Efficient Real-Time Acoustic Monitoring System with Convolution Recurrent Neural Networks

In today’s world, public safety is of greater importance. Based on the National Crime Records Bureau, India has reported that around 3.5 million crimes on a yearly basis in which emergency cases require immediate attention. The existing solutions for the current public safety systems rely on traditional audio classification system which lacks efficiency and accuracy. This project proposes an approach which is more innovative in order to classify audio using Convolutional Recurrent Neural Networks (CRNN) for public safety applications. The audio classification process includes classifying the emergency sounds such as gunshots, screams, car horns, siren from the real-world scenarios. This CRNN architecture combines the use of Convolutional Neural Networks and Recurrent Neural Networks. It is used to extract both the spatial as well as temporal features from the audio and classify them efficiently. Our experimental results provided that this proposed CRNN model achieves an accuracy of 97% in detecting the emergency sounds. The model's performance can be further validated by its F1-score percentage 97.1, recall percentage 97.2 and precision percentage 97.1. These performance metrics confirms that the CRNN algorithm is more robust and efficient than any other algorithms like CNN, RNN, Long Short-Term Memory (LSTM) and conventional Machine Learning algorithms like Support Vector Machine (SVM), Random Forest, etc. By this way, this research’s model is able to make better improvements in public safety and save the lives in an efficient manner.