Optimizing Automatic Speech Recognition for Low-Proficient Non-Native Speakers