Configure the assistants "start speaking plan". Add a little delay and enable smart endpointing. Also, give the instruction in the assistant prompt