Whispering LLaMA: A Cross-Modal Generative Error Correction Framework for Speech Recognition