Abstract: Audio adversarial attacks seriously threaten the security of artificial intelligence application. Most of the methods are only effective for specific attacks, and difficult to deal with different attacks. In this paper, we proved that there was an energy difference between adversarial example and original example, and proposed a detection model Noise-Energy based on the noise energy changes. The experiments show that the detection accuracy of the Noise-Energy model for CW attack reaches 99.5%, and for other attacks is more than 98%. It exhibits strong robustness and excellent generalization.