Interpretability methods are gaining popularity for understanding large language models, but they are underexplored in automatic speech recognition (ASR). This work applies techniques like logit lens and linear probing to ASR systems, revealing how acoustic and semantic information evolves across layers.