Entropy-guided explainability for transformer audio models like Whisper
arXiv·low signal
This paper (arXiv 2606.14647, cs.SD/cs.AI) proposes an entropy-guided attention method to make transformer ASR models such as Whisper more interpretable, since their predictions are otherwise hard to explain. Given how broadly Whisper is deployed, an interpretability hook for diagnosing where and why an ASR model commits to a transcription is practically useful for debugging speech pipelines.