Speech enhancement is the process of removing noise to improve speech quality and intelligibility for applications including hearing aids. Many deep neural networks for speech enhancement have shown great ability in eliminating noise, regardless of its type. In hearing aids, this process may result in removing important noise used in emergency situations, such as fire alarms and car horns. In order to prevent this, a smart speech enhancement architecture is presented in this paper, where a convolution based noise classifier is used to detect emergency noise and activates the speech enhancement model to run in an audio enhancement mode, in which both the emergency noise and the speech are the target system output. The developed speech enhancement model is a deep convolutional recurrent network with several dilated layers to improve feature extraction without increasing network complexity. The results show that the speech enhancement model outperforms state of the art architectures by a 0.22 increase in the PESQ score. Moreover, the smart speech enhancement architecture improves speech and emergency noise quality when evaluated using objective metrics for both normal and hearing-impaired listeners.