Deaf and hard of hearing people have difficulty noticing hazardous sounds around them. A device that enables the brain of a person to recognize the direction of a hazardous sound has already been developed. However, there is an issue that the device needs six seconds to recognize a hazardous sound. In this report, we describe our aim to develop a hazardous sound notification system which recognizes environmental sounds every second and notifies the deaf and hard of hearing people if a hazardous sound is detected. The teaching data is produced using four types of spectrograms converted from traveling noises, bicycle bell sounds, an ambulance siren and a fire engine siren. The accuracy of the deep learning model for each of the cases with log-scale spectrograms and mel-scale spectrograms are compared. Consequently, an accuracy of more than ninety percent is obtained using the deep learning model with mel-scale spectrograms.