Efficient Streaming LLM for Speech Recognition