LaRA: Large Rank Adaptation for Speech and Text Cross-Modal Learning in Large Language Models F i n d i n g s - E M N L P