<p>This article presents a systematic review of artificial intelligence (AI) techniques for detecting violence in audio recordings. The objective was to identify and categorize preprocessing methods, model architectures, hyperparameter strategies, and datasets used between 2015 and 2024. Searches were conducted in major scientific databases, with studies selected through predefined inclusion and exclusion criteria. Data extraction focused on preprocessing pipelines, learning models, training configurations, and dataset characteristics. The review found that convolutional neural networks and lightweight architectures remain dominant, while transformers and multimodal approaches are emerging as promising alternatives. Common preprocessing methods included MFCCs, STFT, and Mel-spectrograms, often combined with data augmentation. Hyperparameters such as batch size, learning rate, and dropout were key drivers of performance. Datasets showed notable limitations, including imbalance, lack of demographic diversity, and scarcity of realistic acoustic conditions. Overall, results highlight both progress and persistent challenges. Future research should focus on developing standardized datasets, incorporating more robust evaluation metrics, and advancing explainability and ethical considerations. This review underscores the potential of AI to contribute to public safety and well-being through reliable audio-based violence detection.</p>