Violence detection in audio: evaluating the effectiveness of deep learning models and data augmentation

Abstract

Human nature is inherently intertwined with violence, impacting the lives of numerous individuals. Various forms of violence pervade our society, with physical violence being the most prevalent in our daily lives. The study of human actions has gained significant attention in recent years, with audio (captured by microphones) and video (captured by cameras) being the primary means to record instances of violence. While video requires substantial processing capacity and hardware-software performance, audio presents itself as a viable alternative, offering several advantages beyond these technical considerations. Therefore, it is crucial to represent audio data in a manner conducive to accurate classification. In the context of violence in a car, specific datasets dedicated to this domain are not readily available. As a result, we had to create a custom dataset tailored to this particular scenario. The purpose of curating this dataset was to assess whether it could enhance the detection of violence in car-related situations. Due to the imbalanced nature of the dataset, data augmentation techniques were implemented. Existing literature reveals that Deep Learning (DL) algorithms can effectively classify audio, with a commonly used approach involving the conversion of audio into a mel spectrogram image. Based on the results obtained for that dataset, the EfficientNetB1 neural network demonstrated the highest accuracy (95.06%) in detecting violence in audios, closely followed by EfficientNetB0 (94.19%). Conversely, MobileNetV2 proved to be less capable in classifying instances of violence.