Language Models as Semantic Teachers: Post-Training Alignment for Medical Audio Understanding