Deep Learning-Based Noise Reduction for Voice and Sound Classification of Individuals in Disaster Situations

This study introduces a noise reduction method aimed at improving the detection of survivors' acoustic and voice signals in disaster scenarios. The proposed approach improves signal quality in noisy environments by applying a deep learning-based Wave-U-Net model and then using a Wiener filter. The method was trained on a dataset that included synthesized urban noises and evaluated using three different models: SimpleCNN, SVM, and XGBoost. The evaluation results demonstrated that the combined noise reduction method significantly improved performance in noisy conditions for the SimpleCNN and XGBoost models, with gains of 3.59% and 9.55%, respectively. However, no significant improvement was observed for the SVM model. These findings indicate that the proposed method is particularly effective for deep learning models in classifying disaster signals under noisy conditions. The study suggests that future research should focus on applying this method in real-time and more complex environments to further validate its effectiveness.