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

Neural networks used for speaker identification can be made more transparent through visualization of their "attention mechanisms" — similar to how humans focus on relevant information, these AI systems highlight which parts of a voice recording they prioritize when identifying who is speaking.

Existing methods for evaluating whether these visualizations are actually reliable and trustworthy have significant gaps — the paper identifies shortcomings in current evaluation approaches, suggesting that previous claims about understanding these AI systems may not have been rigorously tested.

A new evaluation method (Modified RISE-eval) was developed to better assess attention map quality — this improved approach provides a more systematic way to verify whether the visualizations genuinely reflect what the AI system is actually using to make decisions, rather than just appearing plausible.

Explainability in AI requires both visualization AND rigorous evaluation — simply showing what parts of data an AI system focuses on is insufficient; responsible AI development requires validated methods to confirm these explanations are accurate and meaningful.

This work addresses a critical gap in responsible AI: the lack of standardized ways to verify that AI explanations are trustworthy — without proper evaluation frameworks, organizations and users cannot confidently rely on explanations of how AI systems make important decisions.