Explainable AI in Speaker Recognition -- Attention Map Visualisation and Evaluation

This article presents a study on the explainability of neural networks in speaker recognition, focusing on the visualization and evaluation of attention mechanisms. It critiques existing class activation map (CAM)-based methods, proposing a new evaluation algorithm, Modified Randomised Input Sampling for Explanation - Evaluation (Modified RISE-eval), to assess attention maps generated by GradCAM and LayerCAM. The findings highlight the strengths of each method under varying conditions, emphasizing the importance of robust evaluation in enhancing the interpretability of AI systems for practitioners developing speaker recognition technologies.