Speechworthy Instruction-tuned Language Models