Abstract <p>The present work addresses the urgent need to develop technologies aimed at reducing the incidence of violent events in society, particularly in contexts where vulnerable communities face constant fears of lethal violence. This research also explores innovative approaches to early detection and intervention strategies that can prevent violent incidents before they occur. By leveraging advanced data analytics and machine learning techniques, the study aims to identify patterns and risk factors associated with potential outbreaks of violence. Additionally, the project seeks to develop community-based solutions that empower local stakeholders to actively participate in creating safer environments for all residents. In this context, the acoustic approach emerges as a promising aspect due to its intrinsic advantages, such as the ability for automated detection and nonocclusion of the environment. The main aim of this research is to contribute to the problem of identifying scenes of violence by training a machine learning model capable of identifying and classifying scenes such as screaming, physical violence, and firearm discharges. Convolutional neural network architectures were tested, in particular, the ResNet-152V2 and MobileNetV2 networks. The accuracy results revealed that both models achieved similar performances, scoring 87.42 and 86.58%, respectively, in the multi-classification task. The results here proved the viability and efficiency of the proposed method, showing great promise for convolutional neural networks in classifying violent scenes based only on audio. These findings therefore show that there is a very good prospect for audio-based violence detection systems to find applications in various real-world scenarios, such as content moderation for streaming platforms or public safety monitoring.</p>