Current Large Audio Language Models largely transcribe rather than listen