accuracy and robustness of the model.
Patent Information
- Application Number
- CN202610602904.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-30
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies for speaker verification tasks suffer from problems such as coupling of target and non-target information, suppression of non-target information, lack of adaptive optimization mechanisms, and weak theoretical foundation, resulting in poor knowledge distillation performance.
A normalized knowledge distillation method is adopted, which decomposes the distillation loss into the knowledge distillation loss of the target speaker and non-target speaker, and introduces a learnable decoupling weight parameter γ to dynamically adjust the non-target weight. A dynamic weight update mechanism is designed to achieve decoupling and adaptive optimization between the target class and the non-target class.
Significantly reduces equal error rate (EER) and minimum detection cost function (minDCF), improving model accuracy and robustness while maintaining the same model parameter size and computational complexity.
Smart Images
Figure CN122417036A_ABST