Towards Stealthy Backdoor Attacks against Speech Recognition via Elements of Sound