In residential areas, the need to enhance traditional protection systems has become more pressing due to the increasing risks to people’s safety, whether from accidents or intruder attacks. Visual surveillance alone is rarely sufficient, so it’s essential to integrate sound classification with it. Sound classification is crucial in various aspects of life, with its most vital role being the preservation of human lives. This research proposes a system based on the lightweight ResNet-34 model that distinguishes abnormal sounds in the home environment, with a particular emphasis on detecting screams as an emergency signal. Our approach involves preprocessing the sound dataset to create Mel-spectrograms, which are used as inputs for the models. ResNet-34 algorithm is trained with different approaches to accurately differentiate between scream and not scream sounds without requiring additional noise reduction techniques. The results demonstrate that the lightweight ResNet-34 achieved a higher accuracy of 96.8% compared to the models with and without transfer learning, 93.9% and 94.9%, respectively. Moreover, lightweight ResNet-34 achieves approximately a 16.1% reduction in file size compared to the original ResNet-34. Our findings reveal that the proposed system can be effectively deployed in various environments, including schools and private places as well as public ones, and can be expanded to larger datasets to identify as many abnormal sounds as possible.