arXiv:2606.14647v1 Announce Type: cross Abstract: Transformer-based automatic speech recognition (ASR) models such as Whisper are highly accurate, but their predictions remain difficult to interpret. Existing explainable AI (XAI) methods often lack faithfulness and precise temporal grounding. We propose Listening with Entropy-guided Attention for Faithful explainability (LEAF-X), a model-intrinsic XAI framework for transformer-based ASR. LEAF-X combines entropy-guided attention weighting, multi-layer attention rollout, and optional causal ablations to identify low-entropy, high-impact heads an
The increasing prevalence of sophisticated transformer-based AI models like Whisper highlights an urgent need for robust explainability methods to improve trust and troubleshoot performance issues, especially in critical applications.
Improving the interpretability of complex AI models is crucial for their adoption in high-stakes environments, enhancing debugging, auditing, and user confidence, thereby accelerating the deployment of advanced AI systems.
This research introduces a novel, model-intrinsic XAI framework that promises more faithful and temporally precise explanations for transformer-based audio models, potentially setting a new standard for AI interpretability in this domain.
- · AI developers
- · Auditors
- · Regulatory bodies
- · AI-driven product companies
- · Black-box AI models
- · Legacy XAI methods
Increased trust and adoption of advanced transformer-based AI systems, particularly in ASR.
Development of new AI compliance and auditing standards incorporating such explainability methods.
Acceleration of AI integration into sensitive sectors like healthcare and finance, contingent on interpretable models.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI