Adaptive Feature Selection for End-to-End Speech Translation