Post

Presented our work on layer pruning of the Whisper encoder. We show that if we’re careful about which layers we prune away, we can remove 6 (18.5%) of them without major performance drops and even recover performance through simple distillation.