Development of Scream Detection System with Large-Scale Scream Dataset

In this paper, the process of developing scream detection system is written. This research provides a basic technology Scream detection that can detect dangerous situations in public and private spaces can be helpful in public and personal safety, but it has not been studied with large scream data. The previous researches are based on small datasets in which there are less than 1,000 scream samples. To develop a scream detection system, we collected 11,921 scream samples. By utilizing the dataset, we evaluated and compared five different detection models that can accurately detect and classify screams. In our experiments, the model that combines CNN and Transformer outperformed in scream classification accuracy and F1 Score, demonstrating its effectiveness in processing and classifying speech data in real-world situations. Future research directions include improving the performance of scream detection in various environments, building additional datasets and optimizing the model, and developing the ability to recognize various kinds of dangerous situations.