Listening with Attention: Entropy-Guided Explainability for Transformer-Based Audio Models

Transformer-based speech recognition models like Whisper are accurate but opaque: The paper addresses a critical gap in responsible AI by tackling the "black box" problem in widely-used automatic speech recognition systems, where it's unclear how these models arrive at their transcriptions.

A new method (LEAF-X) significantly improves explanation quality: The proposed approach achieves 32% better faithfulness compared to existing explanation methods, meaning the explanations more accurately reflect what the model actually uses to make decisions—a key requirement for trustworthy AI systems.

Better explanations enable auditing and accountability: By producing precise, sparse explanations that show which parts of audio input influenced specific transcribed words, LEAF-X makes it possible to identify when and why ASR systems make errors, supporting more transparent deployment in high-stakes applications like healthcare or legal proceedings.

Technical design matters for responsible AI: Rather than using generic explanation techniques, LEAF-X leverages the internal structure of transformer models themselves, demonstrating that responsible AI requires understanding how systems actually work rather than applying one-size-fits-all interpretability tools.

Stability of explanations is crucial for real-world trust: The paper emphasizes that explanations must be consistent and reliable (showing "most stable attributions"), not just accurate—inconsistent explanations would undermine user confidence in AI systems even if technically sound.